What belongs in the contract when you hire a fractional AI lead?
7 minute read. Updated 2026-08-08.
A fractional AI lead's contract must settle 5 things before work starts: who owns what gets built, what data they can access and for how long, a notice period matched to how much institutional knowledge they hold, who made the IR35 determination and why, and what happens to models, prompts and pipelines if they leave. Silence on any one of these becomes a dispute later, usually at the worst possible moment.
This is not a standard contractor agreement
A fractional AI lead sits closer to the core of the business than most contractors ever get. They see customer data, they build the models that make decisions about pricing, risk or fraud, and within months they often hold more institutional knowledge about how AI is actually used than anyone else in the building, including the people who hired them. A contract template built for a 6-week website project was never designed to carry that weight.
5 things need settling before the engagement starts, not discovered after it has already run into trouble: who owns what gets built, what data they can touch and for how long, how much notice either side gives, who made the IR35 determination and on what basis, and what happens to the work if the lead walks. Leave any one of these open and the cost of that gap shows up at the worst possible moment, usually after the person has already gone.
Who owns what gets built
The default should be straightforward. Anything built for the client, using the client's data, to the client's instructions, and paid for by the client, transfers to the client on payment. That covers the models trained during the engagement, the pipelines that feed them, the evaluation results, and the documentation produced along the way.
What it should not cover is everything the specialist brings with them. Most experienced fractional leads arrive with their own frameworks, prompt libraries, evaluation harnesses and general know-how built up across previous engagements, sometimes over years. A blanket clause claiming all intellectual property for the client tries to take that too, and it usually fails one of two ways: the specialist refuses to sign it, or they sign it and quietly ignore it, which leaves the client no better protected than if the clause did not exist.
The better approach names the categories rather than relying on boilerplate. The models, pipelines and outputs built during this specific engagement transfer on payment. The specialist's pre-existing tools are licensed for use in the deliverables but remain theirs. Anything genuinely ambiguous, a new evaluation method developed mid-engagement, say, gets flagged and resolved at the time it arises rather than reconstructed later from emails.
What they can access, and for how long
Fractional leads move faster with broad access, which is exactly why broad access should not be the default. Scope it to what the specific piece of work actually needs: which systems, which datasets, whether personal data is involved, and for what period.
Where personal data is in scope, UK GDPR and the Data Protection Act 2018 apply regardless of the specialist's employment status or how the engagement is structured. The contract should state who decides whether a data protection impact assessment is needed, and confirm that access is logged rather than assumed.
Revocation has to happen when notice is given, not at the end of the notice period that follows it. A specialist working out a final month with live access to production systems and no active oversight is a risk most businesses would never accept from a departing employee. There is no reason to accept it from a contractor either.
A notice period sized to the risk, not the day rate
Buyers often set notice for a fractional lead the way they would for any other contractor: short, because the person only works 2 days a week and looks easy to disengage. That approach does not survive contact with how these engagements actually run. A fractional AI lead working 2 days a week for 8 months typically ends up holding more context about how and why the models were built than anyone else left behind.
The fix is not simply a longer notice period, which a good specialist may reasonably resist agreeing to. It is a notice period paired with a fixed, contracted handover obligation that survives the notice being served, so the final weeks are spent writing down what currently exists only in one person's head, not starting new work that someone else will have to unpick.
The IR35 determination belongs in the contract, not in a side conversation
Since April 2021, the client, not the contractor, makes the determination of whether an engagement is taxed as employment, for medium and large private-sector clients. That decision, and the reasoning behind it, belongs in or alongside the contract. It should never be settled in an informal conversation and left undocumented.
How to make that determination properly, particularly for a role that in most respects looks like part-time employment, is worth treating as its own exercise and is covered in depth elsewhere. What matters at the contract stage is that it is not left silent, because silence is exactly what gets challenged later, and by then nobody remembers the reasoning that was never written down.
What happens to the work when they leave
A fractional AI lead who leaves without documentation leaves behind a system nobody else can safely change. Models with no record of why a threshold was set where it was. Prompts refined over months of trial with no note of what failed and why. Data connections that only exist in one person's credentials.
The contract should require documentation as a deliverable in its own right, not a courtesy extended if there is time at the end: what was built, why it was built that way, what its known limitations are, and how to run and monitor it day to day. A fixed number of paid handover days, agreed at the start rather than negotiated on the way out, is what turns that from an aspiration into something that actually happens.
What to do about it
- Name what transfers and what stays with the specialist; do not rely on a blanket "all IP transfers" clause.
- Scope data access to the task at hand and revoke it the moment notice is given, not once it expires.
- Attach a fixed, paid handover obligation to the notice period rather than simply lengthening the notice.
- Record the IR35 determination and its reasoning in writing before work starts.
- Make documentation of models, prompts and data connections a contracted deliverable with a fixed number of handover days.
Questions people also ask
Should the fractional AI lead own the code they write for us?
No, not by default. Work built for you, with your data, on your instructions, should vest in you on payment. What you should not claim is the specialist's own frameworks, prompt libraries or evaluation tools built up across other engagements; those stay theirs, licensed for your use. Trying to claim everything usually backfires: either the clause gets struck out in negotiation, or it gets signed and quietly ignored, and you are no better protected than if you had written nothing at all.
How much notice should we ask for from a fractional AI lead?
Less important than the length is what happens during it. A fractional lead working 2 days a week accumulates knowledge that does not exist anywhere else in the business. Pair whatever notice period you agree, commonly a few weeks, with a fixed handover obligation: documentation of what was built and why, delivered as part of the engagement, not negotiated as a favour on the way out.
Who decides if this engagement falls inside or outside IR35?
You do. Since April 2021, medium and large private-sector clients make the determination, not the contractor. Record the decision and the reasoning in writing before work starts. The mechanics of making that determination well, particularly for a role that looks like part-time employment in most other respects, are worth treating as their own exercise.
What if the specialist refuses to hand over their own tools and frameworks?
They are not obliged to, and asking is usually the wrong request. What you need is a licence to use those tools for the outputs they produced for you, not ownership of the tools themselves. Most experienced specialists will agree to that readily; it is the demand for outright ownership of their pre-existing IP that causes standoffs, because it asks them to give away the thing that makes them re-hireable elsewhere.
Where the figures come from
Every rate and salary quoted in this article is a median or percentile of figures advertised in UK job postings over the six months to 8 August 2026. They are not rates paid, and the gap widens at the top of a range.
- IT Jobs Watch, UK contract rates, 6 months to 8 August 2026, read 2026-08-08.
- IT Jobs Watch, UK permanent salaries, 6 months to 8 August 2026, read 2026-08-08.