The AI Agent Procurement Playbook: Contracts, SLAs and Pricing for 2026

Why procurement, not technology, decides who wins

The most quoted number in enterprise AI right now is a failure rate. MIT researchers found that about 95% of enterprise AI pilots produced no measurable profit-and-loss impact. S&P Global reported that 42% of companies abandoned most of their AI projects in 2025. Gartner has warned that more than 40% of agentic AI projects are at risk of cancellation by 2027, and Forrester's root-cause work attributes the bulk of negative outcomes to three preventable causes: unclear success criteria, insufficient tool or data access, and evaluation coverage that drifts over time.

Read that list again, because none of it is about model quality. Every item is a decision you make during procurement. You define success criteria in the contract. You grant tool and data access in the onboarding. You commit to ongoing evaluation in the service level agreement. By the time a pilot is technically live, the outcome has largely been decided by terms that were negotiated, or quietly skipped, months earlier. That is the core argument of this guide: the biggest lever a buyer has over AI agent success is not which vendor they pick, but how they buy.

The three pricing models you will be quoted

In 2026 almost every major agent vendor has converged on the same menu, often offering all of it at once. Salesforce, for example, now markets Agentforce under several pricing models simultaneously. Understanding the trade-offs of each is the first job of any buyer, because the model you choose quietly shapes your risk.

The outcome model is the headline story of the year. Intercom's Fin agent bills roughly $0.99 per billable resolution, at most once per conversation, on top of your existing help-desk seats. Salesforce launched its Agentforce Help Agent with pay-per-resolution pricing at about $2 per resolution, pricing the same outcome at roughly twice Fin's rate, and then moved to acquire Fin for around $3.6 billion in mid-2026. Sierra runs a purer version still, getting paid only when its agent resolves an issue without human intervention, a model that helped carry it past $150 million in annual recurring revenue by early 2026. When three of the loudest vendors in the category all anchor on the resolved outcome, buyers should expect to negotiate on that ground.

The catch inside outcome-based pricing

Outcome pricing sounds like the buyer's dream: pay only for value delivered. The catch is that the vendor usually writes the definition of the outcome, and that definition is where your bill is really set. A "resolved conversation" can mean the customer explicitly confirmed their problem was solved, or it can mean the conversation simply ended without a human agent stepping in, which counts a silent abandonment as a win. Those two definitions can differ by a wide margin on the same volume of traffic.

Before you accept an outcome price, get precise answers to a short list of questions and get them in writing:

• What exactly counts as a billable outcome? Ask for the literal rule, not the marketing phrase.
• What does not count? Confirm that escalations, deflections to a human, and failed attempts are excluded.
• Is there a dedupe window? Fin bills at most once per conversation; confirm your vendor does the same so a follow-up message is not a second charge.
• Who adjudicates disputes? If you believe an outcome was wrongly billed, what is the process and the evidence you can inspect?
• Is a reversed outcome refunded? If the agent "resolves" a ticket that reopens an hour later, does the charge reverse?

This is the same discipline we cover in our deeper look at pay-per-resolution and outcome-based AI agent pricing, and it is worth pairing with a hard-nosed view of the full bill. The sticker price per outcome is only part of what you spend, which is why a proper total cost of ownership view, including integration, monitoring, and the human review the agent still needs, belongs in every procurement model.

What an AI agent SLA should actually guarantee

Traditional software SLAs promise uptime: the service will be available 99.9% of the time. For an autonomous agent, uptime is nearly irrelevant. An agent that is up 100% of the time and wrong 20% of the time is worse than useless, because it acts on those wrong decisions at machine speed. The service level that matters for an agent is the quality and safety of what it does, not merely whether it is running.

A serious agent SLA in 2026 covers a wider surface than a SaaS SLA ever did. Look for commitments across these dimensions, and be wary of any vendor that only offers you availability numbers.

The remediation row is the one most buyers forget. If the agent misroutes a thousand tickets or emails the wrong customers, what do you actually get back? Service credits are the floor. When an agent can take consequential actions, you want explicit liability and indemnification language, not a mutual hand-wave.

The contract clauses that did not exist three years ago

Because agents act, and not just compute, procurement teams are pushing new clauses into contracts that standard SaaS templates never contemplated. The industry has even coined a name for the resulting structure, the Agentic Enterprise License Agreement, or AELA. The Adecco Group signed one of the first publicly reported AELAs in early 2026, a multi-year deal for global access to Salesforce's Agentforce platform across dozens of countries. Whether or not your deal carries that label, the underlying clauses are the ones to fight for.

• Delegation of authority. Exactly what the agent is and is not allowed to do on your behalf, with hard limits on spend, data access, and irreversible actions. This is the single most important boundary in the contract.
• Outcome definitions. The literal, written rule for any billable or SLA-measured outcome, plus how disputes are resolved and what evidence you can inspect.
• Audit rights. The right to inspect logs of the agent's decisions, tool calls, and actions, so you can verify behaviour rather than trust a dashboard.
• Data use and training. Whether your data or your customers' interactions can be used to train the vendor's models, where data is processed and stored, and how long it is retained.
• Liability and indemnification. Who is responsible when the agent acts wrongly, including expanded indemnity that reflects the agent's autonomy.
• Exit and portability. How you offboard, what happens to your data and configuration, and whether outcomes in flight are honoured after termination.

