Nine in ten retailers plan to raise their AI budget this year. That was the headline finding from NVIDIA's latest State of AI in Retail and CPG survey, and it matches what we see across established brands right now. The board has signed off. The pilots are running. The vendors are in the building. A year from now, most of that money will have produced a forecasting dashboard nobody quite trusts, a support bot the team quietly routes around, and a merchandising engine that recommends what already sells.

The reason is almost never the model. It is the layer underneath it. Every operational use of AI worth funding, demand forecasting, inventory intelligence, support automation, dynamic merchandising, runs on your first-party data. When that data sits fragmented across a POS system, a CRM, a marketplace export and three spreadsheets that only one person understands, no model on the market can rescue it. So the AI investment worth making first is not the model. It is the data foundation that decides whether any model you buy afterwards actually works.

This is not a reason to slow down. The shift towards AI-run operations is real, it is early, and the brands that get the foundation right now will compound an advantage that is genuinely hard to copy. But being early only pays if you are also right about where the money goes.

The forecasting problem hiding on your balance sheet

Inventory distortion, the combined cost of stockouts and overstock, reached roughly 1.73 trillion dollars globally in 2025 on IHL Group's numbers, about 6.5 percent of all retail sales. For a single brand that translates into capital tied up in the wrong stock and revenue walking out the door every time something popular is unavailable. It is one of the largest recoverable costs in the whole business, and it barely gets discussed at board level because it is spread thinly across every SKU.

This is exactly the problem AI forecasting is built to attack. Machine-learning demand models read historical patterns, seasonality, price elasticity and external signals together, rather than one buyer's judgement extrapolated in a spreadsheet. McKinsey's work puts the accuracy gain at 10 to 20 percent over traditional methods, and leading operators have cut inventory holdings by 20 to 30 percent while holding service levels. Those are not marginal numbers. On the capital most brands have locked in stock, that is a serious return.

Here is the part the vendor demo skips. Those results come from models fed clean, connected, well-labelled data. Feed the same model your real data, with its duplicate SKUs, its inconsistent product taxonomy, its returns that were never reconciled and its channel exports that disagree with each other, and the forecast lands somewhere between useless and actively misleading. The intelligence was never the constraint. The inputs were.

Why the model is rarely the bottleneck

The market has quietly commoditised the model. Forecasting, recommendation and support-automation capability that was proprietary and expensive three years ago is now available off the shelf, often built into the platform you already run. The hard, differentiating work has moved down a layer, to whether your business can supply those tools with data they can actually use.

Most established brands underestimate how much of that work is still undone. Customer records exist in one system, order history in another, stock in a third, and returns somewhere no analytics tool has ever looked. Each was built for its own job and never designed to talk to the others. That is fine when a human is stitching the picture together once a week. It is fatal when you ask a model to make thousands of automated decisions a day on top of it. Garbage in, garbage automated, at scale and at speed.

This is why we tell brands to treat the plumbing as the project, not the preamble. Connecting your systems so that data flows cleanly between them is unglamorous, it does not demo well, and it is the single thing that determines whether the AI on top returns anything. When we rebuilt the data and systems integrations behind a client's operations, the value was not the integration itself. It was that every tool downstream, forecasting included, finally had one clean version of the truth to work from. We saw the same pattern building the platform behind Astound, where the real gain came from connected data rather than any single clever feature.

First-party data is the fuel, not just the moat

Most boards now understand first-party data as a privacy story. As third-party signal degrades and acquisition costs climb, the argument goes, the data you own becomes your most durable asset. That is true, and it undersells the point. First-party data is not just the thing that protects your marketing. It is the fuel every operational AI system you are about to buy will run on.

That reframing changes how you value the work. Cleaning your product taxonomy, reconciling returns, connecting your POS to your CRM, these read as hygiene projects that never make the priority list. Seen as the precondition for every AI investment on the roadmap, they move to the top of it. You are not tidying data. You are building the thing that decides whether the next three years of AI spend produce a return or a graveyard of half-used tools.

It also changes who owns it. Data foundation is not an IT tidy-up to be delegated and forgotten. It is a commercial decision about where the business is investing to compete, and it belongs in the same conversation as the AI budget it underwrites. Funding the models without funding the foundation is the most common way we see AI money wasted.

Where the next AI pound should go

The sequence matters more than the shopping list. Fund the data layer first: connect the core systems, agree one product taxonomy, reconcile the numbers that disagree, and get to a single clean version of your operational truth. This is the least exciting line on the slide and the one that makes every other line work.

Then, and only then, fund the operational models in order of recoverable cost. Demand forecasting and inventory intelligence usually sit at the top, because inventory distortion is often the largest fixable number in the business. Support automation that genuinely resolves queries rather than deflecting them tends to come next. Dynamic merchandising and personalisation follow, once there is clean behavioural data to drive them.

Be equally clear about what is theatre. An AI tool bolted onto disconnected data will produce confident outputs that are quietly wrong, which is more dangerous than an honest spreadsheet because people trust it. A model chosen for how it demos rather than the decision it improves is a cost, not an asset. And any AI project whose business case cannot name the specific number it will move does not have a business case, it has a budget line looking for a justification.

The brands that will look prescient in three years are not the ones buying the most AI this year. They are the ones spending in the right order, foundation before model, recoverable cost before novelty. That is a less thrilling story than the one the market is selling, and it is the one that actually pays.

If you are weighing up where your AI investment should go, and whether the foundation underneath it is ready to carry the tools you are being sold, that is exactly the kind of decision we are happy to think through with you.