# AI-Native Without Starting Over

Published: 2026-08-21T15:06:30.000Z · Updated: 2026-08-21T15:06:43.000Z · Author: Rod Amora · Canonical URL: https://rodamora.com/blog/ai-native-without-starting-over

> Somewhere in your firm there is an AI tool somebody quietly stopped using, and nobody told you. That is where AI adoption breaks, not in your legacy processes. Existing firms do go AI-native: real training, cheaper models under $500 a month, and somebody paying for the failed first tries.

Somewhere in your firm there is an AI tool you paid for that somebody quietly stopped using, and you don't know which one it is. Nobody complained, nobody filed anything.

The loudest advice says the firm itself is the problem: Diana Hu's April 2026 YC talk, [The Playbook for Building an AI-Native Company](https://www.ycombinator.com/library/OX-the-playbook-for-building-an-ai-native-company), and its June 2026 services sequel from YC's Charlie Warren, who calls buying an existing firm and adding AI on top "generally a trap."

He is half right. Buying licenses changes nothing. But your old business is not the wall, the wall is that nobody paid for the failed first tries.

Existing firms do become AI-native. They pay for real training, 8 to 10 hours over about 3 weeks, they run on cheaper models under $500 a month, and they budget for the weeks when the new way is slower. Here is that path, and the part the playbook never says: a better model is not going to buy you a culture.

## The user who quit the tool and said nothing

We launched a new AI agent tool wired into our project management software. It could pull content from every past engagement with a customer, basically all the consultant notes, so the answers came back specific to that client instead of generic. Better than any ChatGPT or Gemini session, better on every slide anyone would make about it.

Then one user hit a problem. A couple of Google Meet recordings did not import their transcripts into the AI brain. He did not file a bug, he did not tell anyone, he went back to the old system with no integration at all and kept working from there.

We found out days later, by accident. He mentioned he was having a problem with the old system, somebody asked him what he was doing when it happened, and he said he was not using the new tool, he had reverted to the old one because he hit a couple of problems. He mentioned it casually. He was not resisting and he was not hiding it, to him nothing had happened worth reporting, a tool failed twice so he used the one that works.

Now look at what had to go wrong for us to learn any of this. The fallback had to fail too, while he was standing in it, with someone in the room asking the right follow-up question. If the old system had kept working for him, we would still not know.

So the adoption number you can see is the flattering one. A survey would not have caught this and neither would a check-in, because he did not think he had anything to report, and he was right, from where he sat nothing had happened. What does catch it is [looking at where the work actually went](https://rodamora.com/blog/ai-projects-dont-fail-at-your-size-they-go-underground), the file histories and the source trails.

A few broken transcripts were enough reason to discard an AI tool that was better in every other way.

If you have shipped AI tools inside your firm, some version of this has already happened. You just don't know which tool it happened to.

## Nobody paid for the failed first tries

The easy read is that people resist change. Some of it is that, but most of it is the fear of not delivering, of looking incompetent in front of their peers, of being the one who broke the client work by playing with a robot.

A founder starting from zero can experiment all day, because there is nothing else to do. No clients waiting. No old system to fall back on. Sometimes no expectations whatsoever, the world is their oyster.

Your team is not in that position. When a new tool fails the first time, the old system is sitting right there, proven and paid for. Going back costs nothing that anyone tracks, the work still pops out the other end, and no one has to explain anything to anybody.

You didn't plan for what getting better costs: trial and error, slow weeks, work done twice while the new way catches up. Experimentation was never declared part of the job, so it loses to delivery every single week.

That is what was really happening in the story above. The user who went back was not stubborn. He was busy, and nobody had told him the failed first tries were part of the plan.

That asymmetry is the whole reason the greenfield playbook does not transfer:

What happens when a new tool fails

Founder starting from zero

Firm that already has clients

What they fall back on

Nothing, the new way is the only way

The old system, proven and paid for

What going back costs

Not an option

Nothing anyone tracks

Who pays for the slow weeks

The founder, out of runway

Nobody budgeted it, so delivery pays

How you find out it happened

You were in it

Only if the fallback fails too

## The trap is real, and the playbook blames the wrong thing

Warren is describing a real failure. "Don't try to buy your way in," he says, and the temptation he means is "to try and buy an existing services business, add some AI on top, short circuit the revenue. This is generally a trap." If your plan is to buy licenses and leave every process alone, he is right about you.

But the playbook goes one step further. It says the firm itself is the problem, old thinking, old people, old processes, so start over.

That is where I have to disagree, and I disagree from a different chair. Across the 150+ units in my franchise network, roughly 80% are still running chatbots, 15% have moved to rigid workflows that run on triggers, and 5% are running autonomous agents defined by conditions. That is pulled from my delivery data, not an estimate.

That split is a photograph, not a film. I have not measured those firms moving between the tiers over time, so read it as the starting picture and nothing more. Most are early in the climb and that is expected, we are all very early right now, it's okay. What I can tell you is what moving one single task costs, because I have paid that bill myself.

The playbook's authors watch brand new startups. I watch hundreds of firms that already exist. Different seat, different data.

