What a Year of Applied AI at a UK Fabricator Actually Taught Us

A year ago, we started working with a specialist UK steel fabricator that had never had a website, ran entirely on word of mouth, and stored most of the business’s operational knowledge in the managing director’s head.

The work that followed became a custom manufacturing system built iteratively over twelve months, with AI workflows layered in progressively as the platform matured. It’s the kind of engagement we learn the most from, long enough to see the second-order effects, deep enough to get beyond the demos, collaborative enough to hear honest feedback from the people using what we’d built.

Here’s what the year actually taught us. No vendor pitch, just the pattern we’d apply to anyone starting down a similar road.

Lesson 1: Build the system before you add the AI

The biggest operational wins in the first six months had nothing to do with AI. They were boring software: a workflow engine that modelled how jobs flow through the shop, integrations that let different systems share data, dashboards that showed what was happening in real time.

AI got layered in once the underlying platform was solid. Without that foundation, every AI workflow would have had to build its own plumbing, and the project would have collapsed under its own complexity.

What we’d tell a new client: your first investment should almost always be the system that runs your business, not the AI that sits on top of it. If your processes live in WhatsApp, email, and memory, fix that first. AI amplifies what exists. Amplifying chaos just makes the chaos faster.

Lesson 2: Start where the data is already structured

The first successful AI workflow we shipped was supplier invoice reconciliation. It worked because invoices are structured, purchase orders are structured, and the matching rules were known. The payback was measurable in hours per week from day one.

The second workflow was reading technical cutting list PDFs and generating paint notes for the coating supplier. Slightly harder because the PDFs had multiple layouts, but still fundamentally a structured-data problem.

Only later did we tackle the harder stuff, vision models for quality checks, conversational assistants for the shop floor. Those took longer, needed more iteration, and delivered their value more slowly.

What we’d tell a new client: structured data first, always. Invoices, purchase orders, specs, orders, CRM records. Save the unstructured stuff (images, free-text, conversational) for after the structured wins have paid for the project.

Lesson 3: 80% of the work is context, not AI

Every AI workflow we built took longer than the underlying AI model itself needed. The AI part was usually a few days. The remainder, understanding the real-world context, designing the confirmation flows, handling edge cases, building integrations, testing with the actual team, was weeks or months.

The conversational shop-floor assistant is the clearest example. Building a “chatbot that answers questions about jobs” took about two weeks. Making it safe enough to let it start, stop, or move jobs through production stages took another six, because the real engineering was in the confirmation flows. What happens if the radio picks up a wrong word? What happens if two people are logged in to the same terminal? What happens if a job number is ambiguous? None of those are AI problems. They’re systems problems that AI projects uniquely have to solve.

What we’d tell a new client: when you’re scoping an AI project, budget most of the time for context, not AI. Good consultancies know this. Bad ones quote 80% AI and 20% “integration” and then run over.

Lesson 4: Humans stay in the loop for longer than you think

Every AI workflow we deployed went through a shadow-mode period, running alongside the manual process, output compared but not acted upon. In every single case, the period needed to be longer than the initial estimate.

Not because the AI was wrong. Because the trust takes time. A team that’s been processing invoices manually for five years doesn’t hand over to AI on day one, no matter how accurate the AI is. They need a few weeks of watching it get things right. They need to see how it handles the weird invoices (the one that’s in German, the one with the wrong line break, the one with a payment term that’s different from the PO). They need to know what happens when it fails.

What we’d tell a new client: budget for trust. Plan for 2-4 weeks of shadow operation per workflow. Don’t rush the go-live. The value compounds from the trust, not from the technical readiness.

Lesson 5: The best AI projects give time back to people

Every workflow we deployed was designed to reduce time spent on tedious, repetitive, or error-prone tasks. In every case, the team used the saved time on work that was higher-value, the judgement work, the customer-facing work, the strategic work, the work that’s actually hard to hire for.

The managing director ended up with measurably more time for the parts of the business only he can do. The office team stopped being buried in admin and started having capacity for proactive work. The shop floor spent less time looking up information and more time making things.

