跳到正文
Jakob Nielsen · UX Tigers· Jakob Nielsen·· 7 小时前精选AI 评分84

Jakob Nielsen复盘服务即软件趋势:按结果付费重塑UX职责

Service as a Software: The New SaaS Is Ahead of Schedule

文章摘要

专业工作正转变为按结果计费的公用事业,这一模式将交付失败的风险从买方转移至卖方,但也引入了“幽灵解决”等新隐患。Fin、Sierra等厂商已实现数亿美元年经常性收入,证明按解决工单或完成合同收费的模式具备商业可行性。然而,当前多数计量系统依赖客户静默时长来判定任务完成,这种“结果清洗”可能导致模型优化方向偏离真实用户满意度。

对于体验从业者而言,核心职责已从屏幕交互设计扩展至服务契约的设计与审计。这包括定义诚实的“已解决”标准、设计透明的报价与进度检查点、建立有效的申诉与人工接管路径,以及确保在自动化流程中保留人类对异常情况的监督能力。随着采购代理逐渐取代人类成为主要买家,机器可读的服务指标(如成功率、争议率)将成为新的竞争壁垒。从业者需主动介入工作流重构,防止因过度依赖自动化而导致初级人才断层及验证环节失效。

推荐理由

解析按结果付费模式下的风险转移机制,指出“沉默即解决”的计量漏洞,为从业者提供从界面设计转向服务契约审计的具体路径。

正文 · 原文

Summary: AI vendors now sell finished work and charge per outcome, and one pure-play vendor has already passed $400 million in annual recurring revenue. Outcome pricing shifts the risk of paying for failed work to the seller, while the consequences of failure and uncounted “phantom resolutions” remain the buyer’s problem. Designing and auditing the meter is now a core UX job.

(All images in this article made with GPT Image 2.6.)

This is a substantially revised second edition, published in November 2026, of the article I first published on December 19, 2024. I kept the arguments that held, graded the ones that didn’t, and added what we’ve learned since. The original 2024 version is still on UX Tigers if you want to check my reasoning at the time. The jazz music video (YouTube, 2 min.) I made for the first edition still applies.

Turn a tap, and clean water flows: metered, priced by the liter, no plumber required for a glass of water. Professional work is becoming the same kind of utility. State the outcome you want, and finished work flows out of the work faucet while the meter runs.

Nobody needs a statement of work for a glass of water. Professional work is becoming the same kind of utility.

A resolved support ticket costs $0.99. A legal research memo arrives in minutes. A literature review that once justified a consulting engagement costs about as much as lunch. The old SaaS sold you tools and wished you luck. The new SaaS delivers the finished work and, at its best, charges only when the job is done.

When a literature review costs less than lunch, people start ordering it like fries.

Software as a Service (the original SaaS) is the cloud model in which vendors host applications and sell subscriptions to access them. Remote computing services began around 1964 with IBM’s time-sharing systems; the modern SaaS concept emerged in the late 1990s with Salesforce and Concur, and Gmail took it mass-market in 2004. Throughout that history, one rule held: the software helped you do the work, but completing that work remained your responsibility. That rule just broke.

The old SaaS sold you tools and wished you luck. Completing the work remained your responsibility.

Old SaaS vs. new SaaS.

Grading My 2024 Claims

I grade my old predictions in public because futurists who don’t keep score are just entertainers. Here’s the scorecard for the first edition, almost two years later.

Futurists who don’t keep score are just entertainers.

Scorecard: My December 2024 Claims
Claim (December 2024)	Verdict (November 2026)
Intelligence gets ~100x cheaper every 2 years	RIGHT so far, for a fixed level of capability
Services will be packaged and sold as software	RIGHT, and faster than predicted
My list of example products (UiPath, QuickBooks, Bloomfire, hotel chatbots)	WRONG LIST, right category
AI can only facilitate user research; participants must stay human	RIGHT on the participants; TOO NARROW on AI’s other roles
AI will be able to facilitate user research by 2027, if not 2025	RIGHT, within the forecast window
UX consultants displaced within 4–5 years	ON TRACK at the halfway mark
AI won’t cause unemployment among design professionals	HOLDING overall, but beginners are locked out
3 million UX professionals will grow to 10 million by 2034	BEHIND SCHEDULE, reaffirmed below
Service as a software adds $50T to the world economy in 20 years	TOO EARLY, restated below

Of the 5 claims that can be graded today, 3 were right, 1 was mostly right, and 1 was wrong. A decent prediction score in my book! The other 4 are still running, and I return to the jobs forecasts below. Let’s start with the one I got wrong.

My list of examples was a dud. In 2024, I pointed to robotic process automation (RPA), accounting software, and hotel chatbots as early service-as-a-software products. Those were tools that helped humans deliver services faster. The companies that embodied my thesis were already in business, and I overlooked them. Fin, Sierra, Harvey, and Salesforce’s Agentforce had all launched before my article ran.

Worse, Sequoia has since made QuickBooks, one of my picks, its emblem of the tool that the service will replace. You can get the destination right and still name the wrong vehicles.

My 2024 list of example products was a dud. You can get the destination right and still name the wrong vehicles.

On user research, my 2024 line held with two amendments. AI facilitation arrived within my “by 2027, if not 2025” forecast window. In December 2025, Anthropic’s AI interviewer ran 1,250 interviews for a single study. Evaluation participants must remain human because synthetic users still distort results; a 2026 benchmark by Zihan Chen and co-authors found that LLM survey respondents inflate segment gaps 2–4x.

Now for the amendments. First, simulated users do earn their keep in pilot sessions, debugging study plans before you spend time with real customers. Second, heuristic prediction with zero participants (which my Usability Scaling Law analysis projects will account for 2/3 of design quality by 2035) has become a major category I didn’t list. Even so, my humans-vs-AI research rule stands: real customers get the final word, because the purpose of research is to be surprised.

Simulated users earn their keep in pilot sessions. Real customers still get the final word, because the purpose of research is to be surprised.

The consultants are being repriced before they’re being replaced. At a November 2025 press briefing, McKinsey disclosed that about a quarter of its global fees came from outcome-based pricing, in a consulting business that used to bill by a project’s scope and duration. Two months later, its CEO counted 25,000 AI agents in the firm’s workforce, alongside 40,000 humans. EY’s head of growth and innovation has floated a “service-as-a-software” model, in which clients pay for outcomes in place of hours. When the world’s most expensive advisors start selling by the result, the meter has reached the professions.

The consultants are being repriced before they’re being replaced. When the world’s most expensive advisors start selling by the result, the meter has reached the professions.

Re-Running My 2024 Research Request

The first edition contained a small experiment, in which I asked Google’s then-new Deep Research product to scour the literature for recent examples of professional service work being taken over by AI. I noted at the time that this $20 query replaced a research task that consulting firms billed millions of dollars for, minus the snazzy slide deck and the consultant’s polished stage presence.

A $20 query now replaces research that consulting firms billed millions for, minus the snazzy slide deck and the polished stage presence. I used to be a consultant, so I know which part clients paid for.

