I want to write down the things I’d tell myself if I could go back 24 months and start the AI Office practice over. Most of these I learned the hard way. Some of them I’m still learning.
The tools matter less than I thought. The workflows matter more.
In 2024 I spent too much time evaluating models. Which one to use, which features each had, which would last, which would commoditize. By 2026 most of those questions don’t matter — anyone can use Claude or ChatGPT or Copilot, the capabilities are converging, and the differentiator isn’t tool choice.
What matters: the workflow design on top of the tool. How the inputs flow in, how the outputs route to humans, where the approval gates are, what the measurement loop is. None of that depends on model selection. All of it depends on senior judgment.
If I’d known this earlier, I’d have invested less in tool evaluation and more in workflow pattern libraries.
The Champion question is the single most important variable.
I wish I’d put this as a hard gate from day 1 instead of letting it be a “we’ll figure it out” item at signing. The engagements that went sideways in 2025 all had a Champion problem. The engagements that worked had a clear Champion at signing.
If I could fix one thing about how we sold in our first 24 months, this would be it. We’re fixing it now.
Operators learn faster than I expected.
I worried that mid-market operators would be slow to adopt AI fluency — that it would take 6 months to get a team comfortable, another 6 to build real capability. Wrong. Most operators get fluent in 4–8 weeks when they’re using AI on their actual work. The barrier was never their capability. It was the framing — most “AI training” is theoretical, and theory doesn’t transfer.
When we shifted the AI Crash Course toward “build your own AI on your real data,” adoption time collapsed. People come into Day 1 unsure, leave Day 2 capable, and are running real workflows by Day 30.
Owner-led businesses move faster than I expected.
I thought PE-backed operators would lead. They didn’t. Owner-led businesses where the owner personally decided to start moved faster — sometimes weeks faster — because the decision cycle was shorter. The PE governance that I thought was an advantage turned out to be friction.
This changed how we sell. We lead with the owner conversation, not with the PE Operating Partner conversation. The OP conversation matters at the multi-quarter program level; it doesn’t matter at the individual-engagement level for owner-led businesses.
The 12-month KPI guarantee was more powerful than I thought.
When we launched the guarantee on Value Sprints with a committed KPI, I assumed it would close 5–10% more deals. It closed materially more — probably 25–30% more — and shortened cycle time. It also raised our operational bar in a useful way. We can’t be casual about scoping anymore; the guarantee forces discipline we’d have wanted anyway.
Should have launched it earlier. We held it back too long while we got operationally ready. Some clients would have been better served if we’d offered it 12 months sooner.
Buying tools instead of building capability was even more entrenched than I thought.
I expected mid-market operators to be curious about tools. They were also saturated with them. Routinely walking into engagements where the company is spending $30–50K/year on AI licenses with zero business impact. The first conversation in those engagements is usually “stop adding tools; let’s redesign one workflow.”
That stance felt strange the first few times. Now it’s the default. Most mid-market companies need to subtract from their tool spend, not add.
The “small win first” rule is harder to follow than it sounds.
Every client wants to start with the most ambitious AI project. We tell them not to. We tell them to ship the smallest, fastest-payback workflow first to build muscle. About half listen. The other half start with the ambitious project and stall by month four.
I wish we’d been firmer about this earlier. We’re firmer now. If a client insists on starting with the moonshot, we usually decline the engagement.
Hours commoditize.
We launched AI Office with hours on the public pricing page. The reasoning was transparency. The result was that buyers compared us on hours-per-dollar like we were an hourly consultancy. That’s not what AI Office is. We sell judgment, not hours.
Pulling hours off the public page in May 2026 was the right move. Should have done it earlier.
You can’t move slower than the team’s appetite, but you can move faster than they think they want.
In every engagement, there’s a moment around week 6–8 where the team’s enthusiasm crashes. They’ve seen the first build. They’re tired. They want a break before the next one. If you let the cadence slip there, the engagement loses momentum and never recovers.
The right move is to keep the cadence — start the second build before the team thinks they’re ready. They’ll grumble. They’ll also stay engaged. Slow cadence kills more engagements than fast cadence does.
The work compounds.
This one I’m only now appreciating. After 100+ engagements, the patterns repeat enough that engagement #101 is faster, cleaner, and higher-quality than engagement #1. The workflow library, the prompt patterns, the change management approaches, the scoping discipline — all of it is faster the second hundred times than the first.
This is the actual moat. Not the technology. Not the brand. The accumulated pattern recognition from running enough of these.
The implication: a Frogslayer engagement in 2027 is materially better than the same engagement in 2024 — even at the same price. The clients who started with us in 2024 got the early-vintage version. The clients starting now get a more refined product. That’s good for them, and the curve keeps going.
The ask.
If you’re an operator working on AI and any of this resonated, I’d love to compare notes. Reach out directly, or grab a 30-minute call.
— Ross