An AI project should run like a construction job

Anthropic, the company behind Claude, recently published a playbook for how software teams should work with AI. It is written for developers. Most business owners will never read it, and they do not need to. The best idea in it is one owners already understand better than most software teams do, because construction has run on it for a century.

The bottleneck moved

The playbook opens with an observation: the building part of software got fast, and everything around it did not. AI writes working code in hours. The reviews, approvals, and handoffs around that code still move at meeting speed. So the slow part of a project stops being the work itself. The slow part becomes deciding, checking, and handing off.

If you run a trade business, a practice, or a firm, you know that shape. Your crew can rough-in a house in days. The permit, the inspection, and the change order are what stretch the job. AI is now doing to office work what power tools did to field work: the doing gets fast, and everything waits on the approving.

The rule worth stealing

Anthropic's answer runs through the whole playbook: every stage of a project ends in a written artifact, and the next stage starts from it. Their versions are files with names like intent, spec, and plan. Construction has older names for the same chain: the inspection report, the drawings, the building, the maintenance log.

Notice two things nobody in construction ever questions. You pay for drawings, and the drawings are yours even if you never break ground. And nobody frames a wall from a verbal description. The work waits for approved drawings, because fixing a drawing costs a pencil and fixing a framed wall costs a crew.

Software teams needed a playbook to adopt that discipline. Owners already live it.

How we run the same chain

Our method follows the construction sequence on purpose.

• The assessment is the inspection. We walk your operation and write down what we found: where time and money leak, and which fixes are worth doing first. You keep the report. It is useful even if you never hire us again.
• The map is the drawings. We interview your team, trace how work actually moves through the business, and write it down. That document becomes the first layer of your Company Brain, and it is yours. Another builder could work from it. Good drawings stand on their own, which is why the map is priced as its own deliverable and not as a deposit on a build.
• The build follows the drawings. Nothing gets built without a plan you approved. Scope is fixed before work starts, and the finished system is checked against the plan, the way a final walkthrough checks the building against the drawings.
• Run is the maintenance contract. A system that touches real customers and real schedules needs someone watching it, the same way a building needs someone changing the filters. Each review also surfaces the next thing worth building, the way a good maintenance visit catches the next repair while it is small.

What this protects you from

Plenty of AI work is sold as a subscription to someone else's magic. The project starts with a demo and ends with a login, and if the vendor leaves, everything leaves with them.

The construction test cuts through that. Ask any vendor, including us: what do I hold in my hands at the end of each stage? If the answer at some stage is nothing, ask what happens when they leave. A project that ends every stage in something you keep is a project you can pause, resume, or hand to someone else. That is what owning the work means.

Anthropic wrote the playbook so software teams would work this way with AI. Your industry never stopped working this way. When you buy AI work, hold it to the standard you already know: inspection, drawings, construction, maintenance, and something you keep at every step.

Related: The build is now the cheap part