A founder showed me his roadmap last spring. Fourteen slides, eleven of them about AI. Copilot in the dashboard, smart summaries in the reporting module, a natural language layer over the search, an assistant that would “proactively surface insights”. I asked him one question: which of these does a customer already have a budget line for?
Silence. Then, honestly: “None of them. But they all keep asking about AI.”
That gap is the most expensive thing in software right now. Customers ask about AI in meetings because it is the polite question of the era, the way everyone asked about “the cloud” in 2013 and “mobile first” in 2011. Asking is free. Paying is not. And the moment a purchase order needs a signature, the question changes from “do you have AI” to “what does this replace, what does it shorten, or what liability does it take off my desk”. If your feature has no answer to that, it will be admired and not bought.
I have been on both sides of this. I ran innovation inside a bank in Hong Kong, where we built machine learning into lending decisions and hyper-personalization into customer engagement, in an environment where every experiment had to survive risk, compliance, and a committee that had seen fashions come and go. I have since run a P&L as a CEO, which is a very different education. Inside the bank I learned what makes a model defensible. Running a company I learned what makes anything survive a budget cycle. The two lessons are not the same, and the second one is the one most product teams are missing.
The two questions everybody is asking, and why both are unanswerable
Every founder I speak to is carrying the same two anxieties. The first: is AI going to eat my product? The second: where should I actually put it?
Both are unanswerable as stated, which is precisely why they produce so much activity and so little progress. “Is AI going to eat my product” invites an existential answer, and existential answers produce strategy decks. “Where should I put AI” invites a creative answer, and creative answers produce feature lists. Neither produces a decision you can test in ninety days.
I replace both with a narrower test, and it is deliberately brutal. Name the line item that moves. Name the person who owns that line item. Name the number it moves by within one quarter.
Three answers, all specific, all falsifiable. If you cannot produce them in a single sentence, you do not have an AI initiative, you have an AI intention. I have watched this distinction decide the fate of a dozen projects. The ones attached to an existing budget and an existing owner survive contact with the organization. The ones attached to a strategy deck die quietly around month five, usually when the sponsor changes role and nobody inherits the enthusiasm.
Notice what the test rules out. It rules out “improves customer experience”, because nobody owns that line. It rules out “positions us for the future”, because no quarter contains the future. It rules out “our competitors have it”, because that is a fear, not a mechanism. What survives is narrow and unglamorous: this reduces the time a claims handler spends per file from eleven minutes to four, the operations director owns handler capacity, and she is currently over-staffed by two heads against her plan.
That sentence sells. The fourteen-slide roadmap does not.
Customers buy removed cost, shortened cycles, and transferred risk
Strip away the vocabulary and there are only three things a business buyer has ever paid for. A cost that goes away. A cycle that gets shorter. A risk they no longer carry personally.
Removed cost is the easiest to price and the hardest to fake. It shows up as headcount you do not hire, a vendor you stop renewing, a process you retire. Be careful here, because “saves my team time” is not removed cost unless the time converts into something. Ten minutes saved per person per day across forty people is a lovely number in a deck and often worth exactly zero in the P&L, because nobody reassigns the ten minutes. Cost is removed when a manager can point at a line and say it is smaller now.
Shortened cycles are where I have seen the most genuine value from AI, and the least marketing attention. In lending, the difference between a decision in four days and a decision in four hours is not a convenience improvement, it is a different conversion rate, because applicants shop while they wait. In healthcare operations, where I led digital transformation for a private chain expanding across the mainland, the constraint was never the quality of the clinical decision. It was the number of days between a patient’s first contact and their first appointment, and everything that fell out of the pipeline in between. Any technology that compresses a cycle where customers leak out of the funnel has a value you can calculate on the back of an envelope, and defend in front of a CFO.
Transferred risk is the most underrated. Someone in every organization has their name on an outcome that scares them: a regulatory filing, a safety check, an audit trail, a quality release. If your product means that person sleeps better, you have a buyer who will fight internally on your behalf. This is also where AI is most often mis-sold, because a probabilistic system introduced without governance adds risk rather than removing it. The winning position is not “our AI decides”, it is “our AI does ninety percent of the work and gives you an evidence trail for the ten percent you sign”.
Intelligence, in the abstract, is not on the list. Nobody has ever approved a budget for intelligence.
Where AI is a distraction, and how to spot it before you burn a quarter
The pattern I see most often is AI applied to the visible part of the product rather than the expensive part of the workflow. Dashboards get the copilot. The eleven-step reconciliation process that consumes the customer’s Tuesday gets nothing, because it is ugly, undocumented, and full of exceptions. But the ugly process is where the money is. Beauty attracts demos, ugliness attracts contracts.
A few practical signals that you are heading for a wasted quarter.
If the feature is demoed to a buyer but used by nobody in particular, it is a distraction. Every durable feature has a named daily user with a job title. If you cannot name the job title, you have built something for the meeting, not the work.
If the value depends on the customer changing how they work before they see any benefit, the adoption cost will eat the gain. I learned this building consumer hardware in China with the Jean-Michel Jarre venture, taking one product into a range of eight. Hardware punishes you instantly for assumptions about behaviour because the units sit in a warehouse. Software lets you carry the same delusion for two years because a dormant feature costs nothing visible. That is not an advantage, it is a slower feedback loop.
If a competent engineer with an API key could rebuild your feature in a weekend, it is not a product, it is a demo. And here is the part that founders find uncomfortable: this is often true, and it is often fine. Your moat was never the feature. If an AI-native competitor can replicate your summarization in two days, ask what they cannot replicate in two years. The answer is almost always the same three things: the workflow you sit inside and the switching cost of leaving it, the data your customers trust you to hold and the permission that trust represents, and the distribution you earned through years of relationships, certifications, integrations and reputation.
When I co-founded one of China’s early location-based platforms, we crossed three hundred thousand users, and I can tell you that the technical parts of what we built were copied fast. What was not copied was the accumulated behaviour of the network and the operational knowledge of running it at scale. Features are borrowed. Position is not.
So when you ask “is AI going to eat my product”, the honest translation is: is my product actually just a feature wearing a logo? If yes, AI is not your problem, it is your deadline. If no, AI is a tool to deepen the three things that were already your defence.
What to do differently on Monday
Take your roadmap and apply the three-part test to every AI item. Line item, owner, number, inside one quarter. Most teams find that two items out of ten survive. That is not a failure, that is the exercise working. Kill the rest or park them honestly as experiments with an explicit budget cap, which is very different from shipping them as strategy.
Then go and find the owner of that line item inside three real customers and ask a specific question: what does this number look like today, and what would have to be true for it to look different by December? You are not validating a feature. You are validating that a human being with a budget already wants this number to move and has failed to move it with the tools they have. That failure is your opening.
Instrument the before state before you build anything. This sounds obvious and is skipped almost universally. If you cannot state the baseline, you will never be able to prove the delta, and an unprovable delta cannot be renewed. In the bank, the discipline that made models survive was not the modelling. It was that someone had written down what the process cost before we touched it.
Finally, be willing to say that AI does not belong in a part of your product yet. That sentence has become almost unspeakable in board meetings, which is exactly why it carries information. The companies that will look smart in three years are not the ones that put AI everywhere. They are the ones that put it in two places where it moved a number somebody owned, and spent the rest of the quarter widening the moat that AI was never going to give them anyway.
I write from twenty years of building businesses between Europe and Asia. If your company is facing this, start a conversation.