In April 2026 Papa John’s launched Lou, its in-house AI pizza assistant. Three days later we wrote about the five voice AI startup ideas it validated for the wider UK market. The question that landed in our inbox forty-eight hours after publishing was the same from a dozen different UK businesses: should we build an AI agent like Lou ourselves, or buy one off the shelf?
It’s the right question, and almost no-one in the UK apprenticeship space is answering it properly. Let’s try.
The Default Answer Is Wrong
The reflex answer most consultants give is “buy unless you have a strong technical reason to build.” That was good advice in 2022. It’s the wrong starting point in 2026, because the cost of building a competent AI agent has collapsed by roughly 40x in three years.
In 2022 building a useful voice agent meant a six-figure engineering project, a vector database vendor contract, custom prompt engineering, a model fine-tuning budget, and a dev-ops team to run inference. Today it means an ElevenLabs Agents subscription at £0.06 per minute, a no-code workflow tool like n8n or Make, and one person on your team who knows how to wire CRM webhooks together.
That cost collapse changes the build-vs-buy calculation completely. Now the right question isn’t “can we afford to build?” — it’s “do we have someone who knows how to build, and do we want to keep what we build inside the business?”
“Two years ago I would have told nine out of ten employers to buy. Today I tell six out of ten to build — not because the technology is more accessible, which it is, but because building means you keep the capability. When the next wave of AI hits, you’ll have people in-house who’ve done it before.”
The Decision Framework
Here are the seven questions we walk every employer through. Score each from 1 (low) to 5 (high). Higher total score → lean build. Lower total → lean buy.
| Question | Score 1 (lean buy) | Score 5 (lean build) |
|---|---|---|
| 1. Strategic moat? Does this AI tool give you a competitive advantage? | Generic, every competitor has it | Differentiating, hard to replicate |
| 2. Data sensitivity? Will it touch confidential customer or financial data? | Public-facing, low sensitivity | Regulated data (PII, financial, health) |
| 3. Integration depth? How tightly does it need to plug into your stack? | Standalone tool | Deep integration with bespoke systems |
| 4. Workflow uniqueness? Is your workflow standard or specific to you? | Industry-standard process | Bespoke, hard to describe to vendors |
| 5. Existing capability? Do you already have AI-literate staff? | No-one technical in-house | Strong internal AI/data team or apprentices in training |
| 6. Volume of use? How heavily will the tool be used? | Light, occasional use | Core workflow, used daily by many staff |
| 7. Roadmap horizon? How long do you need this capability? | Six-month tactical fix | Multi-year strategic asset |
Total score 25 or higher: build. 15 or lower: buy. In the middle: probably hybrid (more on that below).
Five UK Examples Worth Studying
1. Build — Papa John’s and Lou
Score: 31/35. Strategic moat (personalised group ordering), high integration depth (POS + delivery + loyalty), high volume (every UK store), multi-year roadmap. Papa John’s built Lou in-house because the workflow IS the differentiation.
The lesson: when AI is the product, build it. When AI is plumbing, buy it.
2. Buy — A 30-Person UK Marketing Agency Using ChatGPT Team
Score: 12/35. No moat (every agency uses ChatGPT), no regulated data, no integration depth, generic workflows. The right answer here is a Team licence at £25/seat/month and structured prompt training for staff.
The lesson: when AI is a productivity layer, buy it. Don’t reinvent.
3. Hybrid — A Mid-Market UK Insurance Broker
Score: 22/35. Buys Microsoft Copilot for general productivity. Builds a custom claim-triage agent on top of Azure OpenAI because their underwriting workflow is bespoke and FCA-regulated. Hires through the AI & Automation L4 apprenticeship to grow the in-house team that maintains it.
The lesson: most UK SMEs end up here. Buy the boring layer, build the differentiating one, train people internally to bridge them.
4. Build — UK Estate Agent Group with 12 Branches
Score: 27/35. The lettings phone load is enormous, off-the-shelf voice agents don’t integrate with their bespoke Reapit setup, and the brand voice matters. They built a voice agent on ElevenLabs in eight weeks using two members of staff already mid-apprenticeship.
The lesson: building doesn’t require a software house when the platforms have done the hard part.
5. Buy — A 5-Partner UK Solicitors Firm
Score: 9/35. Regulated, but small volume, no in-house tech team, no appetite to maintain custom software. They subscribe to a vetted UK legal AI platform that’s already SRA-aware. Right answer.
The lesson: regulated ≠ build. Sometimes regulated means buy from someone who’s already done the compliance work.
The Real Cost Comparison
Here’s what most build-vs-buy articles get wrong: they compare licence costs without accounting for the talent question. A £500/month off-the-shelf SaaS isn’t actually cheaper than a £0.06/min self-built agent if the SaaS doesn’t fit and you waste six months in vendor purgatory.
| Cost line | Buy (typical SaaS) | Build (modern stack) |
|---|---|---|
| Software / infrastructure | £500–£5,000/month per use case | £0.06/min usage + £100–£300/month tools |
| Implementation | £5k–£50k one-off | £0 (apprentice time, internally absorbed) |
| Ongoing maintenance | Vendor handles (locked in) | ~5 hours/week of apprentice time |
| Customisation | Limited, vendor-priced | Unlimited, internal |
| Talent built | Zero | Permanent in-house capability |
| Vendor lock-in | High | None |
The real cost of buying isn’t the licence fee — it’s the missed opportunity to grow the AI capability inside your business. The real cost of building isn’t the infrastructure — it’s finding someone who knows how.
