Key takeaways
“A contract allocates risk on the day it is signed. Governance decides whether anyone notices when the risk arrives.”
The AI contract on your desk was probably drafted for a different kind of product. It promises uptime but not accuracy, allows the vendor to change the model without telling you, and leaves data use, liability and exit to default terms most people never read. The price you negotiated is the smallest part of what you will pay. Most of the cost arrives afterwards, as usage growth, rework, lock-in and risk nobody owned. This guide is for procurement, legal, security and business leaders who want to know what to check before signing and how to stay in control afterwards.
The Contract Is Where the Product Is Really Defined
When you buy conventional software, the contract is mostly about price, term and uptime. The product itself is what you saw in the demo, and it stays broadly the same until the next release. An AI contract is different. The thing you are buying is a service whose behaviour can change underneath you, runs on data you supply and produces outputs that nobody, including the vendor, can fully guarantee in advance.
That makes the contract the only place where many of the product's most important properties are fixed — or left open. Four are routinely settled by default rather than by decision.
- What happens to your data. Whether inputs and outputs are retained, who can access them and whether they can be used to improve the vendor's models.
- What the product will be next quarter. Whether the model, prompts or features can change without notice, and whether you can stay on a version you have tested.
- Who is accountable when it is wrong. Whether liability for outputs sits with the vendor, with you or with nobody.
- How you leave. Whether your data, configurations and workflows can be exported in a usable form, and at what cost.
Vendors are generally willing to discuss these points, and many publish clear terms. The difficulty is that buyers sign standard paper in the final week of a procurement cycle, with the commercial negotiation already won and the deadline close. The terms that determine long-term cost are then accepted as written.
Why Careful Buyers Still Sign Before They Understand
This is rarely carelessness. It is a pattern with structural causes, and recognising them is the first step to breaking it.
The people who read contracts are not the people who use the product
Legal reviews liability, indemnity and jurisdiction. Procurement reviews price and term. The business sponsor reviews the demo. Security reviews a questionnaire. Nobody reviews how these pieces fit together, such as whether a retention clause is compatible with the sponsor's intended use of customer records.
Standard terms were written for another kind of product
Many contract templates assume deterministic software. They contain generous uptime commitments but are silent on accuracy, drift or model substitution, because those concepts did not matter when the templates were drafted. Silence is not neutral: it defaults to whatever the vendor's terms of service say.
The pilot becomes the deployment
A trial begins under light terms and an enthusiastic sponsor. Months later, with workflows built around the tool and users depending on it, the production contract arrives. Negotiating leverage is gone, because switching now has a cost.
Urgency does the vendor's work
End-of-quarter pricing, limited-time discounts and fear of competitors moving first all compress diligence. These pressures are legitimate commercial tactics, but they are best answered with a pre-agreed checklist rather than a heroic last-minute effort.
The conclusion is practical rather than critical: an AI contract needs a cross-functional review with a shared set of questions, started before the commercial endgame, not during it.
What an AI Contract Costs You After Signature
The signed price is rarely the price you pay. Most of the expense arrives later, in forms that do not appear on the order form.


- Usage-based growth. Pricing per token, seat or query looks modest during a pilot. Once adoption spreads across teams, spend can rise much faster than planned, and you may have no cap or committed-use discount to protect you.
- Unannounced model change. If the vendor swaps the underlying model, outputs shift. Your prompts, tests and downstream processes may need rework you did not budget for.
- Compliance rework. If data terms turn out to be incompatible with your obligations, you pay to remediate, renegotiate or stop. Rules differ by sector and jurisdiction, so obtain proper legal advice on your specific position.
- Lock-in. The deeper the workflows, prompts and integrations sit inside one vendor's environment, the higher the cost of leaving. Exit terms matter most when you can least afford them.
- Liability you did not expect. If an AI output leads to a customer harm or a wrong decision, a contract that disclaims all responsibility leaves your organisation answerable for something you could not fully control.
- Oversight overhead. Someone must monitor quality, cost and use. If the contract gives you no audit logs or reporting, you either build that yourself or operate blind.
The last item is the one most often missed. A contract can protect you on paper, but it cannot tell you what your people are actually doing with the tool today. For that you need visibility inside your own organisation.
The Clauses and Questions to Check Before You Sign
You do not need to be a lawyer to review an AI contract usefully, though legal and compliance teams — including those in legal and compliance technology — will want a seat at the table. You need to know where to look, and to insist on clear written answers. The following checks work as a working group agenda for legal, security, procurement and the business sponsor.
- Data use and retention. Are inputs and outputs used for training? Where is data processed, how long is it kept and can deletion be confirmed in writing?
- Sub-processors and jurisdictions. Who else touches the data, and will you be told when that changes?
- Model and feature change. Are you notified before the underlying model changes? Can you pin a version for a defined period?
- Performance commitments. Beyond uptime, does the contract include any measurable quality criteria, a testing window or remedies if results degrade?
- Audit and reporting. What logs, usage reports and records will you receive, and in what format?
- Liability and indemnity. What is excluded, what is capped and what indemnities exist for intellectual property or data breaches?