For this second edition, I asked again, this time using Google’s current flagship model, Gemini 4 Carbon. The difference between the two answers tells the story of the last 23 months:

Then vs. Now: The Same Research Question, 23 Months Apart
	December 2024: Deep Research	November 2026: Gemini 4 Carbon
What the products do	Help humans deliver services (chatbots, RPA, analytics)	Execute the service end to end, escalating exceptions and high-stakes judgment to humans
Representative names	Hilton chatbots, DHL routing, UiPath, QuickBooks	Garfield AI, Crosby, Harvey, Basis AI, Cognition’s Devin, Sierra, Fin (formerly Intercom), McKinsey’s 25,000 agents
How they charge	Subscriptions and licenses	Outcome fees (per billed resolution or completed filing), action credits, and fixed-scope tiers
Who answers for the result	The customer’s employees	A split: licensed AI-native firms, vendors whose guarantees are backed by AI insurers, and the buyer’s executives who sign off

In 23 months, the answer flipped from “software that assists service workers” to “software that is the service worker and bills like one.”

The surprise guests on Gemini’s list are two regulated law firms, seated beside the software vendors. In 2025, Garfield AI became the first AI-driven firm authorized by the solicitors’ regulator for England and Wales, chasing unpaid invoices at £2 for a demand letter and £50 for a court claim. Crosby, a registered New York law firm, has run 13,000 contracts through AI agents, with a lawyer checking the final work, at $250–$1,000 apiece. The faucet now holds a law license.

Two regulated law firms now run their work through AI at a fixed price per finished document. The faucet now holds a law license.

The 2024 Deep Research output also shows why I missed the early outcome-priced startups. A literature review finds only what has already been written up. Notice the recursion, too: both editions of this article were partly researched by the work faucets they describe.

Outcome-Based Pricing Went from Startup Experiment to Accounting Guidance

Definition: Outcome-based pricing charges the customer only when the AI completes a defined unit of work, such as a fully resolved support ticket. No resolution, no invoice.

The scorecard can’t show one omission, because the first edition never mentioned it: this pricing model was already on sale in December 2024. Fin had billed per resolution since 2023, and Zendesk announced per-resolution pricing in the summer of 2024. That fall, Salesforce launched Agentforce at $2 per conversation, which put a meter on the work, if not yet on the result. I described what the faucet would deliver and overlooked how it would bill. Two years on, the meter is the liveliest part of the story:

A price of $2 per conversation puts a meter on the work, if not yet on the result.

  • Fin (formerly Intercom): Charges $0.99 per resolved outcome, with unresolved chats free. After passing $100 million in annual recurring revenue (ARR) in April 2026 at 3.5x annual growth, Intercom renamed itself Fin in May and sold to Salesforce for $3.6 billion in September 2026. Fin claims a 76% autonomous resolution rate.

  • Zendesk: Bills roughly $2.00 per automated resolution on a pay-as-you-go plan ($1.50 on committed volume), charging only when no human touches the ticket.

One vendor bills for an automated resolution only when no human touches the ticket.

  • HubSpot: Shifted its Breeze agent in April 2026 from $1.00 per handled conversation to $0.50 per resolved conversation. That one-word edit moved performance risk from buyer to HubSpot.

  • Salesforce: Agentforce passed $1.5 billion in ARR (up 240%+ year over year, bundling Slackbot and sibling AI tools) in July 2026, logging 7 billion “Agentic Work Units.” Its price sheet sells the caterpillar and the butterfly side by side: per user ($125+ per month), per action (about $0.10), per conversation ($2), and, since mid-2026, per resolution ($2). The Fin acquisition now bolsters its outcome-based offering.

One vendor’s price sheet now sells the caterpillar and the butterfly side by side: per user, per action, per conversation, and per resolution.

  • Sierra: Built strictly on outcome pricing, Sierra passed $400 million in ARR in October 2026 after adding more than $150 million in Q3 alone. It openly needles seat-based rivals over their built-in conflict of interest: when an AI agent works, the customer needs fewer human seats.

  • Harvey: Passed $400 million in ARR across more than 3,000 organizations and raised capital at a $15.5 billion valuation in September 2026. Yet Harvey still bills mostly per lawyer seat, so the faucet pours finished legal work while the meter counts chairs.

The broader market is following, though more slowly than the headlines suggest. Kyle Poyar’s 2025 survey of 240 software companies found that seat-based models fell from 21% to 15% of companies within 12 months, while hybrid models jumped from 27% to 41%. Still, only 5% of his respondents named outcome-based pricing as their primary model, although 25% expected to get there by 2028. Even OpenAI, which sells intelligence by the token, is testing pay-on-success deals with a few enterprise customers, as first reported by The Information.

The broader market is following, though more slowly than the headlines suggest: only 5% of surveyed software companies named outcome-based pricing as their primary model.

The development that convinced me the change is permanent came from the accountants. On June 4, 2026, Deloitte published revenue-recognition guidance for outcome-based pricing in agentic AI products, complete with worked examples: a vendor that collects $5 million up front for 250,000 fully resolved refund cases may book each $20 payment only when the corresponding case is resolved, and a fraud-monitoring vendor books $3 per successful autonomous intervention on top of a fixed access fee.

When Big-4 accountants explain how to book a pricing model, that model has stopped being a startup experiment and become infrastructure. Service as a software now has a GAAP treatment (GAAP being the generally accepted accounting principles that auditors swear by). I couldn’t have asked for a stranger compliment.

When Big-4 accountants explain how to book a pricing model, that model has stopped being a startup experiment and become infrastructure.

One catch hides in the fine print, and it’s a design problem more than an accounting problem. What counts as “resolved”? All three published meters can count customer silence as success.

Fin bills an “assumed resolution” when the customer goes quiet for 24 hours after the agent’s last answer. HubSpot counts a conversation as resolved if its agent shared a help article or performed an action and the customer didn’t ask for a human within 72 hours. Zendesk waits out an inactivity window of 2–72 hours, depending on the channel, and then has an AI model read the transcript to verify the resolution.

(Credit where it’s due: Zendesk added that check in May 2026, and Fin drops the charge if the customer returns to the same conversation. But the meter reader still works for the gas company.)

Credit where it’s due for the vendors’ extra checks on what counts as resolved. But the meter reader still works for the gas company.

Silence may mean the customer gave up and switched vendors. The meter counts the silence anyway. Gartner calls it agent washing when a chatbot gets relabeled as an agent. Billing for silence under an outcome-pricing banner is outcomewashing. The billing unit of the work-faucet economy is a measurement artifact, and measurement artifacts are UX territory.

Billing for a customer’s silence under an outcome-pricing banner is outcomewashing.

Why the Outcome Meter Started at the Help Desk

Every outcome-priced success story above sells customer support, while Harvey’s seat licenses show that the flip in what is sold has outrun the flip in how it’s billed. An outcome meter needs three conditions that most professional work doesn’t yet meet.

Condition 1: a system of record already registers the outcome. A ticket is open or closed in the help-desk software. A loan is funded, a claim is paid, a meeting is booked, and a database says so. Where no ledger records success, vendor and customer have nothing to count together, which is why legal memos, strategy advice, and most code still sell by the seat or by the token.

Condition 2: the work is reliable enough for the vendor to carry the failures. Under outcome pricing, every failed attempt comes out of the vendor’s margin. METR, the research group that measures the length of tasks that AI can complete on its own, estimated in spring 2026 that the most capable model it assessed had a 50% chance of finishing software tasks that take human experts 16–20 hours, but reached 80% reliability only on tasks that take 3–4 hours.

A 50% success rate is a demo, 80% is a business, and 99% is a utility. (METR doesn’t even report a 99% horizon.) Support questions take minutes, which puts them well inside the reliable horizon. So the meter was installed at the help desk first.

