Check out our upcoming events | Learn More

Capacity, not contract: how to segment AI customers

Why software companies selling AI should segment customers by their capacity to realize value, not by the size of the contract.

  • Bhavishya Joga
  • October 22, 2026
  • 8 Minute Read
Back to All Insights

KEY TAKEAWAYS

  • Segment customers by their capacity to realize value. Four observable factors predict whether value appears: data condition, the clearance path through risk, tolerance for workflow change, and operator fluency. Contract size predicts none of them.

  • Proof is now a stage, and it needs an owner. Between onboarding and adoption, the customer has to show the system works on their own data at a threshold their risk function accepts. Name an owner and make the evaluation set the deliverable.

  • Measure value per unit consumed. Consumption pricing rewards usage, but usage without an outcome is cost on the customer’s monthly invoice. The vendor willing to cut spend when the ratio is wrong earns the expansion.

For two decades, customer success ran on a proxy: the size of the contract predicts how much help a customer needs.

The proxy worked because software was deterministic. A customer bought seats, configured a workflow, trained users, and then either logged in or did not. Implementation was the main variable, so journeys followed the implementation calendar and segments followed annual contract value. Large accounts received a dedicated account representative, while small accounts were grouped under a shared customer success representative.

AI products retire the proxy, and they do it three ways at once. Value is probabilistic rather than guaranteed, so it has to be demonstrated on the customer’s own data before anyone will rely on it. Pricing is increasingly consumption-based, so revenue moves weekly instead of at renewal. And the unit of adoption is a use case rather than a user, which means one account holds a portfolio of separate adoption curves sitting at different stages.

Exhibit 1

None of this is a reason to abandon the craft. It is a reason to re-cut its two foundations: how customers are grouped, and what the journey between them actually contains.

We advise software leaders to segment customers by their capacity to realize value, not by the revenue they represent. Capacity is observable and it has four parts: the condition of the customer’s data, the clearance path through their risk function, their tolerance for changing a workflow, and the fluency of the people who will operate the system. Those four predict whether value appears. Contract size does not.

Three consequences follow. Conventional segmentation misallocates the scarcest resource an AI software company has. The journey now contains a proof stage nobody owns. And consumption pricing rewrites the customer success mandate in a way that has to be managed rather than celebrated.

The lens throughout is the vendor’s. At each turn we note what the shift means for the enterprise buyer on the other side of the table, because the same analysis tells a CIO what to demand.

Value used to arrive on a schedule. Now it has to be earned in front of the customer, on their data, before anyone will depend on it.

Contract size tells you what a customer pays, not what they can absorb

Tiering by annual contract value assumes that larger customers face harder problems and therefore need more help. However, with AI products, contract size says little about how much help a customer needs.

Census Bureau data puts the point plainly. As of early October 2026, 36.5% of firms with at least 250 employees reported using AI in the previous two weeks, essentially level with the 36.4% of firms with 100 to 249 employees, and against 23.9% of firms with four or fewer.1 Even at the top of the market, nearly two-thirds of large firms were not using it. A vendor’s enterprise tier is therefore populated largely by organizations that are considered to be trailblazers, irrespective of what their contract states.

The cost shows up in the coverage model. Solutions engineering is the scarcest resource an AI software company has, and value-based tiering routes it toward the accounts that pay the most rather than the accounts where three weeks of expert attention would move a use case into production. Meanwhile a mid-market customer with clean data, a permissive risk posture, and an engineer who already writes evaluation sets receives a quarterly check-in and a newsletter.

If you are the buyer Ask a prospective vendor how they decide who receives solutions engineering time. If the answer is a function of your contract value, you are being priced into a queue rather than matched to a need.

The Metis view: segment on capacity to realize value, then staff to the quadrant

Two axes carry more predictive weight than revenue. Capacity is the customer’s ability to convert a working model into a working process: data condition, clearance path, change tolerance, and operator fluency. Headroom is the number of credible use cases still ahead of them.

Exhibit 2

Note what the grid does to a conventional book of business: neither axis tracks revenue. A global bank with a dozen candidate use cases and a nine-month model validation queue is a Lighthouse. A 300-person AI-native firm running the same product, with clean data and engineers who write their own evaluation sets, is a Compounder at a fraction of the contract value. Reading those two in reverse is the expensive mistake, and logo size is what causes it. The quadrant should set the coverage model.

The stages are the same in every industry. Where customers get stuck varies by sector. Design the journey once, then staff it differently by sector.

Six stages, and a loop that never fully closes

Exhibit 3

Three of these six stages are new, or newly heavy enough to need an owner. Validate asks whether the system performs on this customer’s data rather than on a benchmark. Clear asks whether the people accountable for risk will put their name to it. Sustain asks whether accuracy and unit cost hold once real volume arrives.

Conventional onboarding runs through three of them, Qualify, Deploy and Expand, and treats the rest as support tickets. That is why so many programs look healthy in the CRM and stall in reality.

The inner loop matters as much as the full lifecycle. Once an account is live it cycles between clearance, deployment, and sustained operation without returning to qualification. Models change, data shifts, and a workflow redesign upstream can invalidate an accuracy assumption that held last quarter.

Renewal risk in AI products usually starts as an unnoticed trip backward around the inner loop.

