The research, free and ungated

A fractional AI leader is priced from a day rate, not a salary divided down. At the median advertised UK contract rate for Artificial Intelligence roles, £551 a day, 1 day a week costs £2,386 a month. 2 days a week costs £4,772 a month, or £50,692 across a 46-week working year. That figure sits above a pro-rated salary because the day rate covers employer costs, non-billable weeks and risk a salary hides inside someone else's payroll.

The choice comes down to accountability, not skill. A consultancy sells a scoped deliverable and its obligation ends when that deliverable is signed off. A fractional leader carries an ongoing decision and is still there when it goes wrong. Hire a consultancy when you can write the acceptance criteria before you start. Hire a fractional leader when the honest answer is that you will not know what good looks like until you are in it.

Confidence is not evidence of competence, and a hirer who cannot assess the technical work directly needs a different signal. The strongest one is whether a candidate can describe something that failed in production and exactly what they changed afterwards. A good engineer gives you a specific failure, a specific fix and what they'd still do differently. A confident one gives you a general story about success, or blames the data, the users, or the deadline.

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.

Yes, almost certainly, but do not expect to find one by posting the job. Advertised UK rates cover Artificial Intelligence, Machine Learning Engineer, Data Scientist and several other titles; AI governance is not among them, over the six months to August 2026. That absence means the market has not settled on what this job is or what it costs, so buyers who wait for a named, priced role to appear will wait past the point the Act reaches them.

Since April 2021, if you are a medium or large private-sector business, you make the IR35 determination for a fractional AI hire, not the contractor and not their agency. For a role working 2 days a week, closely embedded in your team, the practical risk is that the arrangement looks like part-time employment. Assess control, substitution and mutuality of obligation honestly, document the reasoning, and decide before work starts, not after a tax enquiry.

A properly run hire of a genuine AI specialist takes 6 to 10 weeks: scoping, technical assessment, reference checks, then contract and IR35 determination. Scoping around a discipline and checking references with the person who commissioned the work are worth every day they take. Generic take-home tests and extra interview rounds are not; they add weeks without adding information. Cut those, keep the rest.

A stalled AI pilot almost always has a specific cause, and the cause points to a specific discipline: production failures need AI engineering and MLOps, wrong or hallucinated answers need LLM applications and retrieval work, sign-off delays need AI governance and risk and compliance, and low adoption needs AI product or AI training and enablement. Hiring another generalist rarely fixes any of these. Diagnose the symptom first, then hire the discipline that matches it.

Ask for one project described in enough depth that the candidate can go several layers deeper on any question: the constraint, the decision, the failure, the fix. Treat vague phrases like 'AI transformation' and unexplained titles as warning signs. A genuine NDA covers a client's identity and data, not the reasoning behind a decision, so ask for that reasoning with names removed. Call the person who commissioned the work and ask what broke, not whether they would rehire.

For AI capability you can't yet justify full-time, borrow before you build. A contractor at 2 days a week can cost less than an equivalent permanent salary, using advertised UK day rates and salaries, while letting you learn what the work actually requires. Build once the need is continuous and full-time, when the arithmetic on day rates flips in favour of a permanent hire. Buy only the commoditised parts.

By day 30, a fractional AI engagement should have produced a written recommendation with a rough cost and timeline attached, specific findings on data quality, and an honest answer on whether the original brief is achievable, including a recommendation to stop if that's the truth. If instead you have only meetings and reassurance, set a 7-day deadline for the missing deliverable, then end the engagement if it doesn't appear.

A certification proves someone was exposed to a syllabus and passed a test, not that they can do the work. It carries genuine weight for cloud platform certifications on infrastructure-heavy roles, and for governance certifications tied to a recognised framework, because governance has no established job title or qualification path yet. For hands-on engineering and data science hires, put no weight on it and ask for a work sample instead.