A 50% success rate is a demo, 80% is a business, and 99% is a utility.

Condition 3: the buyer can specify “done” before the work starts. Here my water metaphor springs a leak. Every liter from the kitchen tap is the same H₂O as the last; nobody writes a specification for wet. Knowledge work is different, and 40 years of usability research tell us that humans are terrible at stating up front what they want. In old-school work, a lawyer, designer, or executive often discovered while struggling through a first draft that he or she was solving the wrong problem.

Nobody writes a specification for "wet."

Knowledge work is different, and 40 years of usability research tell us that humans are terrible at stating up front what they want.

Outsource execution to a work faucet, and you risk the Midas trap: getting exactly what you asked for, with a valid invoice attached, before realizing you asked for the wrong thing. Support tickets dodge this trap because “issue refund” needs no creative discovery. Strategy memos and UX redesigns do.

The Midas trap: getting exactly what you asked for, with a valid invoice attached, before realizing you asked for the wrong thing. (In Greek myth, King Midas wished for everything he touched to turn to gold. Unfortunately, “everything” included his lunch.)

These conditions are spreading. Sierra now sells agents that pursue goals over days or weeks and bills for outcomes such as a refinanced loan or a paid insurance claim, which end as database entries that both parties can count. The professions follow a rung behind. Garfield and Crosby have swapped the billable hour for a fixed price per finished document, which puts a meter on the work product, though not yet on its result. The outcome meter spreads in order of verifiability. So to see where the flip arrives next, look for work whose result lands in a system of record.

Seats Were Gym Memberships; Outcomes Are Warranties

Behind the vendor names, one thing happened: the risk of failure moved from the buyer to the seller. Under the old SaaS model, the customer paid for seats whether or not anybody sat in them; Zylo’s 2026 index of enterprise software spending found that the average organization uses only 54% of its SaaS licenses, leaving 46% to gather dust.

Under the old SaaS model, the customer paid for seats whether or not anybody sat in them. The average organization leaves 46% of its licenses unused.

Seat licenses are the gym memberships of enterprise software. Stefano DellaVigna and Ulrike Malmendier’s 2006 study of 7,752 health-club members (PDF) found that people on monthly contracts showed up 4.3 times a month and so paid more than $17 per visit, even though a $10 per-visit pass was available. The gym’s most profitable member is the one who never comes in, and for a quarter century, the SaaS vendor’s most profitable seat was the one nobody sat in.

Seat licenses are the gym memberships of enterprise software. The gym’s most profitable member is the one who never comes in.

Outcome pricing inverts the incentive: a faucet that delivers nothing bills nothing. It also solves an older problem. In a 1970 paper, George Akerlof explained what happens when buyers can’t judge quality before they buy (used cars were his example): good sellers get lumped in with bad ones, prices sink toward the junk, and the market shrivels. The insight later won Akerlof a Nobel Prize.

Akerlof’s lemons problem: when buyers can’t judge quality before they buy, good sellers get lumped in with bad ones, prices sink toward the junk, and the market shrivels.

The professions solved their version of this lemons problem with credentials, from bar exams to the partner’s name on the door.

Outside a few licensed newcomers such as Garfield, AI services have no bar exam, and a plausible-sounding chatbot is a lemon you can’t spot on the lot. So the market reached for Akerlof’s other remedy, the warranty. “No resolution, no charge” is a warranty on expertise. You needn’t judge the machine’s competence in advance when you pay only for competence demonstrated.

The professions solved their lemons problem with credentials. AI services have no bar exam, and a plausible-sounding chatbot is a lemon you can’t spot on the lot.

Why can the seller carry this risk when the buyer couldn’t? The legal scholar Guido Calabresi supplied the rule in 1970, the year of Akerlof’s lemons paper: a loss should fall on the party who can avoid it most cheaply. Here, that party is the vendor, who controls the model, the prompts, and the tuning. The vendor can also pool the risk. One failed ticket in a customer’s 500 is an irritation, but one in a vendor’s 50 million is a line in an actuarial table.

One failed ticket in a customer’s 500 is an irritation, but one in a vendor’s 50 million disappears into the average.

But Calabresi’s rule hits a snag when failure has two parents. The vendor controls the model and the prompts, while the buyer controls the underlying product, the refund rules, and the rotting help documentation. If a retailer ships a broken checkout flow or contradicts its own return policy, no support agent, human or silicon, can resolve the resulting tickets, yet pure outcome pricing makes the vendor eat the compute bill for the buyer’s mess.

The rule that a loss should fall on whoever can avoid it most cheaply hits a snag when failure has two parents.

Economists Sugato Bhattacharyya and Francine Lafontaine at the University of Michigan proved in 1995 (21-page PDF) that whenever both buyer and seller supply unobservable effort, pure pay-for-results contracts break down under double-sided moral hazard.

Free failures corrupt both sides of the pipe. Buyers dump unfixable edge cases into the queue because unresolved chats cost them nothing. Vendors retaliate with cream-skimming: tightening eligibility gates, steering users toward trivial questions that close cleanly, or punting thorny tickets to humans the instant expected inference cost exceeds $0.99. Boasting a 76% resolution rate is easy if you get to disqualify the ugly 40% of the queue at intake.

Boasting a 76% resolution rate is easy if you get to disqualify the ugly 40% of the queue at intake.

Hence hybrid pricing (41% of software firms in Poyar’s survey) is growing faster than pure outcome billing (5%). The same conflict helps explain why real water and power utilities settled on two-part tariffs: a modest meter or diagnostic fee so the buyer keeps his or her own plumbing in order, paired with an outcome fee so the vendor delivers.

But a warranty refunds the fee and nothing more. Outcome pricing moves the price risk to the vendor, since a failure that the meter catches costs the customer nothing. The risk of consequences stays put, in the customer’s lap. When the agent wrongly denies a refund, invents a policy, or closes a ticket on a customer who gave up, the meter records a success, the vendor bills for it, and the client company inherits an angry customer.

The law has already said who answers when the chatbot gives a customer bad information. In February 2024, a Canadian tribunal ordered Air Canada to pay for a bereavement discount that its chatbot had invented, calling the airline’s suggestion that the bot was responsible for its own actions “a remarkable submission.” The award was C$812. The precedent is worth more.

One airline suggested that its chatbot was responsible for its own actions, which the tribunal called “a remarkable submission.” The award was small. The precedent is worth more.

The professions paid steeper tuition for the same lesson. In October 2025, Deloitte agreed to refund the final installment on an AU$440,000 report for the Australian government, drafted with help from GPT-4o, after fabricated citations turned up in it, including an invented quote from a federal court judgment. Whoever signs the work answers for the machine. This is where the licensed faucets shine. Garfield’s regulated solicitors answer for its letters, and Crosby is liable for every contract it reviews. In those cases, the risk of consequences travels with the work.

Whoever signs the work answers for the machine.

An outcome meter is a detector, and like every detector, it can err in two directions:

The 4 Outcomes of an Outcome Meter
	Meter reads “resolved” (billed)	Meter reads “unresolved” (free)
Customer’s problem solved	Earned revenue: the warranty worked as advertised	Unbilled success: a loss for the vendor, presumably rare
Customer’s problem not solved	Phantom resolution: the buyer pays the invoice and bears the damage to its customer relationship	Honest failure: the vendor eats the cost, and a human can take over

The warranty covers the bottom-right cell. The bottom-left cell is where outcomewashing lives, and neither party has an immediate incentive to count it, because the vendor gets paid and the buyer’s dashboard shows a resolution. Shrinking that cell is a UX job.

