Is Your AI Startup a Feature? Five Questions That Give You the Answer

Every major model release absorbs capabilities that were standalone products a year earlier. Summarisation, extraction, transcription, basic code generation and search have all made that journey. Five questions tell you whether your product is next, and the honest answer takes about an hour to reach.

What the kill zone actually is

The kill zone is the space immediately adjacent to a platform's roadmap. A product in it does not compete with another company. It competes with a feature that a platform will eventually ship for free.

The term originated in describing large platform companies and it applies more sharply to foundation model providers, for a specific reason. Their roadmap is capability itself, and capability expands in every direction at once.

A platform adding a feature makes a deliberate product decision. A model getting better absorbs capabilities its makers did not specifically target, which makes the boundary much harder to predict.

The five questions

Answer each honestly. The value of this exercise is entirely in refusing to give yourself credit you have not earned.

1. Is your core capability a better interface to a model?

If the product's value is that it makes a model easier to use for a task, you are building a user interface for someone else's technology. Interfaces are the first thing platforms ship.

2. Could a capable customer rebuild it in a fortnight?

Coding agents have collapsed the cost of building narrow tools. Investors have reported founders replacing entire tool categories with internal builds. If your product is genuinely rebuildable in two weeks, your moat is distribution and support.

3. Does the product work without data you have accumulated?

If a new competitor with zero customers could ship something equivalent tomorrow, you have no data advantage regardless of how much data you hold. The test is not volume. It is whether the data is required.

4. Would a major model release make you unnecessary?

Ask what would happen if context windows doubled, reasoning improved, or native tool use arrived. Products built to compensate for a model limitation disappear when the limitation does.

5. When a buyer chooses you, are they choosing you or the model?

Ask five customers why they bought. If the answers are about the model, you are a reseller with a margin. If they are about the workflow, the integrations or the support, you have a business.

How to read your score

Zero to one exposed answers means you are building on something the platform is not coming for. Keep going and re-run the test after every major model release.

Two to three means you have a business today and a decision this year. The exposed answers tell you exactly where to invest, and the investment is usually deeper integration or a compliance surface.

Four to five means the product is a feature. That is not fatal, and it does require choosing quickly between going deeper into a vertical, becoming infrastructure, or selling while distribution still has value.

The three escape routes

Every company that has survived a platform absorbing its category took one of these.

Go vertical. Industry-specific context, regulatory obligations and workflow depth are expensive for a general platform to replicate and unrewarding for it to try. Vertical AI companies crossing meaningful revenue consistently report that integration depth, not model depth, decides deals.

Become infrastructure. Move below the application layer to something other products depend on. Infrastructure is harder to displace because switching costs are borne by many parties at once.

Own the compliance surface. Audit trails, evidence, retention and regulatory reporting are unglamorous, expensive to build and rarely worth building internally. A product carrying a genuine obligation is difficult to replace with a feature.

The route that does not work is competing on model quality. That is a race against companies that spend more on a single training run than you will raise in total.

Why competing on model quality never works

This deserves its own explanation because founders keep attempting it.

The companies training frontier models spend more on a single run than most startups raise across their entire life. Any advantage built on having a better model is temporary by construction, because the next release resets it.

Worse, the position puts you in direct competition with your own supplier. You depend on their infrastructure, their pricing and their access terms while attempting to beat them at the thing they are optimising hardest. That is not a difficult position. It is an untenable one.

What the evidence supports

The direction is well established and the timing is not.

Gartner projects that roughly 35% of point-product SaaS tools will be replaced or absorbed into agent ecosystems by 2030. That is substantial and it is a decade-long process, which matters for how urgently you should act.

Meanwhile enterprise software counts are still growing. The average enterprise ran 291 SaaS applications in 2025, up from 110 in 2020. Companies are buying and cancelling simultaneously, which is a rationalisation rather than a collapse, examined further in the piece on the software repricing.

The practical reading is that absorption is real, gradual, and uneven. Being in the kill zone is not a death sentence within a quarter. It is a reason to make a strategic choice while you still have the revenue to fund it.

When to run the test again

This is not a one-time exercise, because the boundary moves. A product that passed comfortably eighteen months ago may fail today without anything about it changing.

Re-run it after every major model release from a provider you build on. That is the event that most often converts a defended answer into an exposed one, and it happens several times a year.

Re-run it when a platform ships an adjacent feature, even one that does not compete directly. Feature adjacency is how absorption usually begins, and a platform shipping something near your category is a signal about its roadmap direction.

And re-run it after any funding round, yours or a competitor's. Capital changes what is buildable, and a well-funded competitor with a similar answer set will reach the escape routes before you do.

Founders who run this quarterly report the same benefit. It is not that the answers change often. It is that having a written record of when they changed makes the strategic decision obvious at the point it needs making, rather than a year afterwards.

What has already been absorbed

The pattern is easier to see backwards. These capabilities were standalone products before becoming default features.

The survivors in every row share a property. They stopped selling the capability and started selling the workflow it sits inside, generally before they were forced to.

Where this test fails

Three limitations worth stating.

It systematically undervalues distribution. Plenty of products that fail all five questions have profitable businesses because customers do not switch, procurement is slow, and being adequate and already installed beats being better and unknown.

It also cannot see timing. A product may fail every question and have four good years, which is long enough to build something else or reach an exit. The test tells you the direction of the risk and nothing about the clock.

And the escape routes are not equally available. Telling a horizontal product to go vertical assumes you have the customer relationships and domain knowledge to pick one. Most do not, and choosing badly is worse than staying put.

The test is also silent on the option most founders eventually take, which is selling. A product with real distribution and weak defensibility is frequently worth more to an acquirer than to its own shareholders over time, and recognising that early is a strategic outcome rather than a failure.

Frequently asked questions

Where to start this week

Run the test with someone who will disagree with you.

Answer the five questions alone first, then hand them to your most sceptical engineer or your least sentimental investor and ask them to answer independently. The gap between the two sets of answers is more informative than either set on its own.

Then take the single most exposed answer and write down what you would build in the next six months to change it. If you cannot describe that work concretely, you have found the real problem, and it is a strategy problem rather than a product one.

Diarise the next run for the day after the next major model release. That is when the answers most often move.

Related reading