A good AI brief describes a decision you can't make yet, not a list of tools. State the symptom, the data you actually have, and what changes once the role is filled. Put a budget band in the brief itself, anchored to published market rates, because leaving it out doesn't protect you. It just means you spend weeks interviewing people who were never going to take the job at the number you had in mind.

An AI strategist assesses whether AI can solve your problem and sets the roadmap; that title barely exists in the advertised market, so expect to price it near broad AI rates. An AI engineer integrates existing models like LLMs into your product, priced close to the £550 Generative AI and LLM median. An ML engineer builds and deploys custom models, priced higher at £575 a day or £76,000 in salary. Hire for the symptom, not the label on the CV.

A day rate buys days, not a person. Contract work at the Machine Learning Engineer median of £575 a day has no holiday, sick pay or pension, and stops the moment the work does. A permanent hire at the £76,000 median salary costs more once employer pension and National Insurance are added, and keeps costing that whether or not there's enough work to fill the year. Choose based on whether the need is bounded or ongoing.

Cap technical take-home tests at 2 hours. Longer tests filter for whoever has free time that week, not whoever is best, because your strongest candidates are already billing elsewhere at market rates. Replace the marathon exercise with a paid, capped task built on a real bug, a live walkthrough of the candidate's own past work, and a reference from someone who has actually reviewed their code.

Hire a permanent AI lead once the role needs 3 days a week or more, sustained for longer than 2 quarters. Below that, a fractional hire is cheaper and more flexible. At 3 days a week, using the advertised median contract rate of £551 a day, the annual cost already exceeds the advertised median permanent salary of £71,000, before adding a single employment cost.

Accountability for AI harm sits with the business that deployed the system, not the supplier who built it or the provider whose model it runs on. Contracts can set out who pays afterwards, but they do not change who a regulator or a customer holds responsible first. Name an accountable owner for every AI system in production, in writing, before it goes live, not after something has already gone wrong.

At 200 people, build a register of every AI system in use before forming a governance committee: name, owner, data, decision, human oversight, and last review date. Assign one accountable owner, usually the CTO or head of data. Add a lightweight review step only once the register shows real risk. Committees formed first have nothing concrete to govern.

The useful CV red flag is a specific tell paired with a resolving question, not a reason to reject outright. Watch for CVs with no failures, titles that don't match described tasks, long tool lists with no depth, and outcomes with no named decision attached. Gaps between contracts, short engagements and a non-AI degree are not red flags; they are normal features of fractional work.

Match the hiring channel to the role. Use a marketplace for bounded work you can check yourself, an agency when filling several similar roles and you want the process handled, and a curated register for a single hard-to-assess hire where getting it wrong is expensive. A curated register is genuinely smaller and slower; that is the trade for the reasoning it gives you. Most businesses need more than one channel across a year.

An AI readiness assessment should check four things before you write a job spec: the state of your data, a measurable definition of success, who owns AI-specific risk, and what you already have in-house. In practice the finding is almost always the same: the data is not fit for the use case, and no specialist, however senior, can fix that on day one. Run the assessment first, hire second.

You cannot manage a fractional AI lead daily because they are not there daily, and trying to compensate for that is the wrong instinct. Hold them accountable through written outputs on a fixed cadence, a fortnightly note covering decisions taken, risks flagged and what's next, built into the contract as a deliverable. It gives you a record to check against, and it works far better than status meetings, which measure presence rather than judgement.

Match the symptom to the technique before you hire. If the model doesn't know your business, that's a retrieval problem, fixed by a data engineer and a generative AI specialist, not a fine-tuning job. If it needs to take multi-step actions across systems, that's agent orchestration, needing MLOps and software engineering skill. Fine-tuning is for narrow, stable tasks like tone or format, and it is chosen far more often than it is actually needed.

When a fractional AI hire and an in-house data team disagree, don't referee it by expertise, referee it by accountability. Give the specialist the final say on whether the method fits the business goal, and give the data team the final say on production and everything they'll own after the engagement ends. Most of these disagreements exist because nobody wrote decision rights down before the work started.

