The number that ended the meeting was 22 percent.
That was the year-over-year productivity gain our managed-services provider disclosed, almost in passing, during a contract renewal for a document-review and contract-abstraction function we'd outsourced four years earlier. They were proud of it. They should have been. They had embedded AI-enabled outsourcing tooling into the workflow, retrained a chunk of their delivery team, and could now process the same volume with meaningfully fewer people. The slide was meant to reassure me.
Instead I did the arithmetic in my head and felt the floor tilt. Our contract priced the work on full-time equivalents. We paid for heads. If the vendor needed 22 percent fewer heads to deliver the same output, then 22 percent of the value they'd just created was flowing straight into their margin — and not one basis point of it was reaching my budget line. I was the customer who'd funded the demand, signed the multi-year commitment, and supplied the proprietary data that trained the models. And I was about to renew at the same unit rate.
The measurement that should keep you up at night
Here is the gap, stated plainly, because it's the whole problem in one sentence.
Under an effort-based contract, the vendor's incentive is to bill effort. Under AI, effort and value have come apart. The vendor can deliver more value with less effort, which means an effort-priced contract now rewards your provider for getting better while charging you as if nothing changed.
I went back through three years of our invoices and reconstructed it. Our blended rate had drifted down maybe 4 percent across the term — the normal productivity concessions you negotiate at renewal. But the vendor's internal efficiency, by their own disclosure, had moved roughly five times that. The delta between what they kept and what they passed through wasn't a rounding error. On an annual run rate in the low eight figures, it was a number that would have justified its own headcount in my department, had it ever arrived.
I'm not telling you this to make you angry at vendors. They were behaving rationally inside a contract structure we both signed before any of us understood what AI would do to the delivery economics. The structure was the problem. And if you're holding or renegotiating an outsourcing agreement right now, the structure is your problem too, whether or not anyone has put a 22 percent slide in front of you yet.
Why labor-input pricing breaks, and what's replacing it
The contracts most of us inherited price three things: headcount, hours, and transaction volume. Each of those was a reasonable proxy for value when a human did the work and the work scaled with people. None of them holds when the marginal unit of output is produced by a model that your vendor amortizes across many clients.
So the pricing conversation is moving — unevenly, and with a lot of mutual suspicion — toward structures that try to price the outcome rather than the input. In practice I've seen three show up at the negotiating table.
Outcome and consumption pricing. You pay per processed document, per resolved ticket, per completed transaction, regardless of whether a human or a model did it. This is clean and it aligns interests: the vendor keeps the efficiency it earns, and you stop paying for labor you no longer consume. The catch is that you have to define the unit precisely, and you lose visibility into how the sausage is made — which matters more than you'd think when the model is being trained partly on your data.
Gain-share, or shared savings. You and the vendor agree a baseline cost of delivery, then split the savings that AI generates against that baseline according to a formula. This is the model everyone reaches for first because it sounds fair. It is also the single most disputed clause I've negotiated in the last two years, and I'll come back to why.
Hybrid floor-plus-share. A reduced fixed fee that covers the vendor's committed investment, plus a variable component tied to measured value or savings. This is where a lot of mature deals are actually landing, because it lets the vendor recover the cost of AI capability it has to build regardless, while still giving you a share of the upside that capability produces.
The temptation is to treat this as a menu and pick the cleanest-sounding option. Don't. The structure matters far less than four details inside it, and those details are where the money lives.
How I'd actually decide
Baseline integrity
Every gain-share and hybrid model rests on a baseline — the agreed cost of delivering the work the old way. Whoever controls the baseline controls the split. Vendors have a quiet incentive to set it high, because a generous baseline makes any future savings look modest and shifts more of the gain to them. I now insist on baselining against actual trailing invoices, not a modeled "what it would have cost," and I get the methodology in writing before I'll discuss percentages. If your provider wants to negotiate the split before the baseline, that ordering is a tell.
The gain-share formula and its decay
A flat "we split savings 50/50 forever" clause sounds fair and ages terribly. AI efficiency gains are front-loaded; the big jumps come early. A static split means that in year three you're still paying the vendor half of a saving that's now just the cost of doing business everywhere. I prefer a declining-share schedule — the vendor gets a larger cut of early gains it genuinely drove, tapering as those gains become the market baseline. It rewards the actual innovation without paying rent on it indefinitely.
Benchmarking cadence
The fastest way for a fair deal to rot is to lock the terms and walk away for five years. AI capability is moving faster than your contract term. I now write in a benchmarking review — annual, against a defined peer set — with a renegotiation trigger if the vendor's delivery economics or the market rate move beyond an agreed band. Vendors resist this because it caps their upside. That resistance is exactly why you need it.
Who owns the investment, and the data
Somebody pays to build the AI capability. If the vendor pays, they'll want to keep more of the savings, and that's defensible. But if you're supplying proprietary data that makes their model better — and improves their service to your competitors — that contribution belongs in the value equation. Get explicit on data rights, model ownership, and whether improvements trained on your data can be deployed elsewhere. This is the negotiation point most legal teams under-price, and it's the one that compounds.
Who each model is for
| Structure | Best when | Wrong when |
|---|---|---|
| Outcome / consumption | Output is a clean, countable unit and you trust quality controls | The work is variable, judgment-heavy, or hard to unitize |
| Gain-share | You have clean historical cost data for an honest baseline | You can't agree a baseline, or the function is being transformed beyond recognition |
| Hybrid floor-plus-share | Vendor needs to recover real AI investment and you want shared upside | You want maximum simplicity and minimal monitoring overhead |
If you take one thing: outcome pricing is the cleaner instrument, but most enterprises aren't ready to define the unit precisely enough to use it, so hybrid floor-plus-share is where I'd start for any function with judgment in it. Gain-share is the right answer only when your baseline data is genuinely clean — and far fewer functions clear that bar than vendors will let you believe.
Back to that meeting
We didn't renew at the old rate. We rebaselined against actual invoices, moved to a hybrid with a declining share and an annual benchmark, and wrote data-use terms that should have been there from the start. It took three months and it was not a comfortable negotiation. The vendor kept some of the efficiency — the part they actually earned. We captured the rest. The board got its savings story. I slept better.
The lesson wasn't that vendors are taking advantage of you. The lesson was quieter and more dangerous than that.
When the work gets cheaper to do and your contract doesn't notice, somebody is keeping the difference — and silence is a pricing decision.
