In March, I built a small AI agent to make my training decisions. About 500 lines of Python, and the first API pipeline I had ever built. It pulled my training data, compared it with the race I was preparing for and told me where the next hour of training should go.
Then the race graded it. My first triathlon, Ironman 70.3 Musselman: 4:28:42, 3rd of 120 in my age group, 34th of 1,908 overall.
The time isn't the interesting part. Where the hours went is.
Start from what already compounds
My profile going in was lopsided. I run about 20 km a week. I had swum 53 times in my life, 37 of them this year. And I have ridden 89,000 km on a bike since 2019.
The obvious plan would have been to train everything harder. The agent's job was the opposite: stop rebuilding what already compounds, and put every marginal hour where I had nothing. The bike needed maintenance, not more volume. The swim needed almost all of the new hours.
Most businesses sit on the same shape of asset. Years of data live in the accounting system, the CRM, service logs and spreadsheets. When a technology or AI project starts, the instinct is to start over: a new platform, a new data model, a clean slate. That's often where the budget goes, and it's often the wrong place.
The return is in finding what already compounds and funding only the gap.
Measure the handoffs
The second lesson came from the splits. In a triathlon, the clock keeps running between disciplines, from swim to bike and from bike to run. My two transitions cost me 9:10, against a benchmark of about 4:30.
That's more than four and a half minutes lost not in the water, on the bike or on the run, but between them. It was more time than any fitness I could realistically have bought this season.
Businesses lose time the same way. The waste is rarely inside a process; it's between processes. The quote that waits in an inbox before it reaches production. The order re-keyed from the CRM into the ERP. The month-end report rebuilt by hand because two systems don't share a customer number. Each step looks efficient on its own. The handoffs are where the hours disappear.
They are also the cheapest to fix. Improving a step usually takes new tools or new skills. Fixing a handoff often takes a rule, an integration or a single automated transfer.
What the agent actually did
There was nothing magic about it. It did three things a good advisor does:
- It started from real data, not from a generic plan.
- It compared that data against the goal to find the largest gaps.
- It recommended where the next hour should go, and what to leave alone.
That is also an honest description of a useful first AI project in a company. Not a platform, not a strategy deck: a small tool, built on data you already have, that answers one recurring question better than the current method. A first version can be built in weeks, tested against real decisions, then kept or dropped.
A race result isn't a business case, and I won't pretend otherwise. But the logic transfers well: know your baseline, find the gap and look hard at what happens between the steps.
The rule
Before you fund a technology or AI project, answer two questions in writing.
What already compounds? List the systems, data and skills that already work and that the project should build on rather than replace.
Where are the handoffs? Map where work passes from one person, team or system to the next, and estimate the time or effort lost at each one.
If most of the budget goes to rebuilding what already works, and none of it goes to the handoffs, the plan is aimed at the wrong target.
Checklist before you fund the project
- The data you already have, where it lives and who maintains it
- What already works well and should be kept as is
- The two or three largest gaps relative to the goal
- The handoffs in the process, with the time lost at each one
- The cheapest fix for the most expensive handoff
- A first version small enough to test on real data within weeks
- The measure that will show whether it worked
Where to start
You probably don't need to start over. You need to know what is already compounding and where time leaks between the steps. If you'd like a second pair of eyes on that, a 30-minute call is a good place to begin.