How Performance Based AI Engagements Are Written
By MetaTechAi ยท
A performance based AI engagement only holds up when four terms are fixed in writing before installation: the single metric being measured, the baseline it is measured against, the window the measurement runs over, and the remedy if the number does not move. Add a data access term that puts the buyer in control of the system the metric is read from. Skip any one of these and a "conversions go up or you don't pay" promise is a handshake with extra steps, no matter how confident the pitch sounds.
Tie the written terms to the operating work: managed AI service responsibilities and conversion measurement explains what to scope with the implementation team. For the related checks and handoffs, use our ai sales implementation guide.
This applies to service businesses with sales teams evaluating MetaTechAi or any other provider offering a guarantee. The questions below are what a buyer should ask before signing, so the promise rests on written terms rather than on how the call felt.
Evaluate the operating responsibilities alongside the promise: our AI automation agency overview for service teams explains partner fit. For access, approval and exit evidence to include in the agreement, use the AI infrastructure handoff and ownership checklist.
Why does the metric have to be singular?
Pick one number: booked appointments, qualified calls answered, or closed deals, whichever matches what the sales team actually chases. Name one primary metric that triggers the guarantee. You can track supporting metrics too, but distinguish them from the agreed trigger so neither side substitutes whichever number supports its argument.
Before signing, ask the provider to write the metric definition in one sentence, in language your own team would use on a Monday call.
How should the baseline be set and frozen?
The baseline is the number the guarantee is measured against, and it has to be captured before any part of the system goes live, then frozen in the agreement so neither side can move it later. If the provider proposes setting the baseline after installation, or during a "ramp up" period where changes are already running, the comparison is contaminated from the start.
A defensible baseline covers a period long enough to average out normal week to week swings, pulled from the buyer's own historical records rather than a provider estimate. Write the exact number, date range, and source into the agreement as an appendix, not a verbal understanding. This is also where CRM readiness before an AI sales system matters: a baseline pulled from incomplete records will not hold up to scrutiny later.
What is the measurement window and who can read it?
The window needs to be long enough to cover the buyer's normal sales cycle, not a fixed 30 days chosen because it fits a template. For example, a business whose sales cycle takes several weeks should not judge closed-deal results before those leads have had time to convert.
Just as important is who can read the number. The agreement should name a single system of record, whether that is the buyer's own CRM or a shared dashboard, and state that the buyer has standing, ongoing access to it, not a monthly summary the provider compiles. Ask directly whether the provider can edit historical records in that system after the fact. If edits are possible, require an audit history, agreed correction rules, and a preserved baseline so both sides can distinguish legitimate corrections from changed results.
What counts as a material change that pauses the guarantee?
A guarantee has to name the events that pause it, because a provider cannot reasonably be held to a result the buyer's own actions made impossible. Fair examples include the buyer cutting ad spend that fed the lead volume the system depends on, closing a location, changing pricing that shifts who calls in, or removing access the system needs.
List these specifically rather than leaving a blanket "outside factors" clause either side can invoke whenever convenient. A specific list stops the provider from blaming the buyer for a system that is not working, and stops the buyer from expecting a guarantee to survive changes they made to their own business.
How should the remedy be defined?
The remedy is what happens if the metric does not clear the agreed threshold within the window, and it needs a plain definition, not a phrase like "we will make it right." Possible remedies to negotiate include the client not paying for the period in question, a defined extension at no added cost, or a partial credit tied to how far short of the target the result landed.
Write the remedy as a specific action with a specific trigger: if the metric is below the baseline plus the agreed threshold at the end of the window, then this specific thing happens.
What are fair exit and data return terms?
Every agreement needs an exit path that does not strand the buyer's data with the provider. Ask the agreement to confirm ownership and export access for your CRM data, call records, and lead information, subject to applicable retention and privacy requirements. Define the handoff of account-specific work, including any third-party licenses or provider tools that cannot transfer.
This is the same standard covered in what an AI automation handover should include: a named owner on the buyer's side, usable documentation, and verified access, not a system only the provider's team knows how to run.