## What does it cost to move one task to AI?

Meeting summaries. That is the smallest useful example I have, and it still cost senior time for weeks.

We started checking all the summaries to make sure the thing was doing it properly, not hallucinating anything, not creating facts, not changing data. Senior people, every summary, by hand. We ran that for a few weeks until we were convinced it was reliable enough, and then we just stopped checking.

That is the whole gate. No rubric, no accuracy threshold, no score anybody could put on a slide, just senior people who had read enough summaries to trust the next one. You can start that version on Monday with the people you already have.

Here is the bill for it. Estimated from the artifacts we marked as reworked, tuning burned as much as 200 hours of senior people's time over six months. That is a real cost, small next to the manual labor the system removed, and nobody had it in a budget beforehand.

And I owe you one more thing about that gate, because it works against my own argument. The models evolved a lot while we were checking, the larger ones became more convincing and actually more reliable in those same months. So I cannot hand you the clean story where our tuning earned the whole gain. Part of it walked in the door on its own.

Other tasks followed. Research review now matches quotes to their references on its own and flags what a human has to check. Data analysis moved from heavy human checking to agent reviewers, and competitor analysis and content creation went the same way. Each one walked the same curve, [the four stages of reviewing AI work](https://rodamora.com/blog/the-four-stages-of-reviewing-ai-work): marvel, panic, rubric, maturity. That curve is also the answer to Warren's sharpest point, that "customers will fire you for variance faster than they will fire you for being a bit slower or a bit more expensive than the incumbents." Your client will forgive slow, believe me, they will not forgive a report that is great one week and wrong the next.

No rewrite. No clean slate. One task earned its gate, then the next one did, then the next.

## Does a bigger firm have a harder time going AI-native?

Size is your enemy here, and it was the first thing I thought of when someone asked me whether an existing firm can really do this. The bigger the company, the harder it is. More people means more old systems sitting there to fall back on, more weeks where experimentation loses to delivery, and more places for a quiet revert to hide.

Past a certain size, expect the slow stretch to be longer and plan for it, or find clever ways to speed it up. We have not mapped that part of the way yet, and I am not going to pretend otherwise.

## Which parts of the AI-native company playbook work in an existing firm?

The greenfield advice is mostly good advice, it is just aimed at someone else. Two of those ideas translate straight into a firm that already exists. (A third one, Hu's "almost no human middleware," I have already argued in [Fewer Managers, More Managing](https://rodamora.com/blog/fewer-managers-more-managing): the tracking half of management goes to agents, the judgment half moves to the front, and somebody has to tell people what the freed hours are for.)

**"Make your entire company queryable."** Hu is right that AI needs to be able to read what the firm knows. In an existing firm most of that knowledge is not written anywhere, it is sitting in meeting transcripts nobody opens, client notes in someone's head, phone calls, WhatsApp threads, project files, and decisions that were made out loud and never recorded.

Capturing it is the work you do before you buy the agent, and honestly it might be the most important step in your whole climb. You can start today, no license required. [Document the process first, or the agent scales the chaos](https://rodamora.com/blog/document-the-process-first-or-the-agent-scales-the-chaos), and you have the base of [what makes a service business AI-native](https://rodamora.com/blog/what-is-an-ai-native-service-business).

**"Your company should run as a closed loop."** The working version I have seen is plainer than it sounds, whoever uses the system has to catch mistakes, report them and help the system improve, sometimes that leads to a better prompt, a new skill, a change in a company process, the point is that the loop closes at all and the improvements are small and constant.

Just be careful where the feedback comes from. It has to come from the final artifact, the delivered file or report, the sent message, whatever finished means for that work. We learned that one the hard way, a client status update passed our own checks because a task was closed, and the client wrote back to say the work had never been delivered. Don't measure the middle steps and call it done. More on that in [Cheap Mistakes Go to AI, Expensive Ones Wait](https://rodamora.com/blog/cheap-mistakes-go-to-ai-expensive-ones-wait).

Do those two well and the climb to [Stage 3, AI-native](https://rodamora.com/delivery-model-ladder) on the Delivery Model Ladder is a real climb for a firm that already exists, not a rewrite.

## The bill is smaller than the playbook says

They will tell you that a painful API bill is proof that you are serious. Hu tells founders to "run an uncomfortably high API bill, because it's replacing what would have taken a far more expensive and inflated headcount." StrongDM's software factory sets the bar at [$1,000 in tokens per human engineer per day](https://factory.strongdm.ai/). a16z calls [$1,000 per engineer per month "close to table stakes"](https://a16z.com/there-are-only-two-paths-left-for-software/).

The companies that turned token spend into a scoreboard are already changing the tune. By April 2026 Meta had abolished its internal token leaderboard after a backlash and Shopify had renamed its own leaderboard to a usage dashboard, adding circuit breakers to catch runaway agents, both [covered in Pragmatic Engineer](https://blog.pragmaticengineer.com/the-pulse-tokenmaxxing-as-a-weird-new-trend/). In August 2026 a clip went around of Microsoft telling an engineer to calm down on AI usage, though that one I have only seen through a forum thread, so treat it as gossip that rhymes rather than a reported fact.

