99% Adoption and Still Stuck

You can get your whole org using AI every day and still feel like things aren't moving.

Maybe that sounds surprising. I've seen it happen. This week I spent an hour walking a room of product and engineering leaders through why it happens and how to get past it.

Ramp, the payments company, got to roughly 99% internal AI adoption. I don't know how they measured 99% but let's just take it as "a lot." And the line that stuck with me from their write-up was that it still felt like driving a Ferrari with the handbrake on. The capability was all there. The result wasn't.

They concluded that expecting everyone to grapple with their own setup, MCP servers, CLIs, harnesses, all of this new machinery, was too much to ask. The people who got it working were unicorns, each with a bespoke configuration nobody else could reproduce. The organization never systematized any of it, so the upside stayed with the few who already knew how to configure their environment. Most people didn't get to benefit.

I've also found that the tooling is only half of it. Even once the setup is handled, teams split into two groups. One group points AI at the work they're already doing. Better emails. Faster decks. Cleaner first drafts. Real improvements, and the artifacts look nicer, but nobody feels the 10x that everyone online keeps promising. The other group figures out a better way to work. They assume everyone has the tools, then ask whether the old way of working still makes sense at all. Whether the thing they used to pass around is even the right thing to be making anymore.

Nobody starts there. Reproducing what you already do is the obvious first step. Re-imagining how the work flows is harder, and it shows up in no adoption metric.

The tools changed. The jobs didn't. I come back to that constantly, in my own work and with teams that feel behind.

Product is still responsible for connecting what we do today to where the organization wants to be tomorrow. Engineering is still responsible for getting us there. Those haven't moved. What moved is who can produce what. Product managers are spinning up prototypes. Engineers are drafting PRDs. Support people are shipping working code. The artifacts are crossing lines in ways they never used to.

For most of my career, personal and team capacity was an unavoidable constraint. I could only chase so many things, so my lane was defined by what I could realistically pull off. We don't have infinite capacity now but laws we previously thought were absolute clearly have some give. You could ship software, publish a point of view, and record a TikTok dance before lunch if you decided those were how to connect with the people you're trying to reach. The problem inverted. It used to be "what can we actually do." Now it is "we could do almost anything, so what should we do." Prioritization matters more than it ever has.

The re-imagined version is mostly a loop you tighten one turn at a time.

I start by asking the tool (Claude Code or Codex for me) to just do the thing once. Run the analysis on this batch of voice-agent calls. Good, that worked. Then I save it as a command I can run whenever I want. Then, once I trust it, I schedule it, so the analysis shows up on a cadence without me touching it. Then I point the next step at the output. Based on this analysis, what should we go address? Now there is a small loop running a basic slice of the work on its own.

Stack enough of those and you stop touching every line. You start building the system the work runs inside. The people getting real leverage here aren't the ones with the cleverest prompts. They are the ones who stopped doing the task and instead built the thing that runs it. I put "agentic" in air quotes when I present these things, because the word covers everything from "I automated one command" to "an Iron Man Jarvis-like system that runs my whole life," and the gap between those is enormous.

None of this is a vision-deck promise. The numbers are public.

Intercom's R&D team doubled their output, measured in lines of code and merged pull requests, while breaking changes dropped about 35%. The second number matters, because the fear with robot-written code is that you drown in quality problems. They didn't. But they also didn't get there by turning the faucet on and walking away. Generative AI is really good at generating things. They built the continuous-integration and review system to keep up with the volume, so the extra output didn't just turn into a pile of bugs.

Salesforce's engineering team took a migration scoped at 231 person-days and shipped it in 13. Some of that is simply that orchestrating agents is easier than orchestrating people. Hand a 231-person-day project to a coordinated set of agents and nine women can sorta make a baby in one month, because the testing and the hand-offs are far more controllable. And they didn't get that speed by skipping the checking. They wrote the migration rules down, fed every round of review back into them so accuracy compounded, and came out with fewer incidents. This is not day 1 functionality but it is absolutely achievable.

In both cases the constraint is no longer generation. The spigot is open. The bottleneck moved downstream, to review and verification. That is the most important thing to automate now. Automating that is how the handbrake comes off. Skip it, and doubling your output just routes everything through the same human reviewers. You didn't take the brake off. You handed it to them.

The strongest teams also turn this into team success rather than individual success. That is the part Ramp couldn't crack at first, when the breakthroughs lived with a few unicorns nobody could copy. Shared meeting notes, shared PRDs, shared customer data, all of it reachable. When a new person joins, they can spin up their own assistant against that shared context and get the state of the world in an afternoon instead of a month. So much of product management used to be the manual labor of getting everyone onto the same page. Lift that off, and the hours go back into deciding what is actually worth doing.

The urgent has always eaten the important. You grind through the prep, the synthesis, the today-fire, and the strategic work waits for a calmer week that never arrives. AI is genuinely good at the grinding part. Which means the work only you can do finally gets the time.

The tools changed. Your job didn't. You just got a lot more room to actually do it.

References

• Ramp's 99% adoption, the "Ferrari with the handbrake on" line, and the setup/enablement conclusion (most people couldn't configure their own environments; the few who could had unshareable setups the org never systematized): Seb Goddijn / Ramp, "We Built Every Employee at Ramp Their Own AI Coworker" (the "Glass" write-up), announcement; accessible summary in Michael Burnett, "Ramp: The Rise of the AI-Enabled Operating System"
• Intercom's R&D doubling output while breaking changes fell ~35%: Intercom Engineering, "AI is approving our pull requests: Here's how we made it safe"
• Salesforce shipping a 231-person-day migration in 13 days: Salesforce, "How Salesforce Engineering Became Truly Agentic"