What obligations is it fair for the buyer to accept?
A guarantee only works if both sides carry real obligations. Reasonable buyer-side terms include a lead volume floor tied to the buyer's own historical numbers, since no system can produce results from leads that never arrive. They also include timely response responsibilities for anything the system routes to a human, since a booked appointment nobody calls back is not a system failure.
Access is the other fair obligation: working access to the phone numbers, CRM, and calendars the system depends on, granted before the measurement window starts.
What questions expose a weak offer fast?
A short list of direct questions separates a written agreement from a sales promise:
- Who owns the dashboard the metric is read from, and can the buyer log into it directly?
- Can the provider edit historical records in the system of record after the engagement starts?
- What happens to the guarantee if lead volume drops for reasons unrelated to the system, such as a seasonal slowdown?
- Is the remedy a specific, triggered action, or a general assurance?
- Does the measurement window cover the buyer's real sales cycle?
MetaTechAi states that it agrees on the conversion measure before installation and that conversions go up or the client does not pay. Ask for the metric and the engagement's specific conditions in writing before signing.
Why does a public performance claim need substantiation?
Any performance claim a provider makes in its marketing, whether a case study, a testimonial, or a number on a landing page, has to be backed by evidence under the FTC's advertising and endorsement guidance. That standard exists to protect buyers from unverifiable claims, and it is a useful filter before you ever reach the contract stage.
If a provider's public claims would not survive that standard, treat it as a signal about how carefully the rest of the offer was built. The NIST AI Risk Management Framework provides voluntary guidance for managing AI risks and evaluating trustworthy systems. Use it to inform questions about data handling and oversight; it does not establish contract remedies or certify a provider's sales results.
What should you do before signing?
Write the metric in one sentence. Confirm the baseline is captured before installation and frozen in an appendix with its date range and source. Name the system of record and confirm standing access to it. List the material changes that pause the guarantee. Define the remedy as a specific action with a specific trigger. Confirm your data and account access come back to you on exit.
None of this requires distrust of the provider. It requires a written agreement that still makes sense six months from now, after the sales call has worn off and someone has to explain the result to a business partner. For what to prepare before that conversation starts, see what to prepare before hiring an AI automation agency. MetaTechAi's managed AI sales and marketing services include CRM workflows, dashboards, monitoring, and tuning. Confirm how those services support the metric and written terms for your engagement. To work through your own metric, baseline, and measurement window, book a call and bring your current numbers.
If you are still deciding whether to bring in outside help at all, AI Guy's guide to hiring an AI consultant covers the questions to ask before you get to contract terms.
What else do buyers ask about performance based AI engagements?
Are performance guarantees realistic for a low volume business?
They can be, but both sides must agree how much data is needed and what happens if the result remains inconclusive. A longer window alone does not ensure a reliable result. A business booking a handful of jobs a week needs a longer window than one booking dozens a day, and the agreement should say so in writing rather than assume a standard calendar month works for everyone.
What happens if both sides read the data differently?
This is why the agreement should name one system of record and give the buyer standing access to it before installation starts. If the only place the metric lives is a dashboard the provider controls and the buyer cannot log into, independent verification becomes harder. Agree on shared access, exports, and a dispute process before work begins.
How long does an engagement need to run before a result is fair to judge?
Long enough for eligible leads to complete the buyer's normal sales cycle and for the agreed measurement plan to produce a useful comparison. Capture the baseline separately before installation, and define how to handle an inconclusive result. A provider unwilling to state a window before signing is asking for faith instead of a written term.
Can a provider just show me results from other clients instead of writing terms?
Case studies are useful context, but the FTC's guidance on advertising and endorsements says any performance claim made in marketing has to be substantiated, so a provider quoting results still needs evidence behind them, not a substitute for terms covering your own account.
What is a fair volume floor for the buyer to accept?
A floor tied to the buyer's own historical lead or call volume over a comparable period, written as a condition that pauses the guarantee if volume drops below it for reasons outside the provider's system, such as a paused ad budget or a closed location.