Spend turned out to be a bad way to measure work, which surprises nobody who has run a business. It's like judging a meeting briefing by how many pages it has.

Here is the number from the other side, so you don't make the same mistake. Experiment with the cheapest model you can get away with for a given task, take an output from a stronger model as your baseline and see how the cheap one fares. On cheaper tiers instead of frontier flagships only, under $500 a month in model costs generates hundreds of client artifacts and deliverables while your firm learns where the expensive models actually earn their price.

The bill is never the scoreboard, my fellow entrepreneurs, cost per accepted task is. [A 23x cheaper model bill proved nothing on its own](https://rodamora.com/blog/the-23-ai-cost-cut-we-couldnt-call-a-win) until someone measured what actually got accepted, because you pay for the failed attempts too and those never show up as a line item anywhere. For the per-task math, [your AI employee has a rate card, and you set it](https://rodamora.com/blog/your-ai-employee-has-a-rate-card-you-set-it).

## Should you just wait for the models to get better?

There is a better objection to all of this than Warren's, and it is probably yours. If the models keep getting more reliable on their own, waiting is cheaper. Buy the working version in eighteen months, skip the failed first tries, skip the 200 hours of senior time, and let your client relationships carry you until then.

On the confounder I just handed you, that is not a stupid position. I told you myself that the models improved underneath our gate.

Here is my answer. A better model is not going to buy you culture. The prototyping culture, the commitment to integrate a new tool, people getting used to changing their workflows, rebuilding the way they work, better models and more money are not going to buy you a new way of people working. So you have to start doing that now, because that is what work is going to be: trying, prototyping, building new workflows and systems where AI works better with humans.

And nobody has settled that part yet. There is no OKR for the agent era, no framework you can adopt on a Tuesday and have the whole firm understand by Friday, so in the meantime we all have to get used to working with uncertainty. That interval is not dead time you can skip, it is where the practice gets learned. The firm that waits arrives late to a practice, not to a purchase.

What those weeks of checking summaries bought us was not a working summarizer. It was people who had done the checking, watched it pay off, and now expect their own workflow to change again next quarter. That does not ship with the next model release.

Meanwhile the advantages you do have are real: years of delivery experience, client relationships, trust you earned slowly one engagement at a time, and the fast-moving newcomers have none of that on day one.

I have not watched an AI-native newcomer take clients from an established firm in my network yet, so what follows is my forecast, not a measured case. We might soon see newcomers with strong technology, real delivered value and lower prices, or cleverer outcome-based pricing, taking share fast, and believe me, trust follows delivery. Warren's numbers: traditional services firms top out around 30% margins and the AI-native bet is 50% plus. That gap is what pays for the price pressure coming at you, or for a lot more spending on quality and better outcomes for clients.

A client does not stay because of your history with them. They stay because that history keeps being worth more than the cheaper, faster option in front of them. The day it stops being worth more, the relationship you spent a decade building becomes the other firm's onboarding story.

So treat the trust you have as a runway. It pays for the transition: the training hours, the slow stretch, the failed first tries, the pivots to different workflows, even the waiting when the technology is not there yet. A runway is for taking off and it's a bad place to park.

## What to do Monday

**Fund the exploration hour.** Every firm has a few people who try the tools, help with the setup, and stay open when something breaks. Find those people and give them a real hour a week on the clock, booked like client work is booked, where a failed first try costs them nothing. Adoption fails on stolen hours, not on resistance. [Your team isn't resisting AI; nobody gave them the hours](https://rodamora.com/blog/your-team-isnt-resisting-ai-nobody-gave-them-the-hours).

**Budget the training that sticks.** The pilots that stick get 8 to 10 hours of training spread over about 3 weeks: video courses, in-person sessions, Q&A, and live support from the technical team. That is pulled from my franchise network's delivery data, it is what worked there and not a minimum I am promising you. The version that dies is the single afternoon demo.

**Warn the team about the slow stretch.** Getting from checking everything by hand to a gate you trust takes a few weeks to a few months, and teams often run slower while it happens. I have watched enough rollouts now to expect it, and the firms that pushed through are the ones that got the gain. Say it out loud before the pilot starts, or the first slow month gets read as failure.

**Start on cheaper models.** Under $500 a month buys hundreds of deliverables while the firm learns where the expensive models earn their price. Move up only on the tasks where the acceptance data says they pay for themselves.

**Learn the tools yourself, if you lead.** This is the one I cannot make easier for you. Management has to know the tools well enough to handle the problems when they come, because they come. Hand it to the enthusiast and go back to your calendar, and the team reads that as the real policy, whatever you said at the kickoff.

## Where this stops

Two conditions, and I will state them as facts.

First, the firm needs some slack, money and hours for AI training, however small. A firm with zero room for that is the firm Warren's warning describes, and no framing changes it.

Second, someone in leadership has to own the transition. Usually the enthusiast can carry the tools. They cannot carry the authority.

You can make this climb, I have watched firms do it. Just don't park on the runway.