Why Incumbents Keep Winning Enterprise AI Deals With the Worse Product

Give a worker ChatGPT, Gemini and Copilot at once and 8% pick Copilot. Give the same worker only Copilot and 68% adopt it. Incumbent advantage is that gap, and almost none of it is about product quality. Microsoft changed what sits on the desk, not what the software does. This is the mechanism, measured, and what challengers actually do about it.

What "incumbent advantage" actually buys

Founders use the phrase as a shrug. It is more specific than that, and each part of it can be priced.

An incumbent arrives at an enterprise AI evaluation holding four things a challenger does not. A signed master services agreement. A completed security and vendor risk file. A renewal date the buyer has already budgeted against. And an internal sponsor whose last three purchases from that vendor did not blow up.

None of those are features. All of them are work the buyer does not have to redo. A procurement team is comparing total effort to reach production, and on that measure the incumbent starts a long way ahead.

Here is the position I will defend for the rest of this piece. Product quality is the last gate, not the first, and most AI startups build as though the order were reversed. The evidence is unusually clean, because one large vendor ran the experiment in public.

The preference gap is measurable, and it is enormous

Recon Analytics tracked more than 150,000 US paid AI subscribers between July 2025 and January 2026. The finding that matters is not which tool won. It is what happened when choice was removed.

Where workers could reach all three major assistants, 70% named ChatGPT their primary tool, 18% named Gemini and 8% named Copilot. Where the employer provided Copilot alone, 68% adopted Copilot as their primary tool.

Read those two numbers next to each other. The product did not change. The shelf did.

Same worker, different employer, opposite outcome

This is the cleanest natural experiment available in enterprise AI right now. It isolates distribution from quality, because the population is the same and only availability differs.

It also explains why founders and buyers argue past each other. The founder is looking at the 8%. The buyer is looking at the 68%. Both numbers are true and they answer different questions.

Licences are not adoption

The same study reports workplace conversion, meaning the share of paid subscribers who also have the tool at work. ChatGPT converts at 83.1%. Copilot converts at 35.8% and Gemini at 34.0%.

Treat those as directional. They come from a subscriber panel rather than from licence records, and the vendor sells the full report. The direction matches everything else here, and the sample is large enough to cite.

Distribution is arithmetic, not mystique

Founders talk about distribution as if it were a personality trait of large companies. It is a fraction, and you can compute it.

Microsoft reported that Microsoft 365 Copilot passed 30 million paid seats in the quarter ended 30 June 2026. It reported more than 450 million commercial Microsoft 365 paid seats at its January 2026 results. That is an attach rate near 6.7%.

A 6.7% attach rate would be a disaster for a standalone product. On that denominator it is 30 million paying seats, added in under two years.

The denominator does the work

This is why the same conversion rate means completely different things on the two sides of the table. A challenger converting 6.7% of a 5 million seat reachable market has 335,000 seats and a hard fundraising conversation.

The uncomfortable implication is that an incumbent does not need a good product to win. It needs a product that is not actively rejected, sitting where the work already happens. Rejection is a much lower bar to clear than preference.

The same pattern runs through the rest of the software stack. It drives the absorption of point tools into agent ecosystems, and it sits under the question of whether Salesforce ends the agent era as a platform or as roadkill.

Procurement inertia is the second engine

Distribution puts the incumbent in the room. Procurement keeps the challenger out of it.

Andreessen Horowitz surveyed 100 verified VP and C-level executives at Global 2000 companies for its third annual enterprise AI report, published on 30 January 2026. 88% of those companies had revenue above $1 billion. The finding on vendor choice is blunt: 65% preferred an incumbent solution where one was available, citing trust, integration with existing systems and procurement simplicity.

Notice what is absent from that list of reasons. Nobody said the incumbent product was better.

Buyers say so out loud

This matters because it removes the usual excuse. Founders often assume the buyer is confused, or quietly captured by a vendor relationship. The survey says the opposite. The buyer knows exactly what they are optimising for, and it is not model quality.

The direction of spend reinforces it. In a TechCrunch survey of 24 enterprise-focused investors published on 30 December 2025, the consensus was that AI budgets would grow while the vendor count shrank. Rob Biederman of Asymmetric Capital Partners described a bifurcation in which a small number of vendors capture a disproportionate share.

Vendor rationalisation is not neutral ground. When a buyer cuts from forty tools to twelve, the survivors are disproportionately the ones already inside a renewal. That dynamic is covered from the buyer's side in the analysis of app sprawl and rationalisation programmes.

The last row is the point of the table. Every row above it is settled before anyone opens the product.

Compliance approval is a clock, and the incumbent started it earlier

The third engine is the one founders underestimate most, because it does not look like competition. It looks like paperwork.

Security review used to mean a SOC 2 report and a spreadsheet. In regulated buying it now adds AI-specific questions: model provenance, training data rights, subprocessor disclosure, output monitoring and retention. A vendor who answers those with a roadmap rather than a document does not get rejected. The deal simply stops moving.