The bottom-right cell has news for UX professionals, too. Under seat licensing, the vendor of a confusing product still got paid; lock-in and shelfware saw to that. Under outcome pricing, every abandoned intake, misunderstood brief, and botched handover is unbilled work. My old slogan, UX = Profits, was a claim I had to argue for in budget meetings. Outcome pricing makes it a bookkeeping identity: the vendor’s usability is audited by its own revenue line, one unresolved ticket at a time.

The Flip, Dimension by Dimension
Dimension	Software as a Service (old SaaS)	Service as a Software (new SaaS)
What the customer buys	Access to a tool	A finished outcome
Billing unit	A seat per month	A successful outcome, an action, or a seat
Who bears the risk of failure	The customer (shelfware, stalled rollouts)	The vendor for failures the meter catches; the customer for the ones it misses
What the vendor optimizes	Renewals and seat counts	Higher completion rates and fewer customer disputes
What the service-level agreement (SLA) promises	Availability: the tool is up 99.9% of the time	Completion: the job gets done, say, 92% of the time
Where the UX lives	The screen	The service contract: scope, price, progress, proof, redress
Who places the order	A human with a login	Increasingly, an agent with a mandate
What the buyer must audit	Usage: are the seats filled?	The meter: does “resolved” mean resolved?
Utility precedent	Gas by the burner and the hour	Gas by the cubic foot, electricity by the kilowatt-hour

The Plumbing: 4 of 5 Layers Are Already Built

Besides faucets, a utility needs meters, pipes, a directory of what’s on tap, and inspectors. The first 4 layers were built within two years, and construction of the fifth has begun:

Besides faucets, a utility needs meters, pipes, a directory of what’s on tap, and inspectors.

The Service-as-a-Software Stack
Layer	What it does	2026 exemplars
Faucets	Deliver finished work in a vertical	Fin and Sierra (support); Harvey, Garfield AI, and Crosby (legal); Basis (accounting); agentic coding tools; deep research products; generative UI for design
Meters	Define, measure, and bill claimed outcomes	Per-resolution billing, Agentic Work Units, Deloitte’s recognition guidance, metering engines bought by Stripe and Salesforce
Pipes	Let agents pay and get paid	The x402 payment protocol, Google’s AP2, the OpenAI–Stripe Agentic Commerce Protocol, Visa Intelligent Commerce, Mastercard Agent Pay
Directory	Let buyers (human or machine) find faucets	AI answer engines, agent-readable catalogs, generative engine optimizationInspectors	Certify and insure services, with outcome audits still to come	Regulators licensing AI-native law firms; AIUC-1 agent certification with insurance from the Lloyd’s market; the outcome counts are still unaudited

The meter layer deserves the closest look, because the history of utilities says that’s where the bill gets decided. A utility is born the day its meter becomes trustworthy. The first gas companies had no meters, so they sold gas by contract hours and later by the burner (PDF), and inspectors patrolled at night to shut off customers whose paid-up hours had run out. Per burner was the seat license of the gaslight era.

The first gas companies had no meters, so inspectors patrolled at night to shut off customers whose paid-up hours had run out. Per burner was the seat license of the gaslight era.

Samuel Clegg’s water-sealed meter arrived in 1815, and with it came decades of quarrels over meters that “ran fast,” until Parliament’s Sale of Gas Act of 1859 required every meter to be tested and stamped by a public inspector before it could bill anybody.

Electricity replayed the script. Thomas Edison lit lower Manhattan from Pearl Street on September 4, 1882, but sent no bills until the system ran reliably, and the first electricity invoice, for $50.44, reached the Ansonia Brass and Copper Company on January 18, 1883. Every utility progresses through the same 4 stages: flat rate, meter, meter dispute, and an inspector’s stamp. The AI work faucet has reached stage three. Deloitte, bless its accountants, has told vendors how to book what the meter reads.

Flat rate, meter, meter dispute, inspector’s stamp: every utility, from Victorian-era city gas to AI today, goes through these 4 stages, and the AI work faucet has now reached the third.

The industry has noticed where the value sits. Stripe paid a reported $1 billion for the metering startup Metronome, and Salesforce signed a deal for the metering firm m3ter a week before its deal for Fin. Stripe’s CEO, Patrick Collison, calls metering and billing “the interface between ‘product’ and ‘business.’” I’ll go one step further: for a work faucet, the meter is part of the user interface.

For a work faucet, the meter is part of the user interface. So test it like one.

Electricity supplies one more precedent, and AI is replaying it. The standards that made electrical products safe came from insurers. After the 1893 World’s Fair in Chicago, fire underwriters funded the testing laboratory that grew into Underwriters Laboratories. In 2026, a startup called the Artificial Intelligence Underwriting Company certifies AI agents against its AIUC-1 standard and backs the certificate with insurance from the Lloyd’s market; its first policy covers ElevenLabs’ voice agents for up to $50 million. Insurance supplies the missing half of the warranty, since it’s how the consequence risk gets priced. (It has to be bought separately. Since January 2026, standard policy forms have let ordinary insurers exclude generative-AI claims from general liability coverage, and by mid-year more than 2,000 such exclusions were on file across 49 states.)

But the underwriters test for safety and security. Nobody yet stamps the meter itself. No independent party checks whether a billed resolution was a resolution.

As with electricity, insurers are writing the safety standards for AI. But the underwriters test for safety and security. Nobody yet stamps the meter itself.

Venture capitalists reached the same flip from the revenue side, with smaller numbers. Foundation Capital’s Joanne Chen and Jaya Gupta named the shift in April 2024 (catching outcome pricing before I did) and sized it at $4.6 trillion across 4 job families and outsourced IT; Sequoia’s Julien Bek followed in March 2026, noting that enterprises spend $6 on services for every $1 on software.

Venture capitalists reached the same flip from the revenue side: enterprises spend $6 on services for every $1 on software.

My $50 trillion prediction is about additional annual economic output in 2044, measured in 2024 dollars. Foundation Capital’s estimate is about spending that service vendors could chase. The two figures answer different questions. The existing wage bill indicates the scale of work that AI might transform, but automating work already being done doesn’t make all that work an addition to the economy.

The additional output will come from long-neglected problems finally solved, better products and services, and productive uses for the time and resources released. I retain the $50 trillion prediction as a bet on the scale of those changes. Multiplying today’s knowledge-work bill by future economic growth isn’t enough.

Intelligence on Tap: The Economics Held

The first edition’s economic engine still hums, so I’ll compress it rather than repeat it. High intelligence used to be scarce. Perhaps 2% of the population operates at the IQ 130 level that most advanced business work requires, and each of those people needs a decade of education and experience to become useful, and even then usually in only one specialty. Scarcity meant high prices, and Economics 101 says high prices suppress demand. Nobody pays a $100,000 consultant to solve a $1,000 problem, so the world’s $1,000 problems went unsolved by the millions. These are orphan problems: too small to justify an expert’s fee, too big to ignore, so nobody adopted them.

Nobody pays a $100,000 consultant to solve a $1,000 problem. These are orphan problems: too small to justify an expert’s fee, too big to ignore.