- Pricing mechanics. What triggers price changes at renewal, what are overage rates and can you cap spend?
- Exit and portability. What can you export, in what format, over what period and at what cost?
Mark each answer as acceptable, negotiable or a dealbreaker before commercial talks conclude. A vendor who engages clearly with these points is giving you useful evidence about how the relationship will run.
Governing the Contract After You Sign It
Signature is the start of the work. A good contract sets the rules, but you still need to see whether they are followed and whether your own use stays within them. That means three ongoing disciplines.
Keep visibility of how AI is really used
Contracts bind the vendor and your organisation, not your individual employees. If prompts containing sensitive information flow into approved tools — or into unapproved ones — the terms you negotiated may not protect you. Leaders need to see who is using which capabilities, with what kinds of input. This is the thinking behind prompt governance, and it is why a governance layer belongs alongside the contract, not after an incident.
- Assign an owner for each AI contract. One accountable person tracks renewals, price changes, notifications and vendor updates.
- Test on a schedule. Re-run an agreed evaluation set whenever the vendor announces a model change, and quarterly regardless.
- Review usage and cost monthly. Compare actual spend and adoption against the assumptions in the business case.
- Keep portability real. Store prompts, evaluation sets and configuration in a form you control, so an exit is a project rather than a rebuild.
- Revisit risk tiers at renewal. Use real usage data, not the original forecast, as the basis for renegotiation.
Prospiretech's work includes a governance platform built for exactly this visibility problem — see An Enterprise-Grade Dashboard That Makes Prompt Governance Visible to Leadership. Our AI strategy and advisory work helps organisations choose and structure vendors against business priorities, while MLOps and model engineering provides the evaluation, versioning and monitoring needed to notice when a vendor's model has changed underneath you. For regulated sectors such as financial services, that combination of contractual protection and operational evidence is hard to do without.
An Enterprise-Grade Dashboard That Makes Prompt Governance Visible to Leadership
A contract tells you what a vendor must do. It does not show you what your own people are doing. an enterprise SaaS company in the USA that builds AI governance and prompt management tools for large customers in regulated industries, had a working rules-based governance engine but no interface for the compliance leads, product managers and executives who needed to use it. Prospiretech designed and built the complete customer-facing platform: an executive dashboard showing governance status, a structured prompt submission flow, and a results layer that turns the engine's output into clear pass, fail or escalate views. It connects to the client's engine through REST APIs, and each enterprise tenant runs in full isolation.
- Leadership can see the state of prompt governance across the organisation for the first time, without touching the engine
- Deployed across multiple enterprise tenants with no changes to the underlying engine
- Delivered Okta-ready, supporting enterprise identity provider integration without additional development
- Non-technical users can create, submit and track prompts with no training required
Frequently Asked Questions
What should we check first in an AI vendor contract?
Start with data use, because it is the hardest term to fix afterwards. Check whether your inputs and outputs are retained, used for model training or shared with sub-processors, and where processing takes place. Then look at model change notification, liability and exit. If you only have time for a few checks, choose the ones where a poor answer would cost the most to reverse once workflows depend on the tool. A short, shared checklist makes this quicker than most teams expect.
Can we negotiate standard AI terms, or are they fixed?
More is negotiable than most buyers assume, particularly for larger or strategic accounts and particularly if you raise points early. Data handling commitments, notice of model changes, pricing caps and exit support are common areas of flexibility. Where the vendor will not move, ask for the reasoning in writing, consider technical mitigations such as restricting what data is used, and record the residual risk and who accepted it. A cross-functional review before the commercial endgame also helps legal, security and the business reach the same view on acceptable risk.
Why does model change matter so much?
Because AI behaviour depends on the model underneath. If the vendor swaps or updates it, outputs can shift in tone, accuracy or format even though nothing changed on your side. Prompts that worked may need rework, and downstream processes may break. Contracts that provide advance notice and a way to stay on a tested version give you time to evaluate, while silence leaves you discovering changes through user complaints. Vendors with good products generally accept that sensible visibility and controls benefit both sides.
How do we keep AI costs from growing unexpectedly?
Model costs at higher usage before signing, ask for spend caps or alerts, and understand the pricing triggers at renewal. After go-live, review actual usage monthly against the business case and look for teams or workflows driving growth. Usage-based pricing is not inherently bad, but it rewards adoption faster than budgets are usually revised. Visibility into who uses what is the practical foundation for managing cost. Review the contract against actual usage at least annually, since the way teams use AI tends to change faster than the paper does.
Does a good contract remove the need for AI governance?
No. A contract governs the relationship between you and the vendor; it does not control how individual employees use AI day to day, what they put into prompts or whether usage stays within your policies. You still need visibility, access controls and audit records inside your own organisation. The strongest approach combines protective contract terms with operational governance that produces evidence you can show to leadership, auditors and regulators. Think of the contract as the floor and governance as the evidence that the floor is holding.
Sign with your eyes open, then keep them open
The right contract terms and the right visibility together decide what an AI purchase really costs. Prospiretech will help you review the risks, structure vendors against your priorities and put the governance in place to prove it.