Going fractional changes the ratio of your week more than it changes the work itself. Selling, pricing, negotiating scope and chasing invoices becomes roughly half the job, permanently, not a startup phase you grow out of. What gets easier is autonomy and variety of problem; what gets harder is income smoothness, quality control and the fact that nobody checks your work before the client sees it.

Advertised UK contract rates put the median AI day rate at £551, but that figure prices contract delivery, not decision responsibility. Fractional advisory work means signing off what gets built and owning the consequence if it's wrong, so treat the published median as a floor, not a target. At £551 a day and roughly 8 days a month, a single client at that pattern comes to £4,408; price upward from there once accountability is accounted for.

You cannot show the code, so show the shape of the problem. Describe the architecture and constraints without naming the client, get a named referee to confirm scope and outcome by phone, and rebuild a smaller version of the method on public data so a buyer can see it work. Vague claims about 'delivered ML systems at scale' convince nobody; specific, sanitised detail does.

Since April 2021, medium and large private-sector clients decide your IR35 status, not you. That decision determines whether you are paid gross through your own limited company or taxed like an employee for that engagement. What catches contractors is a blanket inside determination applied without checking real working practices, or a contract that reads outside IR35 while day-to-day work looks like employment. Get the determination and its reasoning in writing before you price the work.

Cloud platform certifications (AWS, Azure, Google Cloud machine learning tracks) and formal AI governance or data protection qualifications carry weight with buyers because procurement and compliance teams can point to them in a tender or an audit. Most vendor 'AI professional' badges and short online course certificates carry almost none, because they test whether you sat through material, not whether you can ship. Spend your time on the first group.

Don't fix a price before you've seen the data. Sell a short, fixed-fee diagnostic first, priced off your advertised day rate, that produces a written scope and a costed proposal for the build. Price the build itself against that scope, anchored to the median day rate for your role and moved toward the 90th percentile only where you can name a specific risk driving it.

Hirers spend less time on your CV than you think and more on 3 things: what a referee says when asked what went wrong, whether you ask about data ownership and failure modes before proposing an architecture, and whether your day rate is defensible against the advertised range for your role. Certifications and polished decks get you shortlisted. Specific references and sharp early questions get you booked.

You can't fix confidentiality, so stop treating it as the obstacle. The real gap is evidence of judgement, and you can build that deliberately: an anonymised case study written around a decision point, one small original project that shows a mistake and the fix, and published scoping or evaluation templates. None of it touches a client's data, and all of it shows more than a generic public-dataset demo ever will.

Two clients is the right number for most fractional AI specialists. One client concentrates risk and can drift into disguised employment under IR35; four clients costs more in context switching and administration than it earns in diversification. At the median advertised day rate for AI work, £551, two clients at 2.5 days each earns more than one client working full time, with far less exposure to a single buyer.

Refuse an AI project when there is no agreed measure of success, the data does not exist yet, or the client has already decided the answer. A failed engagement costs a finite fee; the reputational damage of shipping it does not stop there, because reference calls happen for years. Describe the consequence of proceeding rather than call the client's plan wrong, and offer a smaller, paid diagnostic instead of a flat no.

Machine learning engineers moving into AI governance are entering a market with no published rate to anchor against. Nothing is published for AI governance, AI risk and compliance, or AI safety roles in the UK advertised market, which means the work is being bought under borrowed titles like senior data scientist or MLOps lead. That absence is a genuine risk and a genuine opening: price against adjacent medians, £575 for MLOps or £600 for data science, and argue up.

A first proposal should be two pages that name the decision the client has to make, the option you recommend, what it costs, and what happens if they do nothing. Twenty pages of methodology reads as a hedge against being wrong later. Clients hire people who commit to a position on page one, not people who defer the decision to an appendix.

Before accepting an AI engagement, ask who owns the decision if it fails, what happened to the last person who tried this, whether the budget is signed off or still being argued for internally, and whether you can see a sample of the data before quoting. These questions feel rude. They are the ones that surface a doomed project before you have spent two weeks confirming it yourself.

