How Procurement Changes When You Buy Outcomes, Not Software

Table of Contents

• The Old Procurement Motion Was Built for Seats
• What Actually Changes at the Table
• The Unit of Value Moves From License to Outcome
• You Now Have to Define the Outcome Before You Buy
• SLAs Stop Being About Uptime and Start Being About Quality
• The New Diligence Checklist
• Security and Data Access Questions Seats Never Forced
• Reliability and the Cost of a Wrong Action
• Pricing Models You Will Negotiate Against
• Who Sits at the Table Now
• A Practical Framework for Your First Outcome Contract
• Insights Most People Overlook
• References

The Old Procurement Motion Was Built for Seats

For most of the SaaS era, buying enterprise software meant answering one core question: how many people will touch this thing? You counted seats, multiplied by a per-user-per-month rate, negotiated a volume discount past some threshold, and locked it in for a year or three. The vendor's job ended at access. Whether your team logged in, configured the tool well, or got any value out of it was, frankly, your problem. Procurement's leverage came from the contract's edges, the renewal date, the seat true-up clause, the auto-renewal you remembered to cancel.

That model assumes the software is a capability you operate. Agentic AI sold as a service inverts the assumption. The vendor isn't handing you a capability to operate; they're operating the capability and delivering you the finished work. When you buy an agent that resolves customer support tickets, you don't care how many "seats" exist. You care how many tickets got resolved, how many were resolved correctly, and what it cost you per resolution. The procurement conversation has to follow the value, and the value has moved.

This isn't a cosmetic change to a line item. It reshapes who's in the room, what the contract guarantees, and how you'll know six months in whether you overpaid. Procurement teams that treat an outcome-priced agent like a slightly weird SaaS renewal are going to get burned, usually by signing a deal that looks cheap per task and turns out to have no quality floor.

What Actually Changes at the Table

The Unit of Value Moves From License to Outcome

Start with the pricing unit, because everything else cascades from it. In a seat model, the unit is a person with a login. In an outcome model, the unit is the thing the agent produces: a resolved ticket, a recovered receivable, a qualified lead, a reconciled invoice, a screened candidate. Some vendors price per task (you pay whether or not the agent succeeds) and some price per outcome (you pay only on success). The gap between those two words is where a lot of money lives.

Per-task pricing is simpler for the vendor and riskier for you, you're paying for attempts, including failed ones. Per-outcome pricing pushes risk onto the vendor, which is why vendors who offer it usually attach a tighter definition of what counts as an outcome and a higher unit price to compensate. A support agent priced at $0.80 per "handled" ticket is a different deal from one priced at $1.50 per "resolved without escalation" ticket, even though the second looks more expensive. The first can bill you for every confused half-answer that bounces to a human. Procurement's first job is to interrogate the unit until it maps to value you actually receive, not activity the vendor can manufacture.

You Now Have to Define the Outcome Before You Buy

Here's the part that catches teams off guard. With software, ambiguity was tolerable, you bought the tool and figured out how to use it. With outcomes, ambiguity is the whole ballgame. If the contract says the agent will "improve collections," you have handed the vendor a blank check to declare victory. What does improved mean? Against what baseline? Measured over what window? Net of what costs?

A real outcome contract forces specificity that procurement rarely had to demand before: the metric, the baseline, the measurement method, the attribution logic, and the dispute process when your numbers and the vendor's disagree. That last one matters more than people expect. The vendor's dashboard will report the agent recovered $400K. Your finance system might attribute some of that to invoices that would have been paid anyway. Who arbitrates? You want that written down before money changes hands, not discovered during a renewal fight. This is the same definitional rigor that shows up across the GaaS cluster in discussions of the "agent of record" concept and vendor lock-in, once a vendor owns the measurement, they own a lot of the relationship.

SLAs Stop Being About Uptime and Start Being About Quality

Traditional SaaS SLAs are availability promises: 99.9% uptime, support response within four hours, credits if the service goes down. Those still matter, but they're table stakes for an agent. The SLA that determines whether you're happy is about the quality of autonomous work, accuracy rate, error rate, escalation rate, and the cost of the agent's mistakes.