Example: A usability audit priced at $15,000 is a line item that Alice’s 8-person startup will never approve, so her checkout flow stays broken and she loses customers without ever learning why. Price a comparable audit at $150 from a work faucet, and she buys one per sprint. Multiply Alice by every small business on Earth, and you can see where the growth comes from: the mass of demand that high prices used to strangle. Replacing the purchases that already happen is the smaller prize. Garfield’s clients show what happens when the price falls. An unpaid £900 invoice was never worth a solicitor’s fee, so it got written off. At £2 a letter, it gets chased.

At $15,000, Alice’s 8-person startup never buys a usability audit. At $150 from a work faucet, she buys one per sprint.

Then the price of intelligence collapsed. In the first edition, I reported that GPT-4-level capability had dropped a staggering 240x in price in 21 months. Independent trackers confirm the slope. Stanford’s AI Index documents that the cost of querying a GPT-3.5-level model fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold reduction. Epoch AI’s tracking across benchmarks finds inference prices falling between 9x and 900x per year depending on the capability tier, with the price of GPT-4-level performance on Ph.D.-level science questions dropping 40x per year. One asterisk: these are prices for a fixed level of capability. The frontier model of the day has never gotten cheap, and customers keep choosing it. Yesterday’s intelligence is nearly free by comparison.

The frontier model of the day has never gotten cheap, and customers keep choosing it. Yesterday’s intelligence is nearly free by comparison.

Cheaper = more bought. That three-word law from the first edition still carries the argument. When the price of expertise drops 1,000x, a $100 solution to a $1,000 problem clears the market, and the orphan problems finally get adopted. If demand rises a thousandfold, a thousand times more of them get solved. This is also why enterprise AI bills keep rising while unit prices collapse: falling water prices don’t shrink the water bill if everyone in the building opens 10 new taps. Agentic workflows fire dozens of model calls per task. Exploding consumption coupled with cratering unit costs is what a $50 trillion market looks like in the delivery room.

Cheaper = more bought. Falling water prices don’t shrink the water bill if everyone in the building opens 10 new taps.

Economists have had a name for this since 1865, when William Stanley Jevons noticed that James Watt’s more efficient steam engine had multiplied Britain’s coal consumption rather than curbing it, because cheaper power found a thousand uses that expensive power never could. The Jevons paradox powers the work faucet’s growth. Intelligence is the coal of the knowledge economy. Its price just fell by orders of magnitude, and the world will burn far more of it.

The Jevons paradox: a more efficient steam engine multiplied Britain’s coal consumption rather than curbing it. Intelligence is the coal of the knowledge economy.

The paradox comes with a condition, and that’s where my forecast could fail. Spending grows only if the quantity bought rises faster than the price falls. If expertise becomes 100x cheaper and the world buys only 50x more of it, the market shrinks by half, and my $50 trillion evaporates no matter how much work gets done. But I’m betting the other way because of precedent. The price of computing has fallen by a factor of millions since the mainframe era, and the world’s computing bill has only gone up.

My forecast depends on elastic demand: spending grows only if the quantity bought rises faster than the price falls. A squirrel can eat only so many nuts, however cheap, but companies can use almost limitless cognition to solve new problems and offer new services.

Of course, the buyer increasingly sees no token price at all. He or she sees $0.99 per resolved ticket against a median of $13.50 for a contact handled by a human, according to Gartner’s customer-service benchmarks. The unit of purchase moved from compute to outcomes, which is how utilities mature: nobody buys electricity by the coulomb.

The buyer increasingly sees no token price at all, only a price per resolved ticket. That’s how utilities mature: nobody buys electricity by the coulomb.

Designing the Service, Not the Screen

The venture capitalists analyze the economics of selling work. Almost nobody analyzes the experience of being served by software, and that experience will decide which faucets people trust with their money. A professional service has always included an intake conversation, an estimate, progress reports, a handover, and someone to yell at when it goes wrong, as well as the deliverable itself. Take those away, and the cheaper service degrades into a slot machine that sometimes dispenses documents. (In my taxonomy of AI interaction styles, a work faucet is a Task Agent with a price tag: a bounded job, delegated and delivered.)

Take away the intake conversation, the estimate, the progress reports, the handover, and someone to yell at, and the cheaper service degrades into a slot machine that sometimes dispenses documents.

Jan Carlzon, who ran the airline SAS (the one with wings, not the statistics package) in the 1980s, calculated that his 10 million yearly passengers each met about 5 employees for about 15 seconds, so the airline was judged in 50 million moments of truth a year, 15 seconds at a time. A work faucet has 4 moments of truth: the quote, the checkpoint, the handover, and the redress. Everything else is backstage.

G. Lynn Shostack drew that distinction into the first service blueprints in a 1984 Harvard Business Review article as the “line of visibility” between what the customer sees and what the organization does. Four decades later, the line has an engineering meaning: below it run the agent’s reasoning, tool calls, and retries; above it sits the service. Draw the line too low, and you dump the transcript on the customer; draw it too high, and he or she gets a spinner, then a bill.

Draw the line of visibility too low, and you dump the transcript on the customer. Draw it too high, and he or she gets a spinner, then a bill.

So here are my 10 design guidelines for agent-delivered services. The meter is part of the interface.

  1. Sell the outcome in the customer’s vocabulary. Offer “500 resolved tickets” or “a competitive teardown of 3 rivals,” never “2 million tokens” or “API calls.” The customer is buying work, so name the work.

Sell the outcome in the customer’s vocabulary. The customer is buying work, so name the work.

  1. Quote price, time, and confidence before the meter starts. A human contractor gives an estimate: what it costs, when it’s done, and what could change either number. Agent services must do the same, including an honest success-rate figure for this task type. “About $40, roughly 25 minutes, and I complete this kind of job successfully 92% of the time” beats any spinner ever animated. An unpriced open-ended run is the Slow AI equivalent of a taxi with a hidden meter. A quote is also an expectation, and since satisfaction is delivered quality relative to expectations, an honest 92% sets a more credible expectation than a boastful 100% that the service can’t deliver.

Quote price, time, and confidence before the meter starts. An honest 92% sets a more credible expectation than a boastful 100%.

  1. Stage a cheap fitting before the full meter runs. A bespoke tailor never cuts into the good cloth on the strength of a verbal request alone; he or she pins a cheap muslin mockup first, because clients recognize the wrong cut on sight even when they can’t describe the right one in words. That’s my old usability heuristic, recognition rather than recall, rebuilt for metered work. For any complex job, charge pennies (or nothing) for acceptance prototyping: 3 rough directions, a schema preview, or 5 synthetic edge cases that define “done.” Start the big outcome meter only after the human signs off on the fitting.

Stage a cheap fitting before the full meter runs. Clients recognize the wrong cut on sight even when they can’t describe the right one in words.

  1. Make the guarantee legible at the point of order. “No resolution, no charge” is a pricing model, but displaying it is a UX decision, and it does more to build trust than any amount of interface polish. A guarantee is a design element. Say what it leaves out, too: a refunded fee doesn’t repair the damage from a wrong answer.

Make the guarantee legible at the point of order, and say what it leaves out: a refunded fee doesn’t repair the damage from a wrong answer.

  1. Disclose progress on a schedule. Long-running work needs checkpoints, partial deliverables, and interruptions reserved for decisions only the customer can make, as I detailed in Progressive Disclosure: From Training Wheels to Week-Long AI Agents. Don’t trap the status of a $500 job inside a scrolling transcript.

Disclose progress on a schedule. Don’t trap the status of a $500 job inside a scrolling transcript.

