Support outsourcing contracts have started carrying AI clauses, and most of them are written around a single word that nobody defines. The word is deflection, and whether it means a solved problem or a customer who gave up is usually left to the vendor to decide after the contract is signed.
There is a lot of good material on contracting for AI in general. There is almost nothing on what to put in a support outsourcing agreement specifically, which is unfortunate, because 91% of support leaders report executive pressure to deploy AI this year and 10% describe their own deployment as mature. A great many contracts are being signed in that gap.
This is what to negotiate, in the order it matters, and the wording that makes each clause enforceable rather than decorative.
Define deflection before you agree a deflection target. Put a quality floor next to every automation number. Insist the measurement method is in the contract, not in the vendor’s dashboard. And get a route back to a human written in, because only 27% of customers will try a chatbot again after one bad experience.
Clause one: define deflection, or the rest is decoration
A deflection target is the easiest number in a support contract to hit dishonestly, and nobody has to lie to do it. If “deflected” means the conversation ended without reaching an agent, then a customer who got a useless answer and gave up counts as a success. So does a customer who could not find the handoff button. So does a customer who went to a competitor.
Gartner found that only 14% of issues are fully resolved in self-service, against a far larger share that start there. The difference between those two numbers is the space a loose definition lets a vendor bill into.
Wording that closes it. A conversation counts as deflected only if all three hold:
That third condition is the one vendors push back on, and the pushback is the information. A vendor measuring properly already knows the difference between a resolved conversation and an abandoned one.
Clause two: a quality floor beside every automation number
An automation target on its own creates one incentive: automate more. A quality floor beside it creates the right one: automate more without making it worse.
Pair every deflection or automation percentage with a CSAT or resolution-quality minimum measured on the automated conversations specifically, not blended across the whole queue. Blended is how a falling automated experience hides behind a healthy human one.
The reason this matters more than it sounds: only 27% of customers would try a chatbot again after a single bad experience. A deflection number hit by degrading the experience does not just cost you that conversation. It costs you the channel, and the channel does not come back when the bot improves.
Clause three: the measurement method goes in the contract
Almost every dispute about an AI clause is a dispute about measurement rather than performance. The vendor’s dashboard says 58%. Your read of the same data says 34%. Both are calculated correctly, from different definitions, and the contract names neither.
Four things to write down:
Interaction-level tagging is the practical version of all four: every contact marked as AI-handled, AI-assisted or human-only, available to you as data rather than as a chart. Ask for it by name.
Clause four: a human route, written down
Specify that a customer can reach a person, how, and within what time. It sounds obvious and it is routinely absent, because a visible human route works directly against a deflection target and the contract rarely says who wins that conflict.
There is a demand argument as well as a quality one. 49% of customers say they would have used a chatbot if one had been offered, and 7% actually met one. Willingness is not the constraint. The constraint is trust, and a visible exit is what buys it.
Clause five: what the model may learn from
Your support history is a dense record of what your product gets wrong, written in your customers’ own words, with their details attached. It is also exactly the corpus a vendor would most like to train on.
Three lines worth having:
The subprocessor line matters more each year. A vendor may not train on your data and may still be passing it to someone who does.
Clause six: notice before the model changes
A support AI is not a fixed thing you bought. It is a configuration sitting on top of a model that someone else updates. A change you were not told about can move your resolution rate in either direction, and if you find out from a metric rather than from a notice, you will spend a fortnight debugging your own operation first.
Ask for written notice before a breaking change, and a defined response window if quality drops below the floor after one. Accuracy guarantees are rarely offered and the pushback is normal. Notification is a reasonable thing to insist on.
Clause seven: remedies that are not service credits
Service credits are the weakest lever in any outsourcing contract. A credit against next month’s fee does not repair the customers you lost this month, and both sides know the amount is too small to change anyone’s behavior.
Two things that do work. A pilot period with a clean exit, so the first three months are a test rather than a commitment. And a termination right tied to the quality floor, not to the automation target, so missing the number is survivable but degrading the experience is not.
If a vendor will accept an automation target but not a quality floor with teeth, that asymmetry is the most useful thing you will learn in the whole negotiation.
Negotiate the definitions, not the numbers. A vendor will concede a percentage point far more readily than a definition, which tells you which one they think is load bearing.
One clause that has nothing to do with AI and belongs here anyway
Attrition, with the method attached.
Ask for the agent attrition rate on accounts your size over the last twelve months, and ask how it is calculated. Some vendors report voluntary departures only, which excludes everyone managed out, everyone reassigned internally and everyone whose contract simply ended. The resulting figure can be a fraction of actual turnover and it is not dishonest, it is just answering a different question.
It belongs in a piece about AI clauses because the two interact. Automation absorbs the repetitive work, which leaves agents a queue of nothing but hard problems. That is a harder job, and the attrition number is where you will see it before you see it anywhere else.
Frequently asked questions
What is a reasonable deflection rate to write into a support contract?
Do not start from a rate. Start from the definition, then run a pilot and measure what the vendor achieves under it. Any number agreed before you have both signed up to the same definition is a number about nothing. For context, Gartner found only 14% of issues are fully resolved in self-service, so a confident offer well above that deserves a question about how it is counted.
Should pricing be tied to outcomes rather than hours?
In principle yes, and in practice only if the outcome is defined to the standard above. Outcome pricing on a loose definition of resolution is worse than hourly, because it pays the vendor to close conversations rather than to solve problems. Get the definition right first, then move the pricing onto it.
Our vendor says they cannot commit to accuracy. Is that a red flag?
No, that one is normal and honest. Accuracy on live customer language is genuinely hard to guarantee. What is not reasonable is refusing a quality floor, refusing to define the measurement, or refusing notice before a model change. Those three are process commitments rather than performance ones and a serious vendor can make all of them.
Does any of this apply if we are not buying AI from the vendor?
Most of it does, and the data use clauses apply regardless. Your customers are already using third-party assistants: Gartner found people are about three times more likely to ask a general-purpose AI about a product than to use the company’s own chatbot. The question of what is learned from your support history does not wait for you to buy anything.
Who should own these clauses internally?
Support operations, with legal reviewing. Written by legal alone they become generic AI terms that do not touch deflection. Written by support alone they tend to lack the remedies that make them enforceable. The definition section is the part support has to own outright, because nobody in legal knows what your tickets look like.
The bottom line
82% of support leaders invested in AI last year and 10% call the deployment mature. Those two numbers describe a market where a lot of money has been committed and very little has settled, which is exactly the market where contract wording decides how it turns out.
You will not get every clause here. Get the first three. Defining deflection, putting a quality floor beside it, and naming the measurement method in writing covers most of the ways these arrangements go wrong.
Our RFP kit carries 71 questions with the reason each one matters, including the ones that are awkward for us to answer.
Figures: Gartner surveys of 321 customer service leaders fielded October 2025, 3,566 customers fielded February to March 2026, and 5,728 customers, 2024. Intercom Customer Service Transformation Report 2026, 2,400+ professionals. All are listed with links and sample sizes in The State of Customer Support Operations 2026. This page describes commercial practice and is not legal advice.
