The short version
AI-ready commerce comes down to four things the models can’t supply: product data, systems integration, workflow, and how the organization works.
Nobody at the table spent the evening debating whether the models are capable enough to be useful. Four requirements came up again and again. Each gets a full section below, with three questions to test it in your own business.
-
Product data becomes commercial infrastructure. When an AI assistant shops for someone, it recommends a short list, and a shortlist has no page two. If your product data is thin, you aren’t on it.
-
Adoption is an operating-model problem. Most strategies stop at tools, pilots and training. The companies getting real value changed how work gets done, and engineering alone can’t fix this one.
-
The prototype is no longer the hard part. Working software can now come from anywhere in the business. The bottleneck is everything after: security, testing, integration, governance and deployment.
-
Operations become continuously supervised. The room spent more time on agents that watch: catching broken feeds, price errors and phantom inventory. Teams move from analyzing what went wrong to supervising systems that detect it continuously.
Nobody argued about the AI
We put a group of technology and commerce leaders around one table in New York for an evening, with a single rule: nothing said in the room would be attributed to the person who said it. No presentation, and nothing to buy.
The striking thing was what never got discussed. Nobody spent the evening debating whether the models are capable enough to be useful. The conversation was about everything underneath: what has to be true before a demo becomes something the business runs on.
The companies that win the next few years will be the ones that close that gap. And you can’t buy your way across it.
1. Product data becomes commercial infrastructure
The familiar pattern of a query, a grid of results, and a page of sponsored placements is being replaced by something narrower. When an AI assistant shops for someone, it recommends a short list. A shortlist has no page two. Either the product is on it or it doesn’t exist for that shopper.
The constraint the table kept returning to wasn’t model quality. It was attribute depth. Feeds built for a keyword index carry the fields a keyword index needed, which is a fraction of what an AI system needs to decide whether a product fits a described situation. The emerging remedy is a two-step operation: enrich the attributes, then rewrite the outbound feed so the enrichment survives the trip.
The tension nobody had solved: a meaningful share of retail media revenue depends on paid placement inside the results page, and the results page is the thing being replaced.
What to check in your business
- Take ten products a customer would describe by situation rather than by name. Can your feed answer the situation from attributes alone?
- Pick a competitor. For how much of your catalog could an AI assistant compare your product with theirs, spec for spec? Where their feed lists an attribute and yours doesn’t, you lose by default.
- Run a scenario in which AI recommendations account for 20 percent of your discovery traffic within eighteen months. Which revenue line is cannibalized, and who owns the response?
2. Adoption is an operating-model problem
The most consistent view of the evening was that adoption isn’t primarily a tooling problem. Licenses, pilots and training create access without changing how work gets done. The discussion repeatedly linked deeper adoption to peer learning, clear expectations and redesigned workflows, and this is the one requirement engineering alone can’t fix.
Two mechanisms came up repeatedly. The first is grassroots: a self-selecting group of early adopters who meet, share what worked, and teach their peers directly. The second is blunter: put AI usage into the metrics people are measured on. Optional tools stay optional. Measured behavior becomes ordinary behavior.
What to check in your business
- Name the ten people in your organization who are genuinely fluent. Are they connected to each other, or are they ten isolated experiments?
- Is anyone measured on this, or is it entirely voluntary and therefore subject to the workload of the week?
- When someone finds a technique that works, how does the person three desks away learn about it? If the answer is a Slack channel nobody reads, that’s your adoption ceiling.
3. The prototype is no longer the hard part
Working software can now come from anywhere in the business. A commercial lead with a clear idea of the workflow can describe it and get something that runs.
The bottleneck sits immediately downstream. Enterprise IT now receives a stream of functioning prototypes it didn’t scope and has no established path to secure, test, and deploy. The prototype proves the demand and then stalls, because the organization has a process for building software and no process for adopting software that appeared from the inside.
The deeper point: a prototype that automates the existing process locks the existing process in. The value is in redesigning the workflow around what the technology makes possible.
What to check in your business
- How many working prototypes built outside engineering exist in your business right now? If you don’t know, that’s the finding.
- What’s the path from an employee-built prototype to production, with a security review and an owner? Not the theoretical path, the one that has actually been used.
- For your last three automation projects, did you redesign the workflow or make the old one faster?
4. Operations become continuously supervised
Most public conversation about agentic commerce is about an agent that buys. The room was more interested in agents that watch, because their operational value was easier to make concrete: a feed that stopped updating, a price that moved somewhere it shouldn’t have, inventory that reports stock the warehouse doesn’t have.
Two examples from the table, with identifying details removed. A large travel marketplace kept selling itineraries containing route segments that no longer existed, because no single system owned whether the whole itinerary was still real. A subscription business struggled to explain a revenue decline; the cause was a demand shock from an adjacent market, and its internal reporting didn’t expose the connection.
The human role moves from performing the analysis to supervising a process that performs it, and knowing which answers to distrust.
What to check in your business
- What’s the longest a critical data feed has silently failed before someone noticed? The number you can prove, not the number in the SLA.
- Which questions about your own business currently require a specialist and a queue? Those are the candidates.
- Who’s accountable for noticing a demand shift that starts outside your category? If the answer is nobody, you share the blind spot in the example above.
The same failure, four times
Every one of these is one failure wearing different clothes. A new capability arrives. The organization drops it into the existing structure, keeps the structure, and is surprised the returns are modest. It happened with electricity: factories bought electric motors, bolted them where the steam engine sat, kept the old layout, and got little until they rebuilt the floor around the new machine.
“Companies didn’t capture the value of electricity by appointing a Vice President of Electricity. They rebuilt the factory around it.”
A participant at the roundtable
A product feed built for a keyword index and now fed to an AI system is the same mistake. So is a license rolled out with no teaching. So is a prototype that automates a process nobody re-examined.
Common questions
Commerce and technology leaders, by invitation. The session was deliberately small so that everyone could participate, and no participant or company is identified in the write-up.
No. It reflects a single three-hour discussion among a small group of invited leaders and should be read as a qualitative roundtable report. Remarks are unattributed and identifying details in the examples have been removed or generalized.
Download it from this page. We will use your email to send the report, and if you opt in, we may also contact you about future roundtables.
Tick the box to be considered for a future roundtable when you download the report, or contact us. Sessions run a few times a year in New York and London and are invitation-only.
You don't need all four.
You need the right one first.
Fixing the wrong requirement first is a year spent on the one that wasn't costing you money. The AI Readiness Assessment examines the product data, platform foundations and delivery processes behind your AI ambitions and identifies the constraint most likely to block value. It covers the engineering side of what the roundtable surfaced. The adoption work is the one part no assessment can do for you.
SHARE THIS POST
Latest Insights
Stay ahead of the curve with our expert analysis, industry trends, and actionable advice. Our blog offers fresh perspectives on the challenges and opportunities in the tech landscape, helping you make informed decisions and drive innovation within your organization.