  1. Deliver like a professional, not a printer. Findings first, in one screen; inspectable depth behind it. The handover moment is where a service earns its next order, and dumping 40 undifferentiated pages forfeits it.

Deliver like a professional, not a printer. Findings first, in one screen, with inspectable depth behind it.

  1. Design the redress path before the failure. Who does the customer contact when the agent botches the job? Name the escalation route to a human, state the refund rule, and log every disputed outcome. Redress covers the bill as well as the work, so contesting a charge for a phantom resolution should take one click. A service without redress is a bet.

Design the redress path before the failure. A service without redress is a bet.

  1. Punt with the context attached. When the faucet hands a job to a human, that human inherits the residue: the hardest cases, the angriest customers, and none of the practice the easy cases used to provide (Lisanne Bainbridge called this the irony of automation in 1983, and it bites harder in the AI age). So the handoff must carry the transcript, the attempted fixes, and the customer’s mood. Vendors should also publish their punt rate (the share of jobs handed back to humans) beside their resolution rate. A customer who re-explains the problem to a human has paid twice.

Punt with the context attached. A customer who re-explains the problem to a human has paid twice.

  1. Never let silence masquerade as satisfaction. The inactivity window is a bookkeeping convenience but a flimsy proxy for a happy customer. Ask for a one-tap confirmation, itemize the invoice into confirmed and assumed outcomes, have the assumed ones sampled by somebody who isn’t paid per resolution, and publish the dispute rate. Better yet, bill assumed outcomes at a discount. The vendors who measure honestly will win the enterprise deals, because procurement departments read the fine print for a living.

Never let silence masquerade as satisfaction. The inactivity window is a bookkeeping convenience but a flimsy proxy for a happy customer.

  1. Remember the customer. A returning client shouldn’t have to re-brief the service from scratch. Standing preferences, house style, and past deliverables are the loyalty program of the work faucet, and they’re also the switching cost that makes the business durable.

Remember the customer. A returning client shouldn’t have to re-brief the service from scratch.

None of these guidelines mentions a screen layout. In July 2026, Gartner estimated that AI agents put $234 billion of enterprise application spending, about 20% of the SaaS total, at risk by 2030 and declared that “the user interface is no longer a differentiation.” That’s half the story. The screen is receding as the main design object, and the service contract between human and machine is taking its place. That contract (scope, price, progress, proof, and redress) gets designed whether or not a designer is in the room. Today, pricing managers and lawyers mostly design it by default.

The service flow for agent-delivered work: intake, estimate, checkpoints, delivery, and redress. Every joint in this plumbing is a design decision.

Your Next Customer Is a Machine

The flip has a second twist: increasingly, the entity ordering from the work faucet isn’t a person. Procurement agents comparison-shop for services by price and verified success rate; personal agents order the research, the booking, and the refund chase on their principals’ behalf. (These are the Proxy Agents in my taxonomy of interaction styles.) Gartner predicts that by 2028, 90% of B2B purchasing will be intermediated by AI agents, pushing more than $15 trillion through agent exchanges. Discount Gartner’s enthusiasm by half, and you still need to get ready.

Increasingly, the entity ordering from the work faucet isn’t a person. Procurement agents comparison-shop for services by price and verified success rate.

Thus, every service-as-a-software vendor now serves two audiences. I described the content-side version of this in From PageRank to DeepRank: humans want narrative and reassurance, while machines want structured claims they can extract and compare. For services, that means machine-legible catalogs, published SLAs an agent can parse, and verifiable track records instead of testimonial pages. And it means designing the report your customer’s agent will file about you, because the human principal experiences your service secondhand, in a machine’s summary. That summary is your brand. A faucet that only humans can find will lose to an equally good faucet that agents can find, order from, and audit. Being invisible to AI buyers in 2026 is like being on page 5 of Google in 2006.

Being invisible to AI buyers in 2026 is like being on page 5 of Google in 2006.

Expect the service equivalent of a credit rating to emerge: machine-readable success rates, dispute rates, punt rates, and completion times, audited the way financial statements are. Vendors will resist publishing these numbers, but the first movers who publish anyway will be first in line for agent-mediated orders, because an agent can’t be charmed by a sales dinner or a well-rehearsed demo. Selling to machines is the ultimate test of substance over stage presence. (As a recovering consultant, I confess the profession had it coming.)

Expect the service equivalent of a credit rating: machine-readable success, dispute, and punt rates, audited the way financial statements are. Vendors will resist publishing these numbers.

But a good rating won’t guarantee access to the customer. The purchasing agent is the gatekeeper that decides which providers enter the comparison in the first place.

Power will thus concentrate around whoever controls the customer’s mandate, preferences, and purchasing history. A work faucet may become as replaceable as a washer, while the agent choosing among faucets becomes indispensable. If that agent earns referral fees or favors affiliated providers, better service won’t necessarily win the order. Automating procurement doesn’t repeal the economics of distribution.

So buyers should demand to see the selection criteria and insist on the right to move their briefs, preferences, and outcome histories from one purchasing service to another. Challenger routing, which I discuss below, works only when the buyer keeps enough control to send work elsewhere.

I predict that some of the strongest businesses in this market will own the purchasing relationship while buying execution from competing faucets. Service vendors must prepare for that negotiation as carefully as they calibrate their meters.

Proxy Agents will also kill “assumed resolutions” faster than any procurement auditor. Today’s 24–72-hour inactivity meters profit from human exhaustion. A mortal customer who gets stonewalled eventually gives up and goes to bed, and the vendor sends a $0.99 invoice for his or her silence. Now put the customer’s personal Proxy Agent on the other end of that chat. A machine chasing a $120 airline credit doesn’t get bored or go to sleep. It can spend a fraction of a cent to ping the thread at hour 23, minute 58, and keep chasing until the refund lands in the bank ledger. When a tireless buyer bot meets a stonewalling seller bot, every unresolved loop burns the vendor’s inference margin instead of padding its top line. Outcomewashing survives only while humans do the waiting.

Today’s inactivity meters profit from human exhaustion. A tireless Proxy Agent doesn’t get bored or go to sleep, so outcomewashing survives only while humans do the waiting.

10 Million UX Professionals, Reaffirmed (With a Twist)

The first edition predicted growth from about 3 million UX professionals worldwide in 2024 to 10 million by 2034, and readers have asked whether I’d like to drop that number down the memory hole given the bruising UX job market. I would not. But the grade so far is plain: two years in, the growth hasn’t started, as far as I can tell. No global census of the profession exists, but none of the indicators I follow shows an uptick, so the remaining 8 years must now deliver about 16% a year, up from the 13% that the original path required. I reaffirm the number anyway, with a twist: the composition of the 10 million has changed in my thinking.

Two years in, the growth hasn’t started, so the remaining 8 years must deliver about 16% a year. I reaffirm the 10 million anyway.

UX work itself is already flowing through the faucet. Back in 2024, luggage maker Away used the AI service Outset to interview 78 customers about hard-sided vs. soft-sided suitcases, and the resulting pages converted 25% better than their predecessors. So user interviews, the most-used research method in UX (and second on my list of most valuable user research methods), were already flowing from a faucet two years ago. The tap has only widened since.

UX work itself is already flowing through the faucet: one luggage maker had an AI service interview 78 customers about hard-sided vs. soft-sided suitcases.

