Flexibility is discovered, not bought

Recently, OpenAI began what looks like a deliberate shift towards enterprise. Several of its more experimental consumer products have been wound down, with Sora, the AI video platform, the most visible casualty. The official reasons included compute costs, declining user numbers, and a desire to concentrate resources. The competitive pressure from Anthropic's growing success in coding and enterprise is another reason that is difficult to ignore. Whatever the motivation, the direction of travel is clear enough. A company with one of the most capable and flexible AI systems ever built is choosing to focus it, because flexibility alone does not build a business. Focus does.

The question is why that lesson is so consistently hard to act on, and why so many companies with genuinely capable products arrive at it late, if at all.

The more flexible a product is, the more potential value it can create. And, almost as a direct consequence, the harder it is to buy.

Companies tend to respond to this in one of three ways, and all three have problems.

The first is to explain the breadth more clearly. Founders in particular love to demo their product and show everything it can do. That is understandable and usually counterproductive. Buyers do not become clearer when given more information about a product they cannot yet place. They become clearer when given a sharper frame for it.

The second is to borrow an existing category, or at least its language, to make the product legible. That can be useful, especially early on, because buyers need a familiar frame before they can understand what is different. But if the borrowed category becomes the actual positioning, rather than a bridge into it, it can attract the wrong evaluation processes as easily as the right ones.

The third, and most tempting for founders who believe they have built something genuinely new, is to define a new category altogether. This almost always underestimates how much work the market has to do before it can even decide whether to buy. Creating a new mental model in a buyer's head is an enormous ask, and most companies do not have the resources, the time, or the sustained evangelism it requires.

I have seen this pattern from the inside more than once. A technically strong product can solve a real problem, win immediate enthusiasm from practitioners closest to that problem, and still struggle commercially because the organisation around the buyer does not know how to place it. The users understand the pain and recognise the value quickly. The friction appears elsewhere. IT, procurement, architecture, governance, and formal buying processes all want to know what kind of product this is before they can decide how to evaluate it.

That is where flexibility becomes commercially awkward. A product may genuinely span several categories at once, or combine capabilities that are usually bought separately. The truthful answer is that it does several things and fits no existing box particularly well. But truth is not always helpful in a buying process. A procurement team cannot shortlist “partly several categories, but not exactly any of them”, and an internal champion cannot repeat it clearly in a sentence to colleagues.

Reaching for a broader category label can seem like the solution. It gives the product a recognisable shelf to sit on and a process to be included in. But it also invites evaluation that is not real opportunity. The product gets pulled into RFPs designed around incumbents, measured against criteria that reflect someone else's maturity model, and compared on strengths it was never built to optimise for. Time, energy and morale get consumed by processes that were never genuinely winnable, while the company mistakes activity for progress.

The more useful instinct is usually to go narrower, not broader. The better answer is often to anchor the product in a familiar category while qualifying it with a specific buyer context, workflow, or strength. That framing will not capture everything the product can do, but that is precisely the point. It makes the product legible without pretending it is identical to the incumbents in that category. It attracts fewer conversations, but better ones. It filters out poor-fit opportunities before they consume the calendar, gives the internal champion something concrete to repeat, and provides buyers with a frame they can actually use. That is usually far more valuable than making a broader claim the product has not yet earned.

There is a particular commercial dynamic that makes this worse in practice, and it deserves its own mention.

Salespeople are paid to find and pursue opportunities. With a flexible product, there is almost always a plausible story for any prospect. The product can do something relevant to almost anyone. So the pipeline fills up, conversations are started, proposals are written, and the business looks encouragingly busy.

Usually, that activity is not progress. It is a slow-motion dispersal of energy.

Each pursuit in a different direction pulls marketing into writing different collateral, pulls product into accommodating different requirements, and pulls the whole organisation into describing itself differently depending on who is in the room that week. The sales narrative fragments. References become hard to reuse. And the win rate in each individual area stays stubbornly low, because the product is being evaluated as a partial answer to someone else's well-defined problem rather than as the complete solution to a problem it was built to solve.

A flexible product genuinely can go in many directions. That does not mean it should. The discipline of saying no to plausible but poorly-fitted opportunities is one of the hardest things for a young company to develop, precisely because those opportunities feel like validation. They are usually closer to distraction dressed up as pipeline.

At this point the obvious objection arrives, and it is a good one. The spreadsheet.

If any product succeeded on the basis of flexibility, it is the spreadsheet. Excel and its predecessors became the most widely used analytical tool in the history of computing, ending up in finance, science, operations, project management, journalism and sport, far beyond anything its creators anticipated. It succeeded precisely because it could be applied to almost anything.

So surely the spreadsheet disproves everything argued above?

Look at how it was actually launched, and I think it does the opposite.

When VisiCalc appeared in 1979, it sidestepped the new category problem entirely by borrowing one from the physical world. The word "spreadsheet" already existed, referring to the paper ledger open across two pages of an accounting book. Businesspeople knew exactly what a spreadsheet was. VisiCalc was simply one that recalculated itself.

The problem it solved was equally concrete. Dan Bricklin, watching financial models being built and corrected on ruled blackboards at Harvard Business School, imagined that a computer could make the recalculation automatic. The frustration was real and visible. VisiCalc solved that one thing.

The flexibility was not the pitch. It was what users discovered after buying the product for a concrete reason. The extraordinary range of things the spreadsheet eventually proved useful for was something the market learned through use, not something it was asked to believe upfront.

There is a cautionary postscript. Bricklin and Frankston did not capture most of the value their invention created. Lotus 1-2-3 arrived in 1983 with better execution and more resources, and VisiCalc was effectively finished within two years. Creating a category is not the same as winning it.

Which brings us back to OpenAI.

OpenAI has, of course, already shown how well the first flexible-product challenge can be handled. When ChatGPT launched in late 2022, it did not try to define a new category or educate the market about large language models. It wrapped a genuinely new technology in the most familiar digital interface imaginable. A chat window. In that sense, it pulled the same trick VisiCalc did in 1979, making something novel feel immediately comprehensible by placing it inside a familiar frame.

But the comparison only goes so far. VisiCalc solved a relatively specific problem, and the spreadsheet itself quickly became a clear commercial category. ChatGPT made a far broader capability legible, but not necessarily specific. The chat interface solved the adoption problem. It did not fully solve the later question of what durable business position should be built on top of such a general-purpose system. That is the stage OpenAI now seems to illustrate.

Anthropic did not create the general-purpose AI assistant category. It entered a space OpenAI had already opened up, but kept its compute focused on text and code rather than dispersing it across video, image generation, and consumer social products. The pioneer cleared the path. The challenger that stayed narrower is finding it easier to build a business on it.

The current shift towards enterprise is OpenAI's recognition of exactly that dynamic. Whether driven by strategic conviction, competitive pressure, or both, the logic is the same logic that has applied to flexible products for as long as there have been flexible products.

Markets do not buy flexibility. They buy the removal of a specific, recognisable pain. If the product is genuinely capable, users will discover the breadth for themselves. The job of positioning is not to convey all of that capability at once. It is to make the product legible to the right buyer at the right moment.

Flexibility is discovered. It is not bought.