AI Pricing in 2027: Buyers Will Pay Only for Completed Work
A 2027 forecast predicts AI buyers will pay only for completed work: narrower billable milestones, invoice-level traceability, full-workflow economics and hybrid pricing across Fin, Salesforce, Zendesk and Sierra.

Why it matters
- Zendesk introduced automated resolution tiers in May 2026; Sierra described its outcome pricing model in 2024.
- The forecast predicts vendors will add newly billable outcome categories in sales, IT service and back-office work in 2027.
- Intercom's March 2026 explanation argues an agent can perform useful billable work before human confirmation.
- METR's February 2026 update complicated its early-2025 finding that experienced developers were slower with AI.
- Hybrid pricing — outcome fees alongside seats and platform charges — is predicted to remain common at the end of 2027.
By 2027, buyers of AI agent products will pay for completed work rather than interactions, and that shift will force vendors to define billable outcomes narrowly enough to attribute, verify and dispute, according to a new industry forecast tracking Fin/Intercom, Salesforce, Zendesk and Sierra.
The forecast, published on Unite.AI, argues that outcome pricing will expand in 2027 across four vectors: newly billable work categories, inspectable billing definitions, full-workflow economics and continued hybrid pricing. The stakes are commercial and structural at once: enterprise software budgets are moving from seats to results, and the vendors that can prove a milestone was completed will capture that spending.
What does outcome pricing already look like?
Outcome pricing is not new. Sierra described its model in 2024. Salesforce offers several Agentforce purchasing models, including actions, conversations, licenses and help-agent resolutions. Zendesk introduced automated resolution tiers in May 2026.
These are materially different units, even when each vendor presents it as a way to align spending with value. The article's tracked cohort — Fin/Intercom, Salesforce, Zendesk and Sierra — puts the same purchasing questions in different forms: what work is finished, what the buyer can verify and which costs remain around the result.
The premise is blunt. A buyer does not wake up wanting more AI interactions. They want the service request resolved, the qualified lead handed over or the accounts reconciled. Yet those ordinary outcomes contain awkward questions. When is the job finished? Who decides? What happens when it reopens tomorrow?
Will vendors expand what counts as a billable outcome?
The forecast's first prediction: in 2027, vendors in the cohort will add newly billable categories in sales qualification, internal IT service and back-office reconciliation, because each has recognizable milestones. A message sent or a tool called will not be enough; the offering will have to name the completed milestone for which the customer pays.
The underlying product problem is to find a unit small enough to attribute and valuable enough to price. "Improve customer relationships" is too broad to settle an individual charge. "Validate eligibility and route a qualified prospect according to these rules" is narrower, with an observable completion state. That precision also exposes the cases where a superficially successful action has little business value.
Fin offers a revealing baseline. Its current outcomes documentation distinguishes resolutions, configured Procedure handoffs, qualifications and disqualifications. A configured handoff can be billable; an ordinary escalation triggered by a request for a human is treated differently. Therefore, neither "outcome" nor "handoff" can safely be translated into "the customer's whole problem was solved."
In its March 2026 explanation of the change, Intercom argues that an agent can perform useful work before a human confirmation step. That is a coherent product position. It also means buyers need to specify whether they are purchasing autonomous resolution or a completed subtask in a larger process — and that distinction should survive into the reporting interface and the contract.
The forecast follows the incentive that distinction creates. A supplier that can reliably complete one bounded milestone can enter workflows where the final decision remains human. Expansion will depend on attribution: a call that is booked, a request that is fulfilled, or a reconciliation exception that is correctly classified must correspond to an agreed event. Products will struggle where the event can be counted more easily than its value can be established.
For 2027, the article tells product leaders to watch pricing pages that define newly billable milestones in sales, employee service and back-office work. The important detail is the event that triggers payment, particularly when an agent completes only part of a larger process. Expanding the word "outcome" without clarifying what work the buyer receives would weaken the commercial argument.
Will billing become an inspectable product feature?
The second prediction concerns disputes. Sooner or later, a customer will open an AI invoice and ask why a particular case appears on it. That moment makes verification part of the product. The forecast expects more of the cohort's suppliers to let buyers inspect individual billable events, see how they were classified and challenge a mistake.
A result meter, the article notes, is also a system of judgment. It may infer that a problem was resolved because the customer stopped replying, apply a classifier, or record that a configured sequence ended at a handoff. Each method can be useful. Each can also diverge from the buyer's understanding of a completed job.
The differences are already visible in documentation. Zendesk's resolution-tier description says its automated-resolution unit concerns requests resolved without escalation to a human. Fin's pricing FAQ includes configured handoffs among its outcome types. Comparing the headline price of these units without comparing their definitions would obscure what is being purchased.