Sell the data work as its own priced phase, named as what it is, rather than quietly building a demo on data that will not survive contact with production, or folding remediation into the AI fee for free. If the client will not fund the data phase once they have seen what is actually there, walk from the build. A model built on data nobody has properly assessed is not a shortcut, it is a second, worse project waiting to happen.

Before taking a first AI contract, get 4 things in place: a company structure that suits how you plan to work, professional indemnity and public liability insurance, a written contract for every engagement, and an accountant who understands IR35. Set them up in that order. This describes mechanics, not tax or legal advice; get a solicitor and an accountant to check your specific position before you sign anything or take money.

A good technical assessment for AI work uses a real, scoped problem from the client's own domain, tells you upfront how it will be judged, and pays for anything beyond a couple of hours. It should feel like a preview of working with them, not an audition you have to win at any cost. Push back on open-ended builds, unpaid multi-day work and vague evaluation criteria; accept short, paid, clearly scoped tasks instead.

The safest time to raise a day rate is not a contract anniversary; it's a stated condition, such as a scope increase, a renewal point, or clear evidence the market has moved. Bring the reason before the number, give notice, and know your walk-away rate before you ask. Advertised UK contract rates over the six months to August 2026 are useful evidence in that conversation, but they describe adverts, not agreed pay, so use them to open a discussion, not to dictate one.

Going permanent usually looks better for the client's budget than it does for you, once you account for the rate premium, the spread of clients, and the positioning you built as an outsider. Convert only if the equity or bonus is something you'd genuinely hold, or you were already planning to leave independent work. Otherwise, offer a longer commercial term before agreeing to change your employment status.

Professional indemnity insurance is the core cover for independent AI consultants, but AI work raises the stakes because a flawed model can cause harm that compounds quietly across many customers, far beyond the fees you were paid for the work. Check that your policy explicitly covers AI-related work, read indemnity clauses separately from any liability cap, and take a contract you do not fully understand to a solicitor before you sign, not after.

Billing full days leaves no built-in time to learn, unlike a permanent role with training days baked in, so the gap between what you can invoice for and what you actually know widens quietly. Ring-fence around 1 day in 10, fixed to the same recurring day, and spend it building something rather than reading about it. Increase that allocation around genuine shifts, not every model release, and use contract gaps for focused rebuilding rather than job search alone.

When a client asks you to ship something unsafe, put your concern in writing, name the specific risk, and offer a fix before you refuse anything. Escalate once through the client's own structure, with a deadline for a response. If they proceed without addressing it, invoice for work completed and walk. Never do free remediation to avoid conflict, and never let a verbal assurance replace a written record.

Content marketing is a poor first move for an independent AI specialist because senior technical hiring is bought on trust, not discovered through search, and a blog takes months to earn an audience you don't yet have. Build a referral network instead: former colleagues, peer specialists in adjacent niches, and clients you ask directly for an introduction once the work is done.

Never sign an AI contract with uncapped liability, a broad indemnity for third-party claims arising from model behaviour, an IP assignment that swallows your pre-existing tools, or a warranty that a model will be accurate or bias-free. Ask for a liability cap tied to fees paid, a background IP carve-out, and a standard of reasonable care rather than a guaranteed outcome. Take legal advice before signing any contract you didn't write.

The analogies that survive scrutiny with a non-technical client describe a model as a confident junior who never says 'I don't know', and production monitoring as spot checks on a line with an agreed sampling rate. Retire 'black box' and any comparison to human accuracy, both mislead. Sequence the conversation as strengths, failures, downstream consequences, then mitigation, and get the acceptable failure rate agreed in writing before go live.

Read your current contract for non-compete and IP clauses before resigning, and save enough to cover 3 to 4 months without income, since a day rate carries no holiday, sick pay or pension. Quote your actual discipline's rate rather than the broad AI figure, settle who determines IR35 status in the first client conversation, never start before a contract is signed, and keep selling through the first 90 days.