What's Live in Retail AI Right Now (And What Isn't Yet)

AI budgets are under more scrutiny than they were twelve months ago, and rightly so. A lot of last year's pilots didn't survive contact with a P&L review. As Forte's Chief AI Officer, most of what crosses my desk isn't "what's technically possible", but the much harder question of what's actually deployed, measured, and paying for itself versus what's still a promising demo.

Conflating those two right now is an expensive mistake, and I think it's one a lot of retail organizations are quietly making.

Three different maturity levels, three different risk profiles

Working through a full retail business end to end recently, from demand forecasting through to the shop floor, the maturity gap was clearer than I expected going in.

Some AI is already running quietly in production, has been for a while, and just doesn't get much airtime because it's not customer-facing: demand forecasting that adjusts as sales data comes in, dynamic pricing engines re-evaluating markdown through the day. This is low-risk to invest in, largely because the organizations running it have already worked out the failure modes.

Some AI is earlier-stage but moving fast: computer vision flagging shelf gaps, systems predicting where a delivery will actually land versus the promised window. Promising, increasingly production-ready, but still worth treating as a pilot with a defined success metric rather than a foregone conclusion.

And some AI is genuinely emerging: agentic commerce, where a shopper's own AI assistant does the searching, comparing, and eventually the transacting, is real and moving quickly, but it's earlier than most of the current conversation admits. It also raises a question the industry hasn't settled: once an agent is doing the buying on a customer's behalf, who actually owns that customer relationship? the retailer, the platform the agent runs on, or neither?

Why the distinction matters more than the capability itself

The mistake I see most often isn't picking the wrong AI use case, but picking the right one at the wrong maturity level for the organization's risk appetite. A production-grade forecasting model and an experimental agentic-commerce pilot both count as "AI investment" on a budget line, but they carry completely different risk, timeline, and governance needs. Treating them the same is how a promising pilot either gets killed too early for not delivering production-level ROI, or a production-ready capability gets stuck in pilot purgatory for years out of excess caution.

If there's one filter I'd suggest applying before the next AI conversation in your organization: ask not just "does this work" but "has this been proven at the scale and reliability we'd need to actually depend on it". And be honest about the answer.

I'll be discussing this live with Puma's Falko Saft at Retail Week's AI Summit on September 30: "AI agents: your business's biggest CX tool." If you're trying to separate what's ready to fund from what's still roadmap in your own organization, I'd like to compare notes. Let's find time in London that week.