Think about what an uptime SLA misses. An agent can be 100% available and 100% online while quietly approving refunds it shouldn't, miscategorizing leads, or emailing customers something subtly wrong. Availability didn't fail. Judgment did. So the contract needs quality thresholds, say, a minimum 92% resolution accuracy with a defined remedy when the agent drops below it, and a clear definition of how accuracy is sampled and audited. Negotiating those thresholds is now a core procurement skill, and it's closely tied to agent reliability and the cost of a wrong action, which I'll come back to.

The New Diligence Checklist

Security and Data Access Questions Seats Never Forced

A seat-based tool mostly sat on the receiving end of your data, you put information in, you got reports out. An autonomous agent acts. It reads from your systems, makes decisions, and writes back: updating records, sending messages, moving money, changing states in your systems of record. That action surface is a fundamentally larger security exposure, and your diligence has to expand to match it.

The questions change. Instead of "where is our data stored and is it encrypted," you're now asking: What systems can this agent write to, and with what permissions? Can it take irreversible actions, and which ones require a human in the loop? How is the agent's identity managed, does it get its own scoped credentials, or is it borrowing a human's? What happens if the agent is prompt-injected by malicious content inside a support ticket or invoice? Anyone scoping these deals should be reading current guidance on agent and LLM application risks, like the OWASP Top 10 for Large Language Model Applications, because the failure modes, excessive agency, insecure output handling, prompt injection, are genuinely new categories your existing security review wasn't built to catch. The principle of least privilege, applied to a non-human actor, becomes a procurement requirement, not just an IT nicety.

Reliability and the Cost of a Wrong Action

In a seat world, a software bug produced a wrong number on a screen and a human caught it. In an agent world, a reliability failure produces a wrong action that may already be done by the time anyone notices. That asymmetry should reshape how you price risk into the contract.

Two agents can have identical accuracy rates and wildly different risk profiles depending on what a mistake costs. An agent that mis-tags a lead costs you a little wasted follow-up. An agent that wrongly issues a $5,000 refund costs you $5,000 plus the cleanup. So your diligence isn't just "how often is it wrong", it's "how bad is wrong, and who eats it." Push for liability terms that reflect the blast radius of an error, not boilerplate that caps the vendor's exposure at one month of fees. McKinsey's research on scaling generative AI in the enterprise keeps landing on the same point: the hard part of capturing value from AI is the operating-model and governance work around the model, not the model itself. Procurement is where a lot of that governance gets encoded or skipped.

Pricing Models You Will Negotiate Against

You'll see a handful of structures, often blended, and it helps to know their incentives before you walk in:

• Pure per-outcome. You pay only when the agent succeeds by a defined standard. Best alignment, usually highest unit price, and the vendor will fight hard over the success definition. Watch for cherry-picked outcomes the agent can game.
• Per-task / per-action. You pay per attempt regardless of result. Cheaper per unit, but you absorb the cost of failures and retries. Demand visibility into the success-to-attempt ratio or you're flying blind.
• Tiered consumption with a floor. A committed minimum spend buys a volume of tasks at a discount, with overage rates above it. Feels familiar to SaaS buyers, which is exactly why vendors like it, it smuggles seat-era lock-in into an outcome wrapper.
• Hybrid platform fee plus usage. A base subscription for access to the agent plus per-outcome charges on top. Reasonable, but make sure the platform fee isn't quietly recreating a seat license you swore you escaped.

The strategic mistake is optimizing for the lowest headline unit price. A $0.40-per-task agent that succeeds 60% of the time is more expensive per successful outcome than a $0.90-per-task agent at 95%. Always normalize to cost-per-successful-outcome before comparing vendors, it's the single most clarifying number in the whole negotiation, and it's why the broader shift from software budgets to labor budgets changes how finance should evaluate these deals.

Who Sits at the Table Now

Seat-based software procurement was a fairly contained affair: a software buyer, IT for security sign-off, maybe legal for the MSA. Outcome-based agent procurement pulls in people who never used to attend.

