Brief sections
The leak
Margin by job comes out around the 20th of the following month, after someone exports from the ERP, cross-references the time tracker, and reconciles the spreadsheet where change orders live. The job that went sideways in week one has been sideways for six weeks by the time anyone sees it in the numbers. So decisions get made on feel, the owner can’t step away because they’re the only real-time sensor the business has, and the quarterly review becomes an exercise in explaining variances nobody can fully reconstruct. The report isn’t wrong; it’s late — and late is a form of wrong.
What AI realistically does here
It pulls from the source systems on schedule, assembles the operating picture you’ve defined — not forty charts, the six numbers you actually act on — and delivers it daily or weekly with anomalies called out in plain language: which job’s labor ran past estimate, which customer’s orders fell off their usual pattern, which receivable just aged past your line. Plain language matters — a sentence with a number in it gets read; a dashboard gets a bookmark. The honest limitation: if the underlying data is wrong, faster reporting just delivers wrong numbers sooner. Data-hygiene problems surface in week one, and sometimes fixing those is the actual project.
What has to be true first
- The source systems are queryable — exports, saved reports, database access, an API. “Ask the bookkeeper to run it” is a data source with three-day latency and a vacation schedule.
- A short list of numbers you’d actually act on, with thresholds. “Everything” is not a report; it’s a landfill.
- A reader. If the brief goes out and nothing ever changes because of it, the leak was never reporting.
What we'd install
Reporting builds usually rank high in the opportunity portfolio because the data work they force — connecting systems, defining the numbers — is the same foundation every later agent stands on. The metric definitions go into the SOP library so the numbers mean the same thing next year, the reporting agent runs on cadence with a named owner, and its contribution to the value ledger is the one it makes possible for everything else: problems caught in days instead of weeks. Where it lands in the sequence is on the AI operating partnership.
Worth it?
Time how long your most important operating number takes to produce today, and count the hands that touch it. Then estimate what one sideways job or one quietly shrinking account costs when it runs six extra weeks unseen — that’s the real price of latency. The calculator below covers the labor side with an illustrative annual figure; the scorecard tells you whether your systems can feed the report at all.