Skip to content

AI Integration

Most of the value I bring to AI integration is separating the cases where it genuinely earns its place from the far larger number where it’s theatre — a feature added because the market expects to hear “AI”, solving a problem the business doesn’t have. When AI does fit, building it in well is real, careful engineering. When it doesn’t, the honest move is to say so and build the thing that actually helps. People bring me AI problems expecting a yes; the value is that I’ll give them the true answer, and then build it properly if it’s a yes.

The problems I solve here

The first is the genuine fit: a real problem where AI is the right tool. Extracting structured data from messy documents, classifying or routing large volumes of text, summarising at a scale no person could, surfacing answers from a body of a company’s own content. These are real wins, and the job is to build the integration so it’s reliable, owned, and doesn’t quietly cost a fortune or leak the business’s data — which is most of the actual engineering.

The second is the AI feature that was added for show and is now a liability. Something bolted on to look current that produces unreliable results, costs more per use than anyone modelled, or handles failure badly — confidently returning wrong answers with no fallback. The work is to make it trustworthy or to be honest that it shouldn’t be there.

The third, and most common, is the business about to spend on AI because everyone else is, without a clear problem it solves. They’ve been told they need an AI strategy. The most valuable thing I do there is find whether there’s a real problem AI fits, and if there isn’t, say so plainly rather than building an expensive distraction. That straight read is worth more than any feature.

How I think about AI integration

The first principle is to solve the real problem, not the impressive-looking one. The pressure to add AI is enormous and mostly disconnected from whether it helps. The right question is never “how do we add AI?” — it’s “what’s the actual problem, and is AI genuinely the best tool for it?” Often it isn’t: a simple rule, a better data model, or a plain piece of software solves it more cheaply and reliably than a language model that’s right most of the time. AI is a tool with a narrow set of jobs it’s genuinely best at, not a strategy, and treating it as one is how money gets burned on demos.

The second is that AI integration is exactly where owning your data matters most. The valuable AI applications run on your own data — your documents, your domain, your operation. That makes consolidating and owning your data the prerequisite, not an afterthought, and it raises a real question about where that data goes. Sending your core business data to a third-party model is a decision with cost, privacy, and lock-in consequences that deserve a clear-eyed look, not a default yes. I treat AI as a capability you build into systems you own, on data you control, with deliberate choices about what leaves your walls — not as another rented platform you pour your operation into.

The third is that an AI feature lives or dies on how it handles being wrong. These models are confidently incorrect a meaningful share of the time, and an integration that takes the output at face value will ship wrong answers as fact. The real engineering is in the guardrails: knowing when to trust a result, where a human stays in the loop, what the fallback is when the model fails. The demo is easy and the reliability is the whole job, and skipping it is how AI features become liabilities.

How I’d actually do it

I’d start from the problem, never the technology. Find a real, specific job — say, pulling structured data out of incoming documents that people currently re-key by hand — confirm AI is genuinely the best tool for it rather than a rule or simpler software, and build that one integration well. Reliable, with proper handling of low-confidence results, costs understood up front, and running on owned data. One real win, value banked, rather than a broad “AI transformation” that impresses in a meeting and delivers little.

I’d keep it lean and owned: the model is one component inside systems you control, not the centre of a new dependency you can’t see into. The guardrails get the real attention — what happens when the model is wrong matters more than what happens when it’s right, because the right case takes care of itself. Built this way, AI becomes a sharp, trustworthy capability in a system you own, not a fragile feature bolted on for appearances.

Where this connects

This is one of the specialisms I’m hired for. Useful AI runs on owned, consolidated data, which makes this the natural next step beyond getting your databases right, and the integration work itself sits close to API integrations — wiring a model in reliably is the same discipline as wiring in any other service. For the strategic question of whether AI fits your business at all, that’s the AI consultancy engagement; this page is the hands-on building of it once the answer is yes.

When you need a team instead

A large AI capability built into a product and run continuously — at scale, with ongoing tuning and operation — is team work. That’s my agency, Digital Royalty, with the crew to build and run it. My lane is the honest call on whether AI fits, and the lean, well-built first integration. I’ll tell you straight which one your situation needs — including when the answer is that you don’t need AI at all.

Who this isn’t for

If you want AI added to your product because the market expects it, regardless of whether it solves anything, I’m the wrong call — I’ll ask what problem it’s for first. I’m worth hiring when you want the honest answer about where AI genuinely helps, and the engineering to do it properly when it does.

Talk to me

If you’re weighing an AI feature, or already shipped one you don’t trust, tell me about your problem and I’ll tell you whether AI is the right tool for it — and build it properly if it is.