Design even has an outcome meter ready: the A/B test. A design faucet could bill per measured point of conversion lift, provided its work was reliable enough and the buyer specified the outcome in advance. I expect to see such a service before long.

Design even has an outcome meter ready: the A/B test. A design faucet could bill per measured point of conversion lift.

The mechanism behind the forecast is the same elastic demand that powers this article. When the price of design and research drops 10x, the world buys more than 10x the current amount because the backlog of orphan problems in design is deep. What I now see more clearly is where that demand will pool. Little of it will be traditional UI design, because generative UI is absorbing pixel production, and my declining-ROI analysis explains why handcrafted screens stopped being where the money is.

Generative UI is absorbing pixel production, and handcrafted screens have stopped being where the money is.

The growth is in service design at industrial scale: re-blueprinting corporate workflows around AI. Somebody must decide which work flows through which faucet, design the intake and checkpoint rituals, define what “resolved” honestly means, set the escalation and redress paths, and own the outcome metrics that the meters bill against. Tally those decisions across every workflow in every company, and you get millions of person-years of service design.

History says that this redesign work is the bottleneck, which is why it will pay. Edison opened Pearl Street in 1882, yet the economic historian Paul David found that in 1899, electric motors still supplied less than 5% of the mechanical drive in American factories (PDF), and the productivity payoff waited until the 1920s, when factories had been rebuilt around a motor on every machine. (Yes, every AI pundit with a slide deck has pressed the dynamo into service. It keeps getting used because it keeps being right.)

Yes, every AI pundit with a slide deck has pressed the dynamo into service. It keeps getting used because it keeps being right.

The vendors of work faucets are ahead of schedule. Their customers are on factory time. In McKinsey’s 2026 survey, 80% of respondents said AI had made them personally more productive, but only 37% reported any effect on their company’s earnings. Nearly 3/4 of the high performers (a mere 6% of the sample) had redesigned their workflows from the ground up, against 1/4 of everybody else. That gap is why my forecast for the economy runs to 2044, even though the technology won’t need nearly that long to mature.

The vendors of work faucets are ahead of schedule. Their customers are on factory time: 80% of survey respondents said AI had made them personally more productive, but only 37% reported any effect on their company’s earnings.

But the commodification that squeezes pixel roles is the same force funding workflow roles, and the transition rewards one trait above all: the agency to take action. (Job titles will likely drop the “UX” label anyway, given our field’s chronic vocabulary inflation; the work will outlast the badge.) Redesign one real workflow around an outcome-priced service this quarter, at your own company if nobody will pay you for it yet.

Job titles will likely drop the “UX” label anyway. The work will outlast the badge.

Who Checks the Machine When Beginners Are Locked Out?

Look closely at two facts I reported earlier: Crosby’s business model rests on a human lawyer’s final check of every AI contract, and entry-level postings in fields like UX have shriveled to 4.4% of job ads. Put those two facts in the same room, and you have a time bomb under the hybrid service model.

The fuse burns in two stages. Stage one is low-prevalence blindness, and it’s already blowing up in partners’ faces; ask Deloitte’s Australian office. Psychologists have known since Norman Mackworth’s 1948 radar-watch experiments that human vigilance wilts fast when nothing goes wrong, and Jeremy Wolfe and colleagues at Brigham and Women’s Hospital measured the effect in Nature (2-page PDF).

When target anomalies appeared in 50% of screening trials, observers missed 7% of them, whereas dropping the anomaly rate to 1% caused the human miss rate to jump fourfold, to 30%. Such is the cruel arithmetic of a 99%-reliable work faucet: the better the AI gets, the faster the human verifier’s brain shifts into rubber-stamp mode. Advertising “human-verified AI” without redesigning the verification task is liability theater.

The better the AI gets, the faster the human verifier’s brain shifts into rubber-stamp mode: when anomalies drop from 50% to 1% of cases, the human miss rate jumps fourfold.

Stage two is apprenticeship collapse. For centuries, law firms, accounting partnerships, consultancies, and design agencies ran a hidden cross-subsidy: clients paid junior staff to grind through routine NDAs, audits, and usability tests, and those 5,000 easy reps turned a 23-year-old novice into the 35-year-old senior specialist qualified to catch a subtle flaw or handle an angry escalation.

Lisanne Bainbridge warned in 1983 (in her classic paper on the irony of automation) that machines take the easy work and leave humans the gnarly exceptions without the daily practice needed to solve them. Service as a software turns Bainbridge’s cockpit problem into a generational one. Today’s hybrid faucets live off inherited human capital, meaning senior professionals who earned their scars before 2024. Eat 100% of the junior reps for 8 years, and who staffs Crosby’s final check in 2034?

Today’s hybrid faucets live off inherited human capital. Eat 100% of the junior reps for 8 years, and who staffs the final check in 2034?

Aviation and airport security solved both problems, and knowledge firms must steal their playbook:

  • Slip canary cases into live review queues. Transportation security keeps X-ray screeners sharp by projecting synthetic weapons onto clean bags; miss a fake gun, and you get immediate feedback. Wolfe’s lab showed that bursts of high-prevalence practice with instant feedback reset the human quitting threshold. Every licensed faucet and enterprise escalation team should secretly seed a few percent of the jobs routed through its human signers with faults, such as planted liability traps in contracts or fabricated citations in reports. Audit the human auditor, or don’t brag about having one.

Slip canary cases into live review queues. Audit the human auditor, or don’t brag about having one.

  • Replace billable grunt work with flight simulators. No client will ever again pay $300 an hour for a junior to learn on easy work when a faucet charges $2. Professions must move novice training off the client’s invoice and into deliberate simulation gyms: adversarial AI sandboxes packed with broken flows, hostile stakeholders, and edge-case failures where juniors can log 500 compressed reps in months instead of years.

Replace billable grunt work with flight simulators: simulation gyms where juniors can log 500 compressed reps in months instead of years.

How the Faucets Learn

The first edition asked how AI would acquire real service skills, and I speculated that vendors would pay experienced professionals to annotate thousands of hours of usability footage and design decisions for reinforcement learning. Grade: half right, and the half I missed is elegant.

Right: general foundation models weren’t enough. Specialized practice on real cases separates a competent service from a plausible-sounding one. The mechanism that emerged pairs curated procedure libraries (what the AI labs now call skills) with reinforcement learning on verifiable outcomes, rather than my imagined armies of human annotators. What I didn’t predict is that outcome pricing would create the training incentive automatically.

When Zendesk gets paid only for resolutions and Sierra swallows the cost of every failure, each unbilled outcome creates a financial incentive for more tuning; Sierra describes continuous optimization after deployment as part of the offering. The meter doubles as a gradient. Vendors who charge for outcomes are, in effect, learning from the rewards and penalties on their own revenue line, which is why I expect outcome-priced services to improve faster than subscription-priced ones in the same category. And the moat I proposed in 2024 remains largely unclaimed. Whoever assembles the great labeled corpus of design outcomes (which redesigns improved which metrics) will own the compounding advantage in our own field. My challenge from the first edition stands.

The meter doubles as a gradient: vendors who charge for outcomes learn from the rewards and penalties on their own revenue line.

The incentive to improve performance points wherever the meter points, which makes billing for silence doubly toxic. Charles Goodhart’s law (when a measure becomes a target, it ceases to be a good measure) applies to a meter that reads silence as success. An AI tuned on “the customer stopped replying” will discover that the shortest path to revenue runs through customers who give up, and a model that learns to exhaust people politely will look magnificent on the dashboard. The AI labs call this reward hacking and have fought it inside their models for years; outcome pricing moves the fight onto the pricing page. A dishonest meter trains a dishonest faucet. Thus, Guideline 9 is a training-data decision as much as a courtesy.