Proof is now a customer success deliverable

Between onboarding and adoption sits a stage that deterministic software never required. The customer has to establish that the system works on their data, for their edge cases, at a threshold their risk function will accept. In most vendors nobody owns this. Product owns the model, professional services owns the integration, and the customer success manager owns the relationship. Proof falls between them, which is why pilots stall in a place the health score cannot see.

The demand for proof is not a passing phase of the hype cycle. In Stack Overflow’s 2025 developer survey, 46% of developers reported distrusting the accuracy of AI output against 33% who trust it, with only 3.1% expressing high trust.2 Adoption climbed regardless, with 84% using or planning to use AI tools.2 The most cited frustration was output that is almost right but not quite, named by 66% of respondents.2 Usage is rising while trust falls. That gap is the work.

In practice this means the account team ships evaluation sets, accuracy baselines drawn from the customer’s own labelled examples, and an agreed threshold for human review. It also means serving three people rather than one. The builder wants latency and control. The outcome owner wants a number that moves. The risk owner wants evidence and an audit trail. Journey maps built around a single primary contact fail on the third persona, and the third persona is the one who can stop a deployment.

Exhibit 4

Pilots stall when nobody is accountable for proving the model works.

If you are the buyer Make the evaluation set a contractual artifact. Agree the test cases, the accuracy threshold, and who signs off before the pilot starts, not after it disappoints.

The same journey, a different place to get stuck

Industry decides which stage binds. This is the difference between a cross-industry playbook that is decoration and one that changes how a team is staffed.

Exhibit 5

Where clearance binds, in financial services and the public sector, the customer success team is an evidence producer. Its output is model documentation, validation artifacts, and a defensible audit trail. Enthusiasm is least credible here.

Where proof binds, in healthcare and life sciences, the work is accuracy on the customer’s own population rather than a published benchmark. Expect to run the evaluation twice.

Where data binds, in manufacturing and industrial, nothing ships until the data question is answered. The first competent conversation is about historians, unstructured documents, and operational systems, not about the product.

Where humans bind, in retail and in professional services, the constraint is a calendar and a compensation model. A merchandising team has a window. A billing-hour practitioner has a reason to resist.

What the bottleneck changes about who you hire

  • Evidence producer (financial services, public sector): risk-literate architects who write documentation a second line of defense will accept.
  • Accuracy prover (healthcare, life sciences): measurement specialists who can design an eval on a customer population.
  • Data readiness partner (manufacturing, industrial): field data engineers comfortable with historians and unstructured records.
  • Change agent (retail, legal, professional services): enablement leads who can redesign a workflow and win over the practitioner in it.
  • Cost engineer (software, digital natives): analysts who hold unit economics steady as volume and model choice change.

A good customer success manager on a consumption contract will sometimes tell a customer to use less. If that sounds like a firing offense, the operating model is not ready.

The consumption paradox, and the metrics that resolve it

Consumption pricing works. Snowflake closed fiscal 2026 with a net revenue retention rate of 125% on fourth-quarter product revenue of $1.23 billion, up 30% year over year.3 Revenue of that shape grows when customers use more, which rewrites the customer success mandate from preventing churn to increasing usage.

The rewrite carries a trap. Usage that does not produce an outcome is simply cost, and the customer meets that cost on a monthly invoice rather than at renewal. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.4 Two of those three causes are visible from inside the vendor long before the customer acts on them.

The discipline is to measure value per unit consumed rather than consumption growth, and to be willing to cut a customer’s spend when the ratio is wrong. In our experience that conversation buys more expansion than it costs, because it is the one moment a vendor can prove it is optimizing for the customer’s outcome rather than its own quarter.

Exhibit 6

If you are the buyer Ask your vendor to report value per unit consumed alongside the invoice. A partner who can produce that number is managing your outcome. A partner who cannot is managing your bill.

Re-segmenting a customer base takes a quarter of work. The sequence we have seen hold is unglamorous, and it starts with arithmetic rather than reorganization.

  1. Score, then re-tier. Rate every account on data condition, clearance path, change tolerance, and operator fluency. Re-cut the segments on capacity and headroom, not revenue. The scoring usually settles the argument on its own, because the Lighthouses become visible to everyone at once.
  2. Name an owner for proof. One accountable role, with the evaluation set and the accuracy threshold as its deliverables. Fund it before the next pilot starts. Most companies can staff it from people already on payroll, starting with the solutions architect who keeps getting pulled into pilots.
  3. Change one metric in the executive review. Replace consumption growth with value per unit consumed, and watch how quickly the rest of the conversation reorganizes itself.

If you are the buyer The same three moves work in reverse. Score your own capacity honestly before the next renewal, and you will negotiate for the help you actually need rather than the tier you happen to sit in.

The first two decisions are the difficult ones. The third is the one that makes them stick. Customer success has always been the discipline of turning a purchase into a result. AI has moved the hard part earlier, into the stretch of the journey where value has to be proven before anyone will trust it.

Notes

  1. U.S. Census Bureau, "AI use by employment size class, September 21 to October 4, 2026", Business Trends and Outlook Survey, 2026. ↩
  2. Stack Overflow, "2025 Developer Survey: AI", Stack Overflow Developer Survey, 2025. ↩
  3. Snowflake Inc., "Snowflake Reports Financial Results for the Fourth Quarter and Full-Year of Fiscal 2026", Snowflake Newsroom, 2026. ↩
  4. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", Gartner Newsroom, 2025. ↩