Building Easier AI

Where AI Actually Pays Off in a Business That Builds Things

Every owner I know who builds something for a living has now been pitched AI at least a dozen times. The pitch is always the same shape: a transformation, a journey, a platform. It never survives contact with a Tuesday.

The problem isn't that AI doesn't work in businesses like ours. It's that the people selling it have never cut a purchase order, never chased a supplier for a price that moved, never had a superintendent quit in framing. They're selling a category. You need a payback.

So here's the test I use — on my own companies first, and now on everyone else's.

The question isn't "where could AI help"

Ask a room "where could AI help?" and you'll get forty answers, all of them plausible, none of them ranked. AI could help almost anywhere. That's exactly what makes the question dangerous: an unranked list of good ideas is how you end up eighteen months in with three half-built tools and nothing that closed a single hour.

The better question is narrow and boring: where in this business does a person spend real hours doing something repetitive that still requires a little judgment?

That sentence has three conditions in it, and a task has to clear all three.

It has to be frequent. Something that happens forty times a week is worth automating. Something that happens twice a year is worth writing down, not building. Frequency is what turns a small per-instance saving into a number you can see on a P&L.

It has to require judgment — but teachable judgment. Pure data entry doesn't need AI; it needs a better form, and it probably needed one five years ago. The sweet spot is work where somebody reads something, decides something, and writes something. Reading a plan set and pulling a takeoff. Reading a supplier email and deciding whether the price change affects an open bid. Reading a client's message and deciding whether it's a change order.

Somebody has to be doing it now, and that somebody has to cost money. This is the condition people skip, and it's the one that makes the whole thing real. If no one is doing the task today, automating it doesn't save anything — it adds a capability, which may be worth having, but it isn't a payback. Payback requires a line item that shrinks.

Clear all three and you have a candidate. Clear only two and you have a project that will feel productive and return nothing.

Walk the business the way you'd walk a site

You don't find candidates like that from a conference room. You find them the way you find a framing problem — by walking it and asking the person doing the work.

Spend a day with each seat: the estimator, the PM, the office manager, the field lead. Ask one question and then shut up: what part of your week do you dread?

People will tell you exactly where the money is leaking. They dread the parts that are repetitive and still require attention, because those parts don't let them think and don't let them coast. That dread is a remarkably accurate sensor. Write down every answer. Don't filter yet.

Then rank what you wrote by payback, not by how interesting it is. The interesting ones are almost never the profitable ones.

Do the arithmetic before you spend a dollar

Here's the part almost nobody does, and it takes ten minutes.

Take one candidate. Purchase orders, say. Count how many you cut in a normal week — call it 40. Time yourself honestly on one, start to finish: pull the scope, find the vendor, look up current pricing, key it in, route it. Say it's nine minutes.

That's six hours a week. At a loaded PM rate of $45/hour, it's about $270 a week, or $14,000 a year, on one task, in one seat.

Now ask what a tool would realistically take out — not to zero. Say it drafts the PO and a human approves it in ninety seconds. You've recovered about four and a half hours a week: roughly $10,500 a year.

That's your payback ceiling for that one item. If a build costs $12,000, the honest answer is that it pays back in about fourteen months, and you should probably do something else first. If it costs $4,000, you're whole in five months and everything after is margin.

Run that arithmetic on every candidate on your list and something clarifying happens: most of them die. The two or three left standing are your actual roadmap, and you can defend them to your banker.

What we built, and what it taught us

We ran this test on our own companies before we ran it on anyone else's.

The PO example above isn't hypothetical for us — it's the reason we built an agent that drafts vendor-priced purchase orders by voice, from the truck, so a PM cutting a PO on a lot doesn't have to be back at a desk to do it. We built a permit tracker that syncs municipal permit status automatically instead of somebody refreshing county portals. We've got thirteen custom skills in production running estimating, scope-of-work drafting, and plan revisions.

None of that started with "we should use AI." All of it started with somebody saying they dreaded a Tuesday.

The other thing it taught us: about half of what clears the test shouldn't be built by an outside firm at all. If your own team can be taught to build it, they should — they'll own it, they'll fix it at 6pm when it breaks, and they won't need us for version two. The other half is genuinely worth handing off, because it touches your data in ways that need doing properly the first time.

That's the whole fork. The assessment tells you which door each item goes through. Anyone who tells you which door before walking your business is selling you a package, not a plan.

The uncomfortable part

Some of what you find won't have an AI answer. You'll walk the business and discover the estimator dreads Tuesdays because your cost catalog hasn't been updated since 2023, and no model on earth fixes that. You'll find the PM is drowning because two roles got merged during a slow quarter and never got unmerged.

Good. Those are cheaper to fix than anything on the list, and finding them is worth the walk by itself.

The point was never to use AI. The point was to find where the hours go, and then decide — honestly, with arithmetic — which ones are worth buying back.

Start with the walk. The punch list comes out of it, ranked by payback, and then you'll know what to build.