Before you overlay AI, build the single source of truth.

Across twenty years bridging business and technical teams — retail, fashion, finance, and now hospitality and education — the same pattern keeps turning up. Leaders want the same thing: AI, pointed at our data. Forecasting, anomaly detection, an assistant that actually knows the numbers. The ambition is right. The snag is almost always identical: there is no single "our data" to point it at. There are forty places the data lives, and no two of them agree.

This is the part that never makes the slide. Before AI is worth a penny, you need a single source of truth: one trustworthy place where the numbers live, are defined once, and reconcile. In practice that means two unglamorous jobs done properly — standing up a data warehouse (or a lighter equivalent) and connecting the software already running the business. Skip it and AI just hallucinates faster, over worse data.

Find the system of record — retire the spreadsheet.

The first job isn't technical, it's archaeological. For every important number — sales, stock, margin, headcount — you have to find the system of record: the one place that is genuinely authoritative. In a healthy setup that's a proper application. In most growing businesses, it's a spreadsheet that quietly became load-bearing.

Spreadsheets are where single sources of truth go to die. They get copied, emailed, edited offline, and within a week there are three versions and a chain titled "FINAL_v4". Part of the work is simply naming, for each metric, which system wins — then moving the ones living in spreadsheets into systems that can actually be trusted and integrated.

AI doesn't fix a data mess. It industrialises it.

Three realistic routes — with costs and timelines.

"Build a data warehouse" sounds like one decision. It isn't. There's a sensible route for a five-person firm and a very different one for a multi-site group, and the gap between them is enormous. Here are the three I'd actually put in front of someone, with indicative 2026 costs and how long each takes to stand up.

Indicative only, 2026, GBP approximations from USD list prices; ex-VAT and excluding internal staff time unless stated. Confirm against your own systems and volumes before budgeting. Sources at the foot of the page.

The build behind the multi-site example.

To make the top row real: a group of 1,400–1,800 sites or stores taking both till sales and app sales, with annual revenue in the region of £180–300m, running on disconnected systems and load-bearing spreadsheets. The brief was one trustworthy place for the numbers — and to do it with a small team and no in-house software engineers.

The shape that worked: a Snowflake warehouse with a layered (medallion) model, so raw feeds are cleaned and conformed into one governed, gold-level layer where a "sale" means exactly one thing — whether it came from a till or the app. The finance, CRM and sales systems were integrated alongside the other operational tools, and reporting sits on top in Power BI and Metabase, reading from that single model rather than from forty exports.

Why the integration is the hard part

The warehouse is the easy bit. The graft is in the connections: integrated tills behave differently from non-integrated ones, app sales arrive on a different clock to till sales, and every system has its own idea of a date, a store, and a transaction. Reconciling those into one honest model — so finance and operations finally quote the same figure — is most of the job, and most of the value.

What it costs — by the benchmarks, not by me

Put a figure on it without taking my word for it. Cross-industry, firms spend roughly 3.6–5.7% of revenue on IT (Deloitte's CIO benchmarks), with retail and hospitality typically at the lower end; broader digital-transformation programmes run 5–15% of revenue (Deloitte). A foundational data-platform build is one slice of that, delivered over roughly 12–18 months for a multi-site estate (industry implementation timelines). Independent data-infrastructure studies put a complete mid-market capability — software plus team — at hundreds of thousands to over a million per year.

Stack those together for a group turning over £180–300m and the all-in cost of reaching a single source of truth lands in the low single-digit millions — on the order of £2–3m, or about 1% of a single year's revenue. It sits at the upper end of those ranges precisely because of the number of sites and systems being joined up. The licences are a rounding error in that total; the cost is people and integration.

Then — and only then — the AI.

Once there's one trustworthy model, AI suddenly has somewhere to stand. The questions that were impossible — which sites are quietly underperforming, what's the real margin by channel, where is demand moving — become answerable, because the data underneath is finally consistent. The AI didn't get smarter. The ground beneath it got solid.

That's the order that matters, and the one most businesses get backwards. The exciting layer is the last layer. The single source of truth comes first — whether yours costs five thousand pounds or three million — and it's the difference between AI that works on Monday morning and a demo that falls over the moment someone checks the number.