Customers already have some visibility. Fin's Procedure reporting distinguishes handoff outcomes, while usage controls expose limits and the complications of conversations whose state changes over time. The forecast is for further verification capability, not the invention of billing dashboards.
The next useful step is an event record that connects the invoice to the work: which customer request was handled, what event triggered a charge, which policy classified it, and what happens if the result is later corrected. A buyer should be able to distinguish an incorrect classification from a legitimate charge for a narrower contracted milestone. Those are different disagreements and require different remedies.
There is a commercial reason to expect progress. A supplier can make a strong outcome promise only if it can resolve disputes without re-investigating every conversation manually. Sierra's 2026 discussion of outcome pricing explicitly acknowledges the operational and contractual complexity of attribution — evidence, the article argues, of the work needed behind the pricing model, not that the industry has settled on a common definition.
The practical advice: procurement, finance and the team operating the agent should agree on a shared definition of completion before scaling, inspect a sample of billed events, and establish how disputed classifications will be reviewed.
Will better evidence expose the whole workflow's economics?
The third prediction is the hardest, the article concedes, because public measurement is often thinner than the product claim. An assistant can make the first step faster and still leave the organization with more work. Escalation, checking and reopened cases can consume the apparent saving.
A high automation rate does not reveal whether the remaining cases became harder. A lower cost per conversation can coexist with more conversations per resolved problem. Faster initial handling may be offset by later correction or repeat contact.
Two research results frame the problem. The NBER working-paper version of Generative AI at Work studied a support-assistant deployment and reported higher issues resolved per hour, with substantial variation across worker experience. That concerns human workers using assistance, not a direct estimate of the economics of today's autonomous agents. Its value is methodological: the choice to measure resolved work and examine heterogeneous effects.
There is also a warning against treating perceived speed as measured value. METR's early-2025 developer study found a slowdown in its particular experienced-developer setting. Its February 2026 update explains why later measurements became difficult to interpret because of selection, task choice and changing work patterns. The later update does not establish that every developer is slower — or that the earlier result remains a current universal estimate.
The product lesson is to measure the complete operating loop. A useful cost numerator includes agent fees, relevant platform and infrastructure expense, human review, escalations and rework. The denominator should count cases that meet a stated completion standard over a stated observation window. For a time measure, all human handling associated with those cases belongs in scope, including the correction queue.
That operating view is consistent with Measuring Agents in Production, whose revised 2026 study emphasizes controllable systems and reliability challenges. τ²-Bench supplies a complementary research reminder: completing a support task can depend on coordinating with a user who also changes the environment. A fluent answer and a completed workflow are different evaluation targets.
For 2027, the article tells readers to watch for customer reports that follow cases through escalation, reopening and human correction, rather than stopping at the first automated response. Higher automation rates accompanied by more repeat contacts or heavier exception handling would challenge the forecast's economic premise.
Will outcome billing replace subscriptions?
No — and the forecast is confident on this point. Hybrid bills will remain common within the cohort at the end of 2027: a charge for results alongside seats, platform access or consumption.
The existing baseline makes this plausible. Intercom's pricing distinguishes helpdesk seats and Fin usage; its standalone Fin route has a different structure, so the seat component should not be attributed to every Fin buyer. Salesforce's pricing page exposes multiple purchasing units and configurations. Zendesk's plans combine service-platform packaging with automated-resolution allowances and additional capacity.
The strategic consequence: a cheaper advertised outcome may not produce a cheaper completed case. A lower unit price can sit beside a higher platform commitment or more human handling. A higher unit price can be attractive if it removes enough downstream work. The correct comparison depends on the full workflow and an explicit accounting boundary.
The forecast's high confidence rests on coexistence, not on any prediction that a particular vendor will raise prices. The competing incentive is simplification: a supplier could absorb more supporting costs into its outcome unit to reduce procurement friction. If enough vendors do that, this forecast will fail even if outcome pricing itself becomes more popular.
What should buyers do now?
The article's closing argument is definitional. The strongest AI products will make the buyer's definition of success visible inside the service — including the awkward cases, the correction process and the work people still perform.
A completed outcome can be a powerful commercial unit. It earns that role when both sides can recognize it without needing the sales team to explain the invoice. In 2027, the vendors that make that definition inspectable — event by event, on the invoice — will hold the commercial advantage.
Source: unite.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
718 articles
Related articles
- Zendesk kills per-seat pricing with Outcome AI Platform launch
- Mondrio Raises $3.75 Million to Link AI Pricing and Sales
- AI coding costs are on pace to top developer salaries by 2028
- Zendesk scraps ticket-count thinking with outcome-based AI platform
- Novo launches AI banking tools with approval-based controls