This wasn’t a side effect, it was the explicit goal from day one. Applied AI at an SMB shouldn’t be about shaving labour costs. It should be about reallocating human time from tedious to valuable.

What we’d tell a new client: measure success in time given back, not costs saved. The business impact is the same, but the framing matters. It keeps the project honest and keeps the team bought in.

Lesson 6: Computer vision is harder than document reading

We added AI-assisted quality checking (photographs analysed for defects) in the second half of the year. It took longer than any other workflow to get right, and the accuracy on day one was worse than a trained human inspector.

Iteration closed most of the gap. The model had to learn what your specific inspectors look for, not what a generic “quality AI” considers a defect. And even then, the system ended up as an assistant to the human inspectors, not a replacement. Flagged items get a second look. The humans stay in charge.

What we’d tell a new client: computer vision projects are real but they’re not first projects. Budget 2 to 3 times the timeline of an equivalent document-processing project, and plan for iteration with the team who actually does the job.

Lesson 7: Integration is the hidden cost

Most AI demos happen in isolation. The real projects happen inside the mess of whatever systems the business already runs, accounting software, ERPs, email, CRMs, shared drives, spreadsheets.

We lost count of the number of times an AI workflow was 90% done and the remaining 10% was wrestling with how to get data into or out of some mainstream UK business tool. This is normal. It’s also usually underestimated in project scoping.

What we’d tell a new client: ask any AI consultancy explicitly about integration. “How will this connect to our accounting software?” isn’t a dumb question, it’s usually the question that separates on-time projects from over-budget ones.

Lesson 8: The MD still walks the shop

A year into this work, the managing director still walks the shop every day. He just walks it with better information in his pocket, live job status, current bottlenecks, cash tied up in work-in-progress, quality issues flagged since yesterday.

The work didn’t replace the judgement at the top of the business. It amplified it. That’s the version of AI we think is worth building.

What we’d tell a new client: AI is a tool for decisions, not a replacement for them. If an AI consultancy pitches you a system that removes the human from the loop entirely, be very careful about what you’re buying.

If we started again

If we were dropping the first pin at day zero with everything we know now, we’d:

1. Start with the website and marketing foundation (this did actually happen).
2. Build the core system in months 1-3, not trying to layer AI in yet.
3. Add the first AI workflow on structured data (invoices) in month 4.
4. Add a second structured-data AI workflow (technical document extraction) in month 6.
5. Shadow-mode everything for longer than feels necessary.
6. Tackle the harder stuff (vision, conversational) after month 9, only once the team was comfortable with the earlier wins.

Roughly the path we took, with the minor exception of occasionally being too ambitious too early. The lesson: your first AI project should feel almost boring. If it feels exciting, it’s probably too risky.

What Stephen says

“We never had a website before. Now we’re getting enquiries from Google that we never would have seen.”

, Stephen Chappell, MD, Kingsland Fabrications

The website gets the enquiries. The system, with AI woven in, keeps the shop running smoothly enough to take them on.

Frequently asked questions

Could we do this work for a non-manufacturer? Yes. The patterns (structured data first, integrations are the hidden cost, humans in the loop for longer than you think) apply across sectors. We’ve applied the same principles to trades, retail, and professional services clients.

How much did the full year cost? Bespoke AI-powered platform work at this scale runs in the £15,000 to £50,000+ range, rolled out iteratively in phases over months so clients see value early. We quote every phase upfront with fixed pricing.

How long until a UK SMB sees a return? Focused first projects (invoice processing, document extraction) typically pay back in 1 to 3 months. Full platform work sees compound returns building from the first live release onwards.

Can you write a retrospective like this for our business? Every client project ends with a written retrospective covering what worked, what didn’t, and what to do next. It’s part of how we close out engagements honestly.

If you’re a UK manufacturer, trades business, or SMB wondering what a year of applied AI could actually change in your business, book a scoping call or read the Kingsland Fabrications case study for more on the partnership described here.