The Last Mile Is Ninety Percent of the Road

The phrase "last mile" is doing enormous damage in your planning meetings, and nobody has noticed because it sounds so reasonable.

It implies proportion. A mile out of twenty-six is a rounding error, the bit you jog in after the hard work is done, and everyone in the room hears it that way. So when your tech lead says "we're basically there, it's just the last mile," the room books the celebration.

Then the last mile takes four months, and nobody can quite explain where it went.

It isn't a mile. In AI work it's most of the road, and the reason it keeps surprising people is that the part that got cheap and the part that got measured are the same part, while everything expensive happens somewhere nobody's watching.

What actually eats the quarter

Generation is now trivial. Somebody prompts, something appears, and it appears fast enough that the team's sense of pace gets recalibrated around it — which is the real problem, because everything after generation runs at the old speed.

Integration runs at the old speed. Your new thing has to talk to systems that predate you, through permissions nobody has documented since the reorg, owned by a team with its own roadmap and no particular reason to care about yours.

Data plumbing runs at the old speed, and usually slower, because sample data was clean and yours is not. It has nulls where the schema promised values, three spellings of the same customer name, and a decade of decisions nobody wrote down.

And then there's the category that didn't exist in your last planning cycle: how do you test something that answers differently every time you ask? A regular test suite asserts that a function returns 4. What do you assert about a system that returns a good answer on Tuesday and a slightly worse one on Wednesday, for reasons neither you nor the vendor can fully explain? Somebody has to build an eval harness, and I have never once seen that work appear in the original estimate.

Generation got cheap. Finishing did not, and finishing is the product.

The tool that held three teams hostage

Years ago I built a tool over a weekend to solve a workflow problem, and it took me about two days.

People started using it that Monday. Within a month three teams depended on it, and within a quarter it was breaking regularly, I was the only person who understood how it worked, and I was spending more time keeping it alive than doing the job I was actually hired for. At one point I worked out that a dashboard I'd built in a weekend was costing me eight hours a week.

Nobody asked for a maintenance plan, including me — especially me.

That tool cost more to unwind than it ever cost to build, and the unwinding was the least interesting work I have ever done — no design, no strategy, just weeks of moving things carefully so nothing broke while three teams watched.

Here's the part that matters for you. Every single hour of that was invisible in every system my organization used to see work. It wasn't on a roadmap. It didn't have a ticket. It never appeared in a sprint review, because sprint reviews show what got built, and this was the opposite of building. It looked, from the outside, like I'd simply become slower.

Yikes.

The generation ratio

So measure the thing nobody measures. Take one initiative that shipped in the last six months and split its calendar time into two buckets: time spent generating, and time spent making the generated thing survivable.

That second number divided by the first is your generation ratio, and it's the most honest number in your delivery system.

If generation was two weeks and everything after it was fourteen, your ratio is 7:1. That's not a failure — it might be perfectly healthy for your context. The failure is planning as though it's 1:1, which is precisely what your organization does every time somebody says "just the last mile."

Nicole Forsgren's work in Accelerate made a version of this argument years before anyone was prompting anything: the teams that ship reliably aren't the ones that code fastest, they're the ones with the shortest path from written to running. AI shortened exactly one segment of that path. It did nothing to the rest, and in the eval-harness case it added a segment that wasn't there before.

Do you know your ratio, and does anybody else in your organization know it either?

Make the second bucket visible or it will keep eating you

The reason the last mile keeps ambushing teams isn't that it's unpredictable. It's that the work is entirely unwitnessed.

Every ceremony you run is optimized to show generation. Demos show new things, sprint reviews show completed tickets, and roadmaps show forthcoming features. When was the last time anybody in your company demoed a migration? There is no forum in your entire operating rhythm designed to show the work of making something survivable, so the people doing that work look idle, and eventually they start believing they are.

Three changes, and none of them require a new process:

Demo the unglamorous thing. Once a month, have somebody demo the eval harness, the migration, the permissions model. Not a slide about it, but the actual thing, running in front of people. It reframes that work as work in front of the people who fund it.

Put the ratio in the planning doc. One line: "generation two weeks, everything after fourteen, ratio 7:1." You will be astonished how quickly estimates stop being a negotiation once the number is written rather than argued.

Name a finisher, out loud, at the start. Not "platform will pick it up," but a person, named in the room, before the build begins — because the alternative is that the finishing lands on whoever is least able to say no.

What to stop

Stop calling it the last mile. The phrase is a sedative and your organization has been dosing itself with it for years.

Stop counting demos as delivery. A demo is evidence that something can exist. Delivery is evidence that it does, for someone who isn't you, on a day you're not watching.

And stop letting the finishing work be the only work in your company that happens without an audience.

Do this in your next planning session

Pick the last thing your team shipped. Split its calendar time into generating and finishing, and write the ratio at the top of the planning doc where the next estimate is about to be made.

Then say the number out loud, once, and watch who in the room already knew it.