How to Negotiate an AI Development Contract: Terms That Matter

AI development contracts drafted from software templates miss the terms that matter most: who owns the prompts and evaluation datasets, how data flows to model providers, what acceptance means for a probabilistic system, and what happens when the provider changes the model. This guide covers the terms, the pricing models, and the priorities, drawing on FISTA Solutions' delivery practice. The engagement document is in what is a statement of work and vendor selection in how to choose an outsourcing partner. This article is general guidance, not legal advice; contracts should be reviewed by counsel.

What terms does an AI development contract need?

Term	What it must cover	Why AI is different
IP assignment	Code, prompts, configurations, datasets, labels, fine-tuned models, docs	Value sits in prompts and data, not only code
Data handling	Classification, permitted uses, flow-down to model providers and subprocessors, retention, deletion	Data reaches third-party models
Acceptance	Measurable thresholds on agreed datasets; method; sign-off	Outputs are probabilistic
Model change	Monitoring, re-evaluation, remediation responsibility, deprecations	Providers change models
Security	Access, devices, secrets, logging, incident notice	Prompts and logs hold sensitive content
Pricing and change control	Model, caps, change requests, approvals	AI scope drifts
Knowledge transfer	Documentation, evaluation assets, runbooks, training	Handover determines ownership in practice
Liability and warranties	Proportionate caps; warranties on process, not on model output	Nobody can warrant a model's every answer
Exit and transition	Transition support, asset delivery, account ownership	Continuity of a running system

How should IP be handled?

Assign all work product to you: code, prompts, system configurations, evaluation datasets and labels, fine-tuned models, and documentation. License any vendor pre-existing tools broadly enough to operate and modify the system. State that vendor accelerators applied to your data create no vendor rights in the output. Keep code, infrastructure, and accounts in your name from the start, because contract terms are hard to enforce against assets you never held. Practice is in the ip protection checklist for offshore development.

How should data handling be written?

Classify the data the engagement will touch; state permitted uses; require flow-down of your obligations to model providers and subprocessors with named lists; set retention and deletion; prohibit training on your data unless explicitly agreed; and require incident notice within defined hours. Verify that the model provider terms the vendor uses satisfy your regulatory obligations. Privacy practice is in ai data privacy compliance.

How should acceptance be defined?

As measurements: task success or accuracy thresholds by category on a named golden dataset built during discovery; latency and cost limits at defined load; safety and adversarial tests passed; documentation, model card, and evaluation assets delivered; client owners trained. Each criterion names the evaluation method and the person who signs off. Criteria writing is in how to write acceptance criteria for ai and the test process in the ai acceptance testing checklist.

What should model change clauses say?

Who monitors provider changes during delivery and after; who re-runs evaluation when a model updates; what remediation the vendor owes if a change breaks acceptance thresholds before handover; how provider deprecations are handled; and whether the system is built with a gateway and fallback so changes are configuration. Without these clauses, the first silent model update becomes a dispute about fault. Change management is in how to manage ai vendors.

Which pricing model fits?

Fixed price suits bounded discovery and specification where scope is knowable. Time and materials with caps and transparent rates suits delivery where scope evolves with evaluation findings. Outcome-based pricing suits only engagements with a measured baseline and an agreed attribution method. Hybrids, fixed discovery then capped time and materials, are the most common workable structure. Pilot structure is in how to structure an ai pilot agreement.

How should liability and warranties be handled?

Warranties should cover process and deliverables: specification followed, evaluation performed, acceptance criteria met at handover, security obligations honored. No vendor can warrant every model output, and contracts that demand it either get refused or priced for the risk. Liability caps should be proportionate to fees with carve-outs for confidentiality, data, and IP breaches. Procurement framing is in the AI procurement for CIOs whitepaper.

What should exit and transition terms cover?

Delivery of all assets in usable form; transfer of accounts and infrastructure already in your name; documentation and runbooks current at exit; a transition period with vendor support; and, for team-based engagements, knowledge transfer and optional conversion terms. Continuity of a running system is the point. Ownership paths are in what is build operate transfer.

What should you prioritize?

If you cannot get everything: ownership of all work product, acceptance as measurements, data handling terms with flow-down, and exit with transition. Rate and liability caps matter, but a cheap contract that leaves the vendor owning your prompts and datasets, or that defines acceptance as satisfaction, is the expensive one. Send your terms first; vendors that negotiate ownership and verifiability are telling you something.

How FISTA Solutions contracts

FISTA Solutions signs IP assignment covering prompts, datasets, and models, keeps code and infrastructure in client accounts, defines acceptance as measurements on golden datasets, includes model change handling through gateways and re-evaluation, and delivers knowledge transfer and transition as named deliverables, through forward deployed engineers, staff augmentation, and AI enablement engagements. The record behind the approach is 150+ projects for 50+ companies across 12+ countries.

To contract for ownership and verifiability rather than hope, message FISTA on WhatsApp, or read what is a statement of work for the engagement document beneath the contract.