A 50-staff specialist accountancy firm has decided to invest in AI for client document analysis. The managing partner asks the operations director to start a hiring brief. The brief comes back at £75,000 base. The CFO does the maths. Add 15% employer National Insurance from April 2026, workplace pension at 3%, recruitment fees of £5,000 to £8,000, onboarding overhead, and the year-one cost lands near £110,000 before the new hire produces a single deployable system.
Three months in, no candidate has accepted. Six months in, two have started, both have left within four months for FAANG salaries. The firm has spent £40,000 on recruitment fees and management time, and is no further forward.
A competitor accountancy of similar size signed a £45,000 fixed-price 12-week engagement with a Cambridge-based AI consultancy, deployed a working document classifier, and is now running it in production. Same problem. Two paths. The first firm framed it as “let’s hire someone” and burned six months. The second framed it as “let’s ship the first project” and shipped in three.
The interesting question is not in-house vs outsourced. It is whether AI is becoming a continuing operating capability or a project-shaped need.
The choice you’re facing
In-house means a permanent AI specialist on payroll, accumulating institutional knowledge but carrying fixed cost and retention risk. Outsourced means a UK consultancy on fixed-price engagement, a contractor through a curated platform, or a vendor-led implementation through a partner programme. Neither default is right for every firm.
The 2026 UK context tightens the choice. 97% of organisations report at least one significant AI skills gap. 60% of expert AI vacancies sit in London and the South East. Demand for AI and data skills is forecast to grow over 30% through 2030. Hiring permanently signals long-term commitment but locks the firm into a fixed cost that outlives the project that prompted it. Outsourcing buys execution speed but trades institutional learning for vendor dependence. The question to answer first is what the AI work actually is, not who should do it.
When outsourcing is the right answer
Outsource when you have one major project and no continuing AI pipeline behind it. Outsource when the capability is specialist enough that permanent hiring cannot be justified. Fraud detection, computer vision quality control, AML screening, regulated-sector model governance. Outsource when time-to-value matters more than institutional knowledge.
UK consultancies ship 8 to 16 weeks for £30,000 to £80,000 fixed price for a defined pilot, and £80,000 to £300,000 for production-scale 24-week engagements. Faculty in London works across regulated sectors. BJSS sits in the Microsoft Azure ecosystem as a specialist AI partner. Cambridge Consultants leans deep-tech and responsible AI. Capco runs a financial-services AI lab. Featurespace, deployed in 180+ countries on adaptive fraud-detection ML, is the canonical “buy specialist capability rather than build it” example. For individual contractors, Toptal, YunoJuno and Crossover handle senior AI specialists at £700 to £1,200 per day London rates. YunoJuno is the cleanest path for UK firms because IR35 compliance is built in.
The advantage is speed. A well-scoped pilot moves from kickoff to working deployment in twelve weeks. An in-house hire takes three to six months to onboard before producing deployable work.
When the in-house hire is the right answer
Hire in-house when AI is becoming embedded across operations rather than sitting as a side project. Octopus Energy now has AI handling work equivalent to roughly 250 staff, with AI-generated email replies achieving 80% customer satisfaction against 65% for human staff. Monzo has built an in-house ML platform underpinning fraud, credit, automation and personalisation. In both cases AI is core enough to operations that internal teams are non-negotiable.
For a UK SME the threshold is more modest. The signal is a second or third AI project arriving on the roadmap, not the first. At that point the fixed cost of a permanent hire is amortised across a programme rather than absorbed by a single deliverable. A mid-level ML engineer is £75,000 base per Robert Half, near £110,000 fully loaded year one once 15% employer NI from April 2026, pension at 3%, recruitment fees of £5,000 to £8,000 and onboarding overhead are added in. A senior strategic hire in London commands £100,000 to £150,000 base. The hire only repays itself if the firm has work to keep them on for two to three years. If the project pipeline is one deliverable long, outsourcing is cheaper and faster, even at consultancy day rates.
What it costs to get wrong
The bad in-house hire is the dominant failure mode. 59% of UK organisations made one in the past year per TestGorilla. The pattern is consistent. Strong technical interview, impressive credentials, candidate cannot translate research into business value. Departure or redundancy at month eight to twelve. Total cost £150,000 to £200,000 once recruitment, salary, management time and exit handling are added together.
The failed outsourced engagement looks different but costs comparable money. Scope creep turns a £30,000 fixed price into £50,000. Vendor lock-in on a proprietary feature store or MLOps framework leaves the firm dependent on the consultancy in perpetuity. Insufficient knowledge transfer means the SME has paid for a working model it cannot maintain or evolve. Anthropic’s April 2026 shift to usage-based pricing for Claude enterprise illustrates how vendor pricing changes can destabilise outsourced engagements built on third-party APIs overnight, and the same pattern is now visible across the major model vendors.
The defensive move on either path is the same. Document everything. Insist on knowledge transfer as a contract line item, not an afterthought. Negotiate for portable, open-source technology rather than proprietary platforms. None of this prevents failure. It bounds what failure costs.
What to ask before you decide
Six questions decide it. Is this one project or the start of a programme? Is your data ready, twelve months of history, accuracy above 90%, completeness above 85%? Who is the internal sponsor regardless of who does the build? Have you negotiated explicit IP assignment and data processing agreements? Have you specified open-source, portable technology rather than proprietary platforms? What is the exit cost if the engagement ends or the hire leaves?
The pragmatic answer for many UK SMEs in 2026 is sequence, not either-or. Outsource first to validate that AI delivers value in your context. Hire one senior internal lead at £80,000 to £120,000 once the second or third project is on the roadmap. Use that lead for judgement and architectural oversight, and keep outsourced execution capacity that scales up and down with demand. The ICO’s guidance on AI and data protection applies regardless of which path you take, so build the governance overlay early.
That is the pattern Octopus Energy followed, and the one many UK accountancy and professional-services firms will recognise as workable. If you cannot answer the first question with conviction, you are not ready to hire or to procure. You are ready to scope.