Governance, the EU AI Act, and shadow agents

Only about one in five organisations reports having a mature governance model for autonomous agents, and data quality is the single most cited blocker to deployment. That governance gap is not just an internal problem; it is increasingly a regulatory one. For any EU-facing deployment, your agent procurement needs to map to your obligations under the EU AI Act and existing data-protection law, which means requiring vendors to support your record-keeping, transparency, and human-oversight duties rather than leaving them to you after the fact. Our guide to the EU AI Act and business automation breaks down what those duties look like in practice.

There is also a quieter risk on the buyer's side of the table: shadow agents. Just as shadow IT crept in when teams bought SaaS on a credit card, individual departments are now spinning up agents outside any procurement or security review, each with its own data access and its own bill. A central procurement standard, even a lightweight one, is how you keep that sprawl from becoming your next incident. If governance is where you are weakest, tackle that before you widen the number of vendors and agents in your estate.

A procurement checklist you can use this week

You do not need a fifty-page framework to buy your first agent well. You need to walk a short, disciplined path and refuse to skip the steps that feel tedious, because those are exactly the steps that decide whether you land in the 11% or the 89%.

1. Define the outcome first. Write down, before you talk to any vendor, what success looks like in measurable terms and what a failure looks like. If you cannot define it, you are not ready to buy per-outcome.
2. Scope a narrow pilot. Pick one process, one clear metric, and a small budget cap. A tightly scoped pilot with a defined success bar is the cheapest way to learn how an agent behaves on your real data.
3. Model your worst-case bill. Ask exactly how a billable unit is defined, then price your peak month, not your average. Negotiate hard spend caps and per-day ceilings into the contract, not just the dashboard.
4. Demand an SLA with teeth. Require commitments on resolution rate, accuracy, escalation, latency, and cost, plus remediation and liability when they are missed.
5. Pin down the clauses. Delegation of authority, audit rights, data-use terms, liability, and exit. Have security and legal review anything that touches revenue, customer data, or regulated processes.
6. Instrument before you scale. Agree how you will measure the agent in production and who owns the ongoing evaluation. Evaluation drift is a documented killer; plan against it from day one.
7. Keep a human gate on consequential actions. Money, access, and customer-facing actions at scale should route through a checkpoint until the agent has earned trust on your data.

None of this is exotic. It is the same rigour you would apply to onboarding any contractor who could spend your money and talk to your customers. The only novelty is that the contractor is software, moves at machine speed, and bills by a unit someone else defined. That combination is exactly why the buying decision, not the technology, is where the value is won or lost.

Build versus buy, and where a marketplace fits

Not every agent needs a six-figure enterprise contract. For many focused tasks, buying a ready-made workflow or assembling one on a low-code platform gives you full control over the logic, the data access, and the cost, without a per-outcome meter running on someone else's terms. The trade-off is that you own the guardrails, the monitoring, and the maintenance yourself, which is real work but keeps the economics transparent.

The pragmatic pattern for 2026 is to match the buying model to the stakes. Use a self-built or marketplace workflow where the task is well-defined and you want cost certainty, and reserve the heavier vendor contracts, with their AELA-style clauses, for agents that genuinely act autonomously on high-value processes. Before you sign either way, it is worth reading how buyers get burned so you can spot the same traps early; our guide to buying an AI agent without getting burned pairs naturally with the contract discipline in this playbook.

FAQ

What is outcome-based pricing for an AI agent?

It charges you only when the agent produces a defined result, such as a resolved support conversation, rather than for seats or raw usage. Intercom Fin bills around $0.99 per resolution, Salesforce's Agentforce Help Agent launched at about $2, and Sierra is paid only when its agent resolves an issue without human help. The definition of a valid outcome is where the real negotiation lives.

Why do most AI agent pilots fail to reach production?

Gartner has reported that roughly 89% never scale, and MIT found about 95% of pilots delivered no measurable P&L impact. Forrester traces the failures to unclear success criteria, insufficient tool or data access, and evaluation drift, all of which are procurement decisions you make before the pilot goes live.

What is an Agentic Enterprise License Agreement?

It is a contract structure built for autonomous agents rather than seat-based software, bundling platform access with delegation-of-authority limits, outcome-based SLAs, audit rights, and expanded indemnification. Adecco signed one of the first public AELAs, for Salesforce Agentforce, in early 2026.

What should an AI agent SLA measure?

Resolution or completion rate, accuracy against an audited sample, escalation behaviour, latency, and a hard cost cap, plus remediation and liability when targets are missed. Uptime alone tells you the agent was running, not whether it was making good decisions.

How do I avoid a runaway bill?

Negotiate hard spend caps, per-run and per-day ceilings, and alerting into the contract. Prefer hybrid pricing with a predictable base fee, confirm failed or reversed outcomes are not billed, and model your worst-case month rather than your average.

Should I buy per-seat, per-usage, or per-outcome?

Per-seat is predictable but disconnects cost from value; per-usage tracks effort rather than results; per-outcome aligns cost with value but only when the outcome is cleanly measurable. A hybrid of a base fee plus metered outcomes is often the safest first deployment.

What data and compliance terms do I need?

Where data is processed and stored, whether it can be used to train the vendor's models, retention and deletion, sub-processor disclosure, and for EU deployments, alignment with the EU AI Act and data-protection law. Add liability terms for when the agent acts wrongly.

Do I need a lawyer to buy an AI agent?

For a small, capped pilot, standard terms are usually fine. Once an agent touches revenue, customer data, or regulated processes, have legal and security review the delegation, outcome, liability, audit, and data-use clauses, because these are exactly the ones standard SaaS templates handle poorly.