AI development agreement for a bespoke model

A client commissioning a bespoke model (a fine-tuned classifier, a domain-specific assistant, a model trained on the client's own data) needs an agreement that a software development contract does not provide: who supplies the data and on what basis, which base model sits underneath and what its licence allows, who owns the resulting weights, what performance counts as acceptance, who tests for bias, and how the client runs the model when the developer is gone. The agreement can be drafted for either side. I draft or review the AI provisions for the agreement for a fixed fee of £595, delivered in five working days.

Who this is for

AI developers and consultancies in England and Wales building models for clients, and businesses commissioning a model trained on their own data who want to own what they are paying for.

What matters in a bespoke model development agreement

The scope, the data and the model the client is paying for

The agreement describes the model to be built (its purpose, its inputs and outputs, the base model or architecture, the deployment environment), the data the client will supply (its format, volume, quality and the labelling the developer or the client will do), the data the developer will supply or source (with the licences it holds), the development stages (data preparation, training, evaluation, deployment, documentation), the deliverables (the model weights, the training code and configuration, the evaluation results, the documentation, the deployment), and the timetable; the agreement states what is fixed, what is a target and what happens when the target is missed.

The base model, its licence and the terms that flow into the result

Most bespoke models are fine-tuned on a base model (open-weight or proprietary) whose licence sets what may be done with the result: some permit commercial use freely, some restrict it by field or by the user's size, some require attribution or the same licence on derivatives, some prohibit use to train competing models, and proprietary providers' fine-tuning services keep the weights on their platform under their terms; the agreement identifies the base model, warrants the developer's compliance with its licence, flows the licence's restrictions into the client's rights, and addresses the position where the base model's licence changes or the provider withdraws it, since the client's model is only as free as its base model.

Ownership of the model, the weights and the training artefacts

The agreement assigns to the client the weights of the fine-tuned model, the training configuration, the evaluation results and the documentation (subject to the base model's licence), licenses or assigns the developer's tooling and pipeline code as negotiated (the developer usually keeps its generic tooling and licenses it; the client owns anything built specifically for it), and states that the client's data remains the client's and is returned or deleted; the Copyright, Designs and Patents Act 1988 names the person who made the arrangements as the author of a computer-generated work (section 9(3)), a provision untested on modern tools, and protects only works that are original, and model weights may be protected as a computer-generated work, as a database, as confidential information or not at all, so the assignment covers all rights however arising and the confidentiality obligations protect the weights regardless.

Performance, acceptance and the metrics that define done

Acceptance is defined by stated metrics on a stated evaluation set (accuracy, precision and recall, latency, cost per query, task-specific measures), with the evaluation set agreed and held back from training, the threshold stated as a target with a range, the consequences of falling short stated (further iterations within the fee, a reduced fee, termination with the deliverables to date), and the client's own testing on its live data provided for; the metrics and the evaluation set are the agreement's core.

Bias, testing and the obligations the client inherits

putting personal data into an AI tool is processing under the UK GDPR, so it needs a lawful basis, transparency to the individuals, an Article 28 agreement with the provider and, for uses that affect people, an Article 35 impact assessment, and a model trained on the client's personal data makes the developer the client's processor (with the data processing agreement) and the client the controller responsible for the basis, the notice and the assessment; where the model will make or shape decisions about people, when an AI tool alone decides something with legal or similarly significant effects for a person, Article 22 of the UK GDPR as amended by the Data (Use and Access) Act 2025 permits it outside special category data only if the person is informed, may make representations, may obtain human intervention and may contest the decision, and the agreement requires the developer to test for bias across protected characteristics, to document the model (its data, its limits, its intended use), and to support the safeguards, while the client remains responsible for the use; where the model will be placed on the EU market, the EU's AI regulation's obligations are allocated between the parties.

Exit, hand-over and the model the client can run without the developer

The agreement provides for the client to run the model independently: the weights, the code and the documentation delivered in a form the client can deploy, the deployment environment in the client's own accounts, the dependencies and licences listed, a knowledge transfer period, support and retraining services at stated rates, and the developer's deletion of the client's data and model copies on termination with certification; trade secrets, confidential information and clients' data stay out of any tool whose provider may train on or retain the input, a line the Trade Secrets (Enforcement, etc.) Regulations 2018 (which protect only secrets kept by reasonable steps) and the business's confidentiality obligations both require, and the agreement's confidentiality terms cover the client's data and the model in both directions; liability sits within a cap connected to the fee, with the developer's warranties limited to the stated metrics and the steps it has committed to.

What it costs

AI provisions in a contract, £595. Review or draft AI-specific terms in a supply or customer agreement. Five working days.

AI use policy, £995. Internal policy plus the accompanying contract provisions. Five working days.

Email me at geoffrey@caesar.co.uk with what you need and any documents, and I will confirm the scope and the fee in writing, usually the same working day.

What you get

• An internal AI use policy your staff will follow, rather than a document that bans everything and is quietly ignored
• Contract provisions covering output ownership, training data, confidentiality and liability
• A defensible position to give customers who ask what your AI governance looks like
• Review of your AI suppliers' terms, including what they reserve the right to change
• Practical guidance on what to permit, what to restrict and what to prohibit outright
• One round of amendments

What is not included

• Technical AI assurance, model testing or bias auditing
• Sector-specific regulatory approval
• Advice on non-UK regulatory regimes beyond flagging where they apply
• Ongoing regulatory monitoring

Questions I am often asked

Will we own the model the developer builds for us?

The fine-tuned weights, configuration, evaluation results and documentation, subject to the base model's licence, which may restrict commercial use or require the same licence on derivatives. The agreement identifies the base model and flows its restrictions into your rights.

How is acceptance defined for a model?

By stated metrics on an agreed evaluation set held back from training, with a target and a range, and stated consequences for falling short. 'Best efforts' buys nothing measurable; a fixed accuracy on unseen data cannot be warranted.

Who is responsible for bias in the model?

The developer tests for it and documents the model; the client, as controller and user, is responsible for the decisions it makes with it and the safeguards Article 22 requires. The agreement allocates both.

Related guidance and services

• AI contracts and AI use policies, £595, the service this page describes
• SaaS and technology contracts, £995
• Data protection agreements and privacy terms, £795
• Training data licence agreement
• AI consultancy agreement

This page is general guidance for businesses in England and Wales, not advice on your own circumstances. Last reviewed: October 2026. Email geoffrey@caesar.co.uk.