Goodhart’s law applies to a meter that reads silence as success. A dishonest meter trains a dishonest faucet.

How Smart Buyers Play the Tap: Challenger Routing and Monoculture Risk

Once software loses its seats and its screens, two strategic rules of enterprise buying flip upside down.

First, outcome pricing can weaken single-vendor lock-in. Under seat licensing, buying Zendesk and Intercom at the same time was expensive; paying for two chairs per employee doubled your bill, and retraining 2,000 agents on a new UI was purgatory. When no human touches the screen and you pay per resolved ticket, routing 80% of live traffic to Faucet A and 20% to Faucet B can cost the same per resolution as relying on a single vendor, assuming equal unit prices and no minimum charges. Call this challenger routing, borrowed from how e-commerce companies split card traffic across payment processors. A permanent 20% challenger lane gives the buyer a live comparison. If Vendor A claims an 82% resolution rate via silent timeouts while Vendor B hits 71% with higher repeat purchases and half the angry callbacks, you shift the valve on Tuesday morning without retraining a soul.

Call this challenger routing: a permanent 20% challenger lane gives the buyer a live comparison, and you shift the valve on Tuesday morning without retraining a soul.

Second, shared gradients breed competitive monoculture. A faucet that learns from 3,000 corporate customers optimizes toward the global mean of what works across all of them. For hygiene workflows (payroll, tax filings, password resets, overdue-invoice chasers), convergence is wonderful; nobody wants “artisanal” accounts payable. For differentiation workflows (flagship product UX, pricing strategy, brand positioning, outbound sales), pouring finished work out of the same conversion-metered tap as your 50 closest rivals pushes everybody toward the same bowl of oatmeal. We already watched AI sales-email faucets poison their own well within 18 months because every prospect’s inbox filled with the exact same statistically “optimal” pitch.

Shared gradients breed competitive monoculture: pouring finished work out of the same tap as your 50 closest rivals pushes everybody toward the same bowl of oatmeal.

Tap water belongs in the kitchen sink, not on a Michelin restaurant’s wine list. Pipe the 80% of corporate toil where variance is a bug into shared utility faucets. Guard the 20% where variance is your moat, and use AI there as an instrument under human art direction, never as an autopilot billed on a pooled industry average.

Tap water belongs in the kitchen sink, not on a Michelin restaurant’s wine list. Guard the 20% of work where variance is your moat.

Conclusion: The Tap Is Open

In 2024, this article was a forecast. In 2026, it’s a field report. Professional work is becoming a metered utility, and the guarantee is now written into the price tag. Accountants, payment networks, underwriters, and now the first regulators have filed their paperwork. The risk of paying for a failed job has moved from buyer to seller, and usability has moved with it onto the vendor’s income statement. But the risk of the failure that nobody catches hasn’t moved, and it’s a design problem. What to do about it depends on who you are:

  • If you’re a UX professional: move your center of gravity from screens to service contracts. Pick one workflow at your company, route it through an outcome-priced service, and design the intake, checkpoints, and redress yourself. Then audit the meter: pull 50 cases billed as resolved and count how many of those customers would agree. That project teaches the skills the next decade pays for.

Audit the meter: pull 50 cases billed as resolved and count how many of those customers would agree.

  • If you lead a SaaS product: your pricing page is now a design surface, and picking between “handled” and “resolved” is the most consequential word choice in your company. Decide which outcomes you’re willing to guarantee, meter them honestly, and say so where the order is placed. Don’t outcomewash: every silence you bill today is a dispute you’ll lose to a procurement agent tomorrow.

Your pricing page is now a design surface. The choice between “handled” and “resolved” moves the performance risk from the buyer to you.

  • If you run an agency or consultancy: package your judgment as a product with a meter on it, because the market for unmetered billable hours is closing. Sell the work, keep the accountability, and let AI carry the volume.

Package your judgment as a product with a meter on it, because the market for unmetered billable hours is closing.

  • If you buy services: read the meter’s definition before the price. Ask every vendor three questions: What counts as resolved? What share of billed outcomes did customers confirm? Who besides the vendor has checked the count? And keep the consequence risk in view when you sign, because the warranty covers the fee and nothing else. Your own liability insurer may no longer cover it either, so read that policy, too.

The warranty covers the fee and nothing else. Your own liability insurer may no longer cover AI failures either, so read that policy, too.

What would you buy first from a work faucet, and what would it take for you to trust the meter? Tell me in the comments. Your answers are the user research that this article argues should be cheap.

And since grading my own predictions is now a house tradition, here’s the ledger for future editions. Bookmark it and grade me; I’ll do it publicly either way, as I did above with the 2024 edition.

8 Predictions for Service as Software

  1. By the end of 2027, at least 15% of software companies will name outcome-based pricing as their primary model, up from 5% in 2025, as measured by Kyle Poyar’s annual monetization survey or its nearest equivalent. (Confidence: 65%.)

  2. Before 2028 arrives, at least one of the three big support-agent vendors (Fin, Zendesk, or HubSpot) will stop billing silence at full price, either by charging only for resolutions that the customer confirms or by discounting the assumed ones. (Confidence: 60%.)

  3. By year-end 2028, at least 3 companies will each report more than $1 billion a year in revenue billed per outcome, as opposed to per seat, per token, or per action. Sierra, at $400 million, is the closest example discussed here. (Confidence: 55%.)

  4. By the end of 2029, at least 3 of the 10 largest agent vendors will publish outcome rates (resolution, dispute, and punt rates) attested by an independent auditor, much as SOC 2 reports attest to security controls today. (Confidence: 65%.)

Grading my own predictions is now a house tradition, so each one comes with a confidence level. Bookmark the ledger and grade me.

  1. By 2030, more than half of AI-native vendors will report that customers pay them out of services budgets, up from 35% in Kyle Poyar’s 2026 survey, where 15% already drew on headcount budgets. (Confidence: 65%.)

  2. By 2030, at least 10% of enterprise SaaS spending, upward of $100 billion a year, will be billed per outcome. Gartner predicts that 40% will have shifted to usage-, agent-, or outcome-based pricing by then; I’m betting on the strictest of the three. (Confidence: 50%.)

  3. By 2034, the world will have at least 10 million UX professionals (whether or not they’re called that), and more of them will design services and workflows than screens. Interim check: 5 million by 2030. (Confidence: 55%.)

  4. By 2044, service as a software will contribute on the order of $50 trillion annually to the world economy, in 2024 dollars. Interim check: $1 trillion a year by 2030. (The original bet, unchanged.)

The first SaaS era handed you a wrench and wished you luck with your pipes. The new one fixes the leak and charges you only if the dripping stops. Gas needed 44 years to get from its first meter to a meter with an inspector’s stamp. I give the work faucet 3 years. The tap is open. Go design the meter.

The tap is open. Go design the meter.

原作者保留版权;来源:Jakob Nielsen · UX Tigers。本站维护者确认获准转载,非公共领域或开放许可证声明。 原文来源(转载授权由站点维护者确认) · 中文译文由 CxHub 客户体验社区生成,并非官方译文。

来源:Jakob Nielsen · UX Tigers · uxtigers.com