Regulatory dates make this worse in a way that is easy to misread. The EU AI Act originally applied high-risk obligations from 2 August 2026. The Digital Omnibus, Regulation (EU) 2026/1744, entered into force in July 2026 and moved that Annex III deadline to 2 December 2027. Transparency duties under Article 50 and the penalty regime still started in August 2026.

A delay is not relief for a challenger. Large buyers keep their internal control frameworks regardless, and the incumbent already staffs the function that produces the evidence. Moving the date extends the window in which a well-resourced vendor answers questions a small one cannot.

My view, stated plainly: compliance readiness is a go-to-market asset and most AI startups book it as a cost centre. That is the most expensive misclassification on an early enterprise balance sheet.

The 9x problem, and why your demo misleads you

There is an older explanation underneath all of this, and it predates AI by two decades.

John Gourville of Harvard Business School described the pattern in Eager Sellers and Stony Buyers, published in Harvard Business Review in June 2006. Buyers overvalue what they already own by roughly three times, through loss aversion and the endowment effect. Builders overvalue what they have made by roughly three times. The two biases multiply into a 9x mismatch in perceived value.

Gourville's practical estimate is that a new product needs a relative advantage of about three times the incumbent before a switch happens. Not 30% better. Three times better, on the dimension the buyer already measures.

What the buyer is actually pricing

A switch is never a product swap. It is retraining, prompt libraries rebuilt, integrations rewired, an audit trail restarted, and a personal reputational bet by whoever signed. Those costs land on the buyer. The benefit is a model that answers slightly better.

This is why a demo that wins in the room loses in the committee. The demo shows the product delta. The committee prices the switching cost, and the committee decides.

Where this argument is weakest

If the story above were complete, no AI startup would ever win an enterprise deal. That is plainly false, and the counter-evidence is strong enough to name properly.

Startups are taking the application layer anyway

Menlo Ventures surveyed 495 US enterprise AI decision makers between 7 and 25 November 2025. At the application layer, startups captured 63% of revenue against 37% for incumbents. Incumbents held 56% at the infrastructure layer, where scale genuinely matters.

The same report found that 76% of AI use cases were bought rather than built, up from 53% a year earlier. Buying more, from more startups, is not the behaviour of a market locked shut.

Distribution does not lock a position in

Recon Analytics drew the opposite conclusion from its own data to the one a Microsoft investor would prefer. Its reading was that distribution advantages do not lock in market position, because workers keep evaluating alternatives whatever the employer licensed.

There is a sharper version in the a16z data. Microsoft Copilot's net promoter score among software developers dropped 48 points after those developers tried Cursor. Exposure to a better product is corrosive to a default, and defaults are what incumbents sell.

What nobody can settle yet

The unresolved question is timing. Incumbents win the default slot and challengers win on preference, so the outcome depends on how fast preference converts into procurement decisions. Nobody has a credible model of that lag.

Two caveats about this post's own evidence. The Recon Analytics figures come from a subscriber panel and the full report is sold commercially. The a16z and Menlo samples are 100 and 495, which is small for a global claim, and both firms hold positions in the market they measure.

How challengers actually break through

The three engines above suggest three counters. None is "build a better model", which is table stakes and does not move a committee.

Win a number the buyer already reports. Not a metric you invented, and not a productivity multiplier. Find the figure that appears in an existing operational review, move it, and let the person who owns it carry the renewal argument internally.

Arrive with the compliance file finished. Treat the AI governance questionnaire as a product surface, not as legal overhead. A challenger who answers in a week while the incumbent takes six is briefly the easier vendor, and that is the only window that exists.

Then price against the switch, not against the incumbent's list price. The buyer pays for change management whether you acknowledge it or not. Absorbing part of that cost explicitly beats a discount, because it targets the thing actually blocking the decision.

The second route is the strongest and the most temporary. Whether it holds depends on the absorption test in the kill zone piece.

The vertical route deserves its own caution. Selling into one industry narrows the buying committee and shortens the compliance conversation, and the trade is examined in the analysis of vertical AI eating horizontal SaaS. The counter-case, where a funded challenger takes real share from a data-rich incumbent, is worked through in the Harvey and Thomson Reuters comparison.

One thing worth saying to founders directly. If your deal reviews keep ending with "great product, wrong time", you are not losing on features. You are losing on the gates in the table above, and more model work will not change that.

Frequently asked questions

Where to start this week

Take your last five lost enterprise deals and sort the losses by gate rather than by competitor. Legal, security, budget line, sponsor, product. If four of the five died before anyone ran a serious product evaluation, your roadmap is not the problem.

Then pick one open deal and ask the champion a single question. What number does your manager already report that this would move? If they cannot answer it, you do not have a champion. You have an enthusiast, and enthusiasts do not survive procurement.

Related reading