Fleet 2.7.0 is out.See what's new →
FleetFleet
Use case

AI Workflows for Board Reporting

A board update is high-stakes recurring knowledge work: the same document every month, assembled from sources that change constantly — metrics exports, OKR docs, hiring plans, runway models. Someