AI agent outcome pricing
who verifies it

On 23 September 2026, TechCrunch reported that Ema raised a $77M Series B. The least interesting number in the story is the amount. The customer list is the interesting part: PwC, KPMG, NTT DATA, Wipro, Hitachi. Professional services and IT services firms appear on the customer list of a third-party agent platform. In the same report, Ema described how it charges: not by software seats, and not by the number of AI tokens a customer consumes, but tied to the completion of tasks and business outcomes. Outcome pricing is no longer the differentiator. The differentiator moved.

1. Why services firms are on Ema’s customer list

Start with the facts. Ema is headquartered in Mountain View, with offices in Bengaluru, London and Vancouver, and about 200 employees. Its founders are Surojit Chatterjee (previously Google and Coinbase) and Souvik Sen (previously Okta).

The $77M Series B was led by Creaegis, with Accel, Section 32 and Prosus participating. Total funding reaches $140M. The valuation is more than 4x the 2024 round; the company declined to disclose an absolute figure. The round is all primary equity, with no debt and no secondary sales.

Adoption figures are also company-reported: more than 50 active enterprise contracts, more than 1M active enterprise users, more than 5M actions and queries, 50x revenue growth over two years, bookings above $150M — note the definition, which the company describes as the total value of multi-year contracts over two to three years, not ARR — net dollar retention around 180%, more than 90% of customers expanding beyond their initial use case, and gross margins near 80%. Every one of those numbers is self-reported. TechCrunch did not independently verify them.

The customer list is the part worth recording. PwC, KPMG, NTT DATA, Wipro and Hitachi appear on it. At least four professional services and IT services firms are buyers of this third-party agent platform.

One discipline is required here: this is not consultancies abandoning their own builds. It is layering. The same buyers run two tracks. They build client-facing delivery capability themselves — the recorded in-house moves at Accenture, BCG, McKinsey, EY and Genpact still stand — and they buy third-party platforms to fill gaps in their own product coverage. Both facts hold at once.

That changes how you position to a services firm. You cannot say you replace their in-house build. You can only say: when you deliver the result to your end client, who independently verifies this particular run.

2. The billing axis changed; the differentiator did not

Ema’s own wording on pricing is specific: it does not charge customers based on software seats or the number of AI tokens they consume, and its pricing is tied to the completion of tasks and business outcomes.

That is a third billing axis, beyond seats and tokens, and it has been validated by both capital and adoption: a $77M Series B, a valuation above 4x the 2024 round, more than 50 active enterprise contracts and more than 1M active enterprise users.

The conclusion runs the other way. Once the default billing model on third-party platforms is completion, outcome pricing stops being a reason to choose a vendor. It becomes the entry ticket. Buyers do not award points for charging by outcome — it is the baseline, not the bonus.

The gap therefore moves onto the word complete. Once payment hangs on completion, the contract has to state what counts as complete and who certifies it. Today a buyer can obtain exactly two readings: the vendor’s self-reported completion rate, or the buyer’s own assessment against its business definition. Neither is independent evidence.

This is not a new topic, but it has changed category. It used to be a governance problem. It is now a pricing feasibility problem — because of the next set of numbers.

3. Same goal, 30x the cost

According to Business Insider, McKinsey’s research finds that the same goal completed by different agents can cost 30 times as much. McKinsey senior partner Lari Hämäläinen puts it plainly: every day there is a 30x difference in the cost.

The same report carries another set of figures: agentic AI can cut human time on transformation-office work by 35% to 40%, and sometimes by more than 70%, but enterprises have to measure value in order to justify the cost.

The 30x figure upgrades who verifies from a compliance question into a pricing feasibility question. The logic is short: when the denominator is not comparable, an outcome-priced clause cannot be drafted.

The same word, complete, costs one unit on one execution path and thirty on another. If the contract says price per completed task, the two sides are negotiating against different denominators. The buyer is asking what the outcome is worth; the seller is asking what this run actually cost. Across a 30x band, those two questions can sit an order of magnitude apart.

Drafting that contract requires per-task evidence first: which path this run took, what it cost, and who independently certifies it. Miss any one of the three and the price never reaches paper.

On sourcing: 30x is McKinsey’s research finding, not our measurement. The first-party pages on mckinsey.com were unreachable inside this research window, so we cite the figures as reported by Business Insider and keep the first-party URLs for verification.

4. Budgets tightening, spending still rising

The third cluster of facts comes from McKinsey’s State of AI 2026, again as reported by Business Insider. Roughly one in five organisations say AI spending has already started constraining their operating costs. About one third report that AI spend has passed 10% of their technology or telecom budget. At the same time, 60% plan to spend more next year.

A fourth figure has a different scope and has to be recorded separately: 32% of surveyed organisations gave up at least one software purchase because they could build it themselves. The scope is at least one product or feature, at least once — not one third of enterprises switching to building in-house.

Those two thirds are different thirds. One measures AI spend crossing 10% of a technology or telecom budget. The other measures whether a purchase was ever abandoned. Merging them produces a conclusion that no survey question actually supports.

Read the budget constraint and the continued increase together and the direction of procurement becomes clear. The money keeps rising, but the signature gets more careful and more evidence-oriented. A CFO is not asking how many tokens the agent burned. A CFO is asking what the money bought, and whether what it bought can be checked by a third party. That points at the same gap as the 30x cost spread.

5. Framework one: the three preconditions for outcome pricing

Put the three clusters together and a reusable test falls out. For outcome pricing to actually hold, three preconditions have to hold at the same time: attributable, reproducible and verifiable.

Precondition one, attributable. The result traces to one specific execution, with its steps, its tool calls and its model version. Not an aggregate claim that the system improved by some percentage. Aggregates cannot serve as a payment basis in procurement, because they map to no individual item anyone can dispute.

Precondition two, reproducible. Re-running the same goal on the same input lands the execution path and the cost on the same order of magnitude. If a 30x spread cannot be explained, pricing becomes a lottery: what the buyer pays depends on which path the agent happened to take, and the buyer cannot see the path.

Precondition three, verifiable. Someone other than the payee decides that completion happened. This is the precondition most often skipped, because the first two look like engineering problems and this one looks like a trust problem. It is not.

The three depend on each other: attributable is the precondition for reproducible, and reproducible is the precondition for verifiable. Drop any one and outcome pricing degrades back into seats plus a discount story.

The test is directly usable. Take an outcome-priced contract currently under negotiation and ask three questions of each clause: can the definition of complete here be traced to one specific execution? Can that same path be re-run? Who holds the authority to certify it? If you cannot answer two of the three, the billing axis in that contract is decoration.

6. Framework two: three layers of verification authority

The second half of the framework is who holds verification authority. There are three layers, and the current distribution is very uneven.

Layer one, vendor self-report. This is today’s default: the completion rate is generated unilaterally by the party being paid. Its structural defect is not dishonesty, it is that it cannot falsify itself — the denominator is defined by the vendor, the certification standard is held by the vendor, and when something goes wrong the vendor is the only source of information.

Layer two, buyer self-assessment. The buyer checks whether the result was right against its own business definition. That is a step up from self-report, but it still is not independent evidence. The buyer can see whether the result was right; it cannot see the execution path, and therefore cannot distinguish an agent that took a path three times more expensive from one that took the shortest path. Pricing disputes live precisely in that difference.

Layer three, independent evidence. A per-run runtime record persisted to disk: path, tool calls, token consumption, elapsed time, failures and retries. The format is readable by both sides, which is what lets both sides accept it as a shared denominator. Evidence at this layer belongs to neither party.

Today the overwhelming majority of outcome-priced contracts sit at layer one, with a minority at layer two. Layer three is the gap.

One clarification: layer three is not a new line of business for auditors, it is an engineering problem. It asks whether the path was recorded during execution, whether the recorded format can be read by both buyer and seller, and whether the system emits unverified or a default value when the record is missing.

7. Contract level: incomparable denominators cannot be drafted

A standard outcome-priced contract has four steps: define complete, agree the pricing unit, agree the certification method, agree dispute resolution. The gap opens at steps one and two.

Step one stalls immediately. Take a concrete case: an enterprise buys completed invoice processing, priced per item. What counts as complete? Extracting fields by OCR, or landing in the ERP with reconciliation passed? The first can be self-certified by the agent. The second requires a reading from a downstream system. The unit price between those two definitions can differ several times over, and the buyer cannot get the second definition onto the negotiating table.

Step two stalls on the denominator. When the same goal can cost 30x, price per completed task is an incomparable denominator on the buyer’s side. Did we complete ten thousand items this month because the agent got more efficient, or because it took a more expensive path? Without a path record, there is no answer.

The buyer is left with two options: accept the vendor’s self-reported completion rate, or push the pricing back to seats or tokens — back to a denominator it can see.

That is the commercial meaning of the gap. For outcome pricing to hold on the buyer’s side, an independent verification layer has to exist first. Without it, outcome pricing degrades back into usage pricing inside the procurement process — not because buyers dislike paying for outcomes, but because buyers cannot verify them.

8. Where this judgement stops applying

The argument above has firm boundaries. Past them, it does not hold.

1. When the task has one path and an enumerable result. Filing 200 contracts into three categories, for instance. There is almost nothing to dispute about completion, vendor self-report is sufficient, and a verification layer only adds cost.
2. When cost does not vary with path. Fixed price, subscription and in-house builds produce no 30x spread. The denominator is comparable and a verification layer adds nothing.
3. When the buyer is the executor. In-house agents and internal delivery have no vendor-buyer separation of verification authority. What remains is an internal audit question.
4. When a single deal is far smaller than the cost of verification. If the marginal cost of evidence exceeds the amount in dispute, sampling beats full logging.
5. Where regulation already mandates an audit trail. Verification authority is already assigned by law. The remaining work is format standardisation, not ownership.

There is also a priority rule: if the agent runs reversible, low-cost internal work, the verification layer ranks below cost governance. Know what it cost first, then argue about whether to pay.

9. The Asia-Pacific timing window

Ema states it plans to expand into new markets over the next year, particularly across Asia-Pacific, along with South America and parts of the Middle East. That fixes an observation window: twelve months from 23 September 2026, so before 23 September 2027.

The highest-probability landing path runs through existing customers and the company’s own base: NTT DATA in Japan, Wipro in India, and Ema’s headquarters in Bengaluru. Stack those and the first named Asia-Pacific move most likely lands in that band.

The trigger conditions are worth writing down. A first named Asia-Pacific customer, a local office, or local pricing — any one of them appearing means this judgement should be re-assessed.

Why this matters for buyers in China: pricing and contract shapes in Asia-Pacific tend to follow head-office templates first, and the definition of complete plus the certification method travel with those templates. The verification gap arrives in local negotiations attached to the contract template.

10. Our judgement

Three sentences. First, outcome pricing is no longer the differentiator. Second, the differentiator is which path this particular execution took, what it cost, and who independently certifies it. Third, verification authority is where the gap currently sits.

For buyers, one executable action: in the next negotiation, split the definition of complete and the authority to certify it into two separate clauses, and require the execution path record as an attachment to payment.

For sellers: be able to evidence per task before you sell per outcome. Without that capability, outcome pricing moves risk back onto the seller, because when a dispute starts the only information source is the seller itself — and self-reported data from the party being paid is not what disputes get settled on.

References: Ema’s funding, customer list and pricing language come from TechCrunch, 23 September 2026 (techcrunch.com/2026/09/23/ema-raises-77m-as-ai-starts-eating-into-enterprise-software-and-services), verified in full text in this window. All Ema operating figures are company-reported and were not independently verified by TechCrunch; bookings means the total value of two to three year multi-year contracts and is not ARR; the valuation is disclosed only as more than 4x the 2024 round, with the absolute figure withheld, and we do not estimate it. McKinsey’s 30x cost spread, the 35% to 40% and 70%-plus reductions in human time, and the State of AI 2026 budget figures are cited as reported by Business Insider (businessinsider.com/mckinsey-ai-agent-costs-variable-expensive-2026-9); first-party pages retained for verification: mckinsey.com/capabilities/quantumblack/our-insights/is-that-ai-agent-worth-it-agentic-economics-and-the-modern-operating-model and mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai (mckinsey.com was unreachable inside this research window and was not independently verified). State of AI 2026 methodology as reported: n=1,719 across 97 countries, fielded 2026-05-04 to 2026-06-08, secondary methodology check at digitalapplied.com/blog/a-third-of-companies-skipped-buying-software-and-built-it. The 30x figure is McKinsey’s research scope, not our measurement. The Asia-Pacific window and trigger conditions are our observation framework built on TechCrunch’s wording, not a company commitment.