The line-of-business owner whose metric the agent affects, head of support, head of collections, head of sales ops, is now a primary stakeholder, because they own the baseline and the definition of success. Finance shows up differently too, because the spend is starting to look less like a software subscription and more like variable labor cost, which is the heart of the CFO reframing of agents as opex labor rather than software spend. And whoever owns risk and compliance has real teeth now, because the agent is taking actions in regulated workflows. The buyer inside the enterprise is changing, and procurement has to orchestrate a wider, more cross-functional table than the seat era ever required.

This also changes vendor evaluation. You're no longer just assessing features and a roadmap; you're assessing the vendor's ability to operate a system that performs your work reliably over time. That's closer to evaluating an outsourcer or a BPO than evaluating a software company, which is why incumbents in those categories are nervous, and why the comparison to traditional outsourcing keeps surfacing across the cluster.

A Practical Framework for Your First Outcome Contract

If you're scoping your first GaaS deal, here's a sequence that's served teams well:

1. Define the outcome in measurable terms before you talk price. Write the metric, baseline, measurement window, and attribution rule. If you can't define success crisply, you're not ready to buy on outcomes yet.
2. Demand a pilot with a real baseline. Run the agent against a holdout or a pre-period so you can attribute lift. Vendors confident in their agent will agree; vendors who resist are telling you something.
3. Normalize every quote to cost-per-successful-outcome. Make the vendors show their success rate, or run the pilot long enough to measure it yourself.
4. Negotiate quality SLAs with teeth. A minimum accuracy threshold, a sampling and audit method, and a concrete remedy, credits, rework, or termination, when it's missed.
5. Scope the agent's permissions to least privilege. Enumerate exactly what it can read and write, gate irreversible actions behind human approval, and require its own scoped identity.
6. Write the dispute process for measurement. Decide now who arbitrates when your numbers and theirs disagree, and what data both sides will rely on.
7. Plan your exit before you're locked in. Who owns the agent's accumulated context and learned configuration? Can you leave without the vendor holding your workflow hostage? This is the lock-in question that defines the relationship's later years.

None of this is exotic. It's the discipline procurement already applies to high-stakes services contracts, transplanted onto a category that's still being sold with SaaS-flavored slide decks. The teams that win are the ones who notice the mismatch and refuse to procure an outcome as if it were a license.

Insights Most People Overlook

• The cheapest per-unit agent is often the most expensive per outcome, and vendors know it. Headline unit pricing is the single most effective misdirection in agent sales. A low per-task rate paired with an undisclosed success rate lets a vendor look competitive while billing you for a flood of failed attempts. Always force the conversation onto cost-per-successful-outcome. If a vendor won't share their success rate, that silence is the answer.

• Outcome contracts quietly transfer your domain knowledge to the vendor. To define success, you have to codify how your business judges good work, your collections logic, your lead-qualification criteria, your support quality bar. That codified judgment lives in the vendor's system now. Over a couple of years, the vendor may understand the operational mechanics of your workflow better than your own remaining staff do. That's a deeper, stickier lock-in than any seat contract ever created, and almost nobody prices it into the deal.

• "Per-outcome" pricing can perversely incentivize the vendor to cherry-pick the easy outcomes. If the agent gets paid per resolved ticket, it has every reason to resolve the trivial tickets enthusiastically and quietly punt the hard ones to your humans, then bill you for the easy wins while leaving you the expensive long tail. Smart procurement defines outcomes across the full difficulty distribution, not just the cases the agent prefers to take.

• Uptime SLAs are now a distraction, and vendors lean on them precisely because they're easy to hit. A vendor waving a 99.99% availability number is answering a question you should no longer be asking. An agent that's always online and frequently wrong is worse than one that's occasionally offline and reliably correct. Redirect every SLA conversation from availability to decision quality, and watch which vendors get uncomfortable.

• The procurement skill that matters most is no longer negotiation, it's measurement design. The leverage in an outcome deal is set the moment you define what counts as success. A team that writes a tight, gameable-proof outcome definition has already won most of the negotiation before pricing comes up. A team that lets the vendor draft the definition has lost it. Invest in the people who can design metrics, not just the people who can haggle on rate.

References

• OWASP Top 10 for Large Language Model Applications
• McKinsey, The State of AI