The Awkward Middle What Pre-ChatGPT Software Companies Are Up Against

There's a specific kind of company that's having a strange few years. It did everything right. It found a market, built a product people pay for, and grew into the $10M to $100M ARR range. Real revenue, real customers, a real team. And it did all of this before late 2022, when ChatGPT and the wave of AI coding assistants reset everyone's expectations about how fast software can get built and what it should cost.

These companies aren't failing. But they're stuck in an uncomfortable spot. They're too big to pivot like a startup and too small to absorb disruption like an incumbent. Here's what they're actually dealing with.

A company that hit scale before 2023 has years of code written by humans, for humans, under deadline pressure. It works, but nobody built it with AI tooling in mind. There's no consistent structure for an assistant to navigate, the documentation is thin, and a lot of the important knowledge lives in the heads of a few senior engineers.

The frustrating part is that AI coding assistants deliver the least benefit on exactly this kind of code. They do their best work on new projects and clean, well-organized systems. Point one at a sprawling, under-documented monolith and it stumbles, invents context that isn't there, and produces changes that senior engineers then have to babysit.

“

So the companies that need a productivity boost the most are the ones worst positioned to get it.

A lot of these companies set their prices around what it used to cost to build software the old way. Then a new crop of competitors showed up, putting out comparable features for a fraction of the engineering spend, and happy to price that way too. Sometimes they give away as a free feature what the older company sells as a whole product.

That squeezes pricing power from below. The older company can't just drop prices without breaking the math that its business was built on, and it can't justify charging a premium when a venture-funded rival is undercutting it to grab market share.

There's a second shift underneath that one, and it's bigger. The annual license and the per-seat model are falling away. Buyers increasingly expect to pay for what they actually use, measured in tokens, calls, or compute, not in headcount. And the whole idea of a "seat" is getting shaky, because the thing using your software might not be a person anymore. It might be an AI agent hitting your REST API thousands of times an hour, with no login, no dashboard, and no interest in a per-user plan.

“

You can't charge a seat fee to something that was never a seat.

Most of these companies have no real answer for this. The honest internal position is "we'll figure it out when a deal forces us to," which sounds pragmatic and is actually a problem waiting to happen. The software was never instrumented to bill on usage, so the telemetry to count tokens or meter API consumption simply isn't there. The pricing team, the product team, and the data team each assume someone else owns it. Then a customer shows up wanting to pay per call for agent access, and there's no model to quote, no system to measure it, and no way to get ready in the time the deal allows. The deal either stalls or gets signed on terms the company can't actually support.

Every leadership team in this range is having some version of the same argument. Do we rebuild our core systems to work with AI from the ground up and collect the long-term savings, or do we bolt AI onto what we already have and keep shipping?

Neither option is clean. A rebuild pulls your best engineers off revenue work for months or years, on the bet that the payoff shows up before a competitor takes your customers. A retrofit risks attaching AI features to an architecture that fights them, leaving you with a product that feels stitched together instead of genuinely rethought. Startups never face this fork because they have nothing to tear down.

This is the part that doesn't show up in a strategy deck. Any real shift these companies make comes with an upfront tax. Retraining the team, untangling the old architecture, getting buy-in across departments, reworking the security review for a new class of tooling. None of it is free, and all of it takes time before a single customer sees a benefit. Meanwhile the AI-native competitor isn't paying that tax. It's already moving at the speed the new tools allow, and it pulls further ahead during exactly the window when the older company is slowed down by the work of catching up.

That's the trap. The bigger the change, the longer the initial drag, and the more ground a faster rival covers while you're mid-turn. So leadership keeps choosing small, safe moves that don't trigger the delay, which is precisely how you fall behind in slow motion.

Most of these companies built their roadmap by listening closely to their customers, and that habit served them well for years. The problem now is who they're listening to. Their main contacts tend to be the buyers and the departments they've always sold to, and a lot of those teams are behind on AI themselves. They're describing the workflow they have, not the one that's coming.

Inside those same large enterprises, the people actually working out where things are headed sit somewhere else entirely. The internal AI labs, the research groups, the teams running experiments with the new tech. Those teams rarely sit in the buying conversation, so their signal never reaches the roadmap.

That breaks the flywheel. The company faithfully builds for the departments it hears from, and some of those departments are the ones most exposed to being automated or restructured as AI lands. You end up perfecting a product for a workflow that's on its way out, guided by customers who are as far behind the curve as you are. The feedback loop that used to keep you aligned with the market is now quietly steering you away from it.

Good engineers now expect to work with modern AI tooling, and they judge employers partly on whether the day-to-day actually lets them. A company with a locked-down legacy stack, security rules written before any of this existed, and a codebase that doesn't cooperate with assistants can feel like a step backward to someone used to a faster way of working.

At the same time, the gap between an engineer with good tooling and one fighting legacy friction keeps getting wider. That makes headcount harder to justify, ambitious people harder to keep, and hiring harder when you're competing against companies that were AI-native from day one.

But the harder problem isn't hiring these people. It's that the ones already inside the building are moving 10x faster and immediately slamming into everything around them that isn't. The engineer using AI well can produce a week's worth of old work in an afternoon, and then it sits. The release cycle still takes months. Legal isn't set up to review anything involving AI and defaults to "let's wait." Security has no process for a new class of tooling, so approvals stall.

Customers now expect AI features. Plain-language interfaces, smart automation, instant summaries. They treat these as the baseline, not as something special. A product that was clearly the best option in 2021 can feel dated in 2025, not because it got worse, but because the floor moved up.

The pressure to ship something visibly "AI" is intense, and it tempts these companies into stapling a chatbot onto the corner of the dashboard and calling it done, instead of doing the slower, harder work of rethinking the product around what's now possible.

A company in this range is often venture-backed or owned by private equity, with investors who have watched AI-native startups do more with smaller teams. The pressure to show the same kind of efficiency, to flatten the org and prove that AI is widening margins, lands on a company whose costs and architecture were set up for a different era. The demand is real even when the fast way to meet it isn't.

There's no clever third option here. It comes down to a decision, and most companies in this range are making it by default instead of on purpose.

The first path is managed decline. You don't call it that, of course. You cut costs, you win a few deals while others quietly churn, and you tell everyone you're "focused on efficiency." You sign a contract with OpenAI or Anthropic so the team feels like the company is serious about AI, then you steer everyone toward Sonnet or an older, cheaper model to keep the burn down. You ship a chatbot in the corner of the dashboard. From the outside it looks like a company addressing AI. On the inside, the real strategy is to stay presentable long enough to get acquired. You stay stuck below $100M, with no realistic path to break out, and you sell before the slide becomes obvious. For some founders and boards that's a rational outcome. Just be honest that it's the plan.

The other path is to do the thing almost nobody does at this size. Go back to founder mode. The Sergey Brin move, where the founder drops back into the building and starts making real decisions again instead of managing the machine from a distance. The organization heads back to its startup roots and makes actual bets. You stop optimizing for the customers and departments that are themselves on the way out, you reorganize around the AI-native reality instead of bolting it on, and you accept that this means short-term pain and real risk of being wrong.

The question is whether you spend that asset trying to look stable on the way to an exit, or whether you bet it on actually breaking out. The middle, where most of these companies are sitting right now, is the one place the decision gets made for you.