The Talent Question (and Why Apprenticeships Solve It)
Here’s the honest answer to “should we build?” — yes, if you have someone who can. The barrier to building competent AI agents in 2026 isn’t money or compute. It’s skills.
The roles that build modern AI agents in UK SMEs aren’t senior ML engineers on £120k. They’re mid-career operations people, project managers, marketing analysts and finance assistants who’ve learned prompt engineering, agent orchestration, basic API workflow and AI governance — the exact stack covered by the AI & Automation Specialist Level 4 apprenticeship. For organisations that want a faster route or already have some technical staff, the accelerated 8-month version compresses the same content, and the AI Prompting Accelerator is a shorter accelerator focused specifically on the prompt-engineering layer.
“Every employer who tells me they need to hire an AI engineer is, eight times out of ten, actually looking for an AI-literate operations person. The L4 apprenticeship trains your existing staff into that role — funded by the levy, not bought from outside.”
The Hybrid Path Most UK SMEs End Up On
If your score landed between 15 and 25, you’re probably in the hybrid bucket. That’s the most common answer for UK businesses between £5m and £50m turnover, and it looks like:
- Buy the productivity layer — Microsoft Copilot, ChatGPT Team, Google Workspace AI for the whole business (we run dedicated Microsoft Copilot and Google Gemini short courses for the staff using them)
- Build the differentiating workflow — the one specific agent or integration that maps to how YOU work
- Train the team that connects them — via apprenticeships, so the cost is levy-funded
This is what TESS Group itself does internally, and it’s what we recommend to most clients.
What to Do This Week
Before another vendor pitch lands in your inbox:
- Score yourself against the 7 questions above. Be honest about question 5 (existing capability).
- Identify the one workflow where AI would matter most for your business this year. Not five — one.
- If your score is 25+ on that workflow, before you start hiring, look at your existing team and ask: who could do this with the right training?
- If your score is 15 or below, just buy. Move on.
- If your score is in the middle, plan for hybrid. Buy the layer, build the workflow, train the people.
How TESS Group Helps with the Build Side
If your decision was “build” or “hybrid,” the question becomes: where do the building skills come from? Most UK SMEs grow them via one of these routes.
| What you need | Best fit | Length |
|---|---|---|
| Mid-career staff trained as full AI builders | AI & Automation Level 4 apprenticeship | 15 months, levy-funded |
| The same content compressed | AI & Automation L4 Accelerated | 8 months, levy-funded |
| Specifically the prompt-engineering layer | AI Prompting Accelerator | Short programme |
| An advanced AI build skill upgrade | AI Advanced Accelerator | Short programme |
| A team-wide bootcamp on generative AI | Generative AI Bootcamp | Multi-day |
| Building AI-ready operating habits across staff | Building AI-Ready Teams | 1 day |
| Workflow automation skills specifically | Automating Workflows short course | 1–2 days |
| Mastering Microsoft Copilot or Google Gemini | Microsoft Copilot / Google Gemini | 1 day each |
Hybrid in approach as well as in policy: we run apprenticeships and short courses in parallel so different parts of your team can move at different speeds. Apprenticeships vs short courses explains how to combine them, and the programme finder gives a tailored recommendation.
Frequently Asked Questions
What is the cost difference between building and buying AI in 2026?
For most UK SME use cases, building costs roughly £100–£400/month in infrastructure plus internal time, versus £500–£5,000/month for off-the-shelf SaaS. The bigger difference is talent: building grows AI capability inside the business, while buying does not.
Do I need software developers to build an AI agent?
Increasingly no. Modern platforms like ElevenLabs Agents, n8n, Make and Zapier let non-developers build production AI agents using prompts and visual workflows. The skill needed is prompt engineering and integration design, which is exactly what the AI & Automation Level 4 apprenticeship trains.
How long does it take to build an AI agent in-house?
For most narrow use cases (a voice receptionist, a sales qualifier, an order-taker, a multilingual support agent), 4–8 weeks from start to production with one or two internal staff working part-time. Bigger transformations (replacing a customer-facing system end-to-end) take 3–6 months.
Is it safe to build AI in regulated industries like financial services or healthcare?
Yes, with proper governance. UK regulated sectors can build AI safely if there are documented controls around data handling, consent, audit trails, and human-in-the-loop oversight for any decisions touching regulated activity. AI governance, ethics and risk are core modules in the AI & Automation Level 4 apprenticeship for exactly this reason.
What if my AI build fails — is buying safer?
Both can fail. Off-the-shelf SaaS fails when it doesn’t fit your workflow and you waste 6 months in vendor purgatory. In-house builds fail when nobody owns the maintenance after launch. The mitigation in both cases is the same: have AI-literate people internally who can either choose the right vendor or maintain the in-house build.
Can the Apprenticeship Levy fund AI capability building?
Yes. The AI & Automation Specialist Level 4 apprenticeship is fully funded through the Apprenticeship Levy for UK employers. Existing staff can be enrolled — they don’t need to be new hires — and the apprenticeship explicitly trains the skills needed to build, ship and govern AI tooling internally.
Build AI Capability In-House
The AI & Automation Specialist Level 4 apprenticeship trains your team to build, ship and govern AI tooling. Fully funded through the Apprenticeship Levy.
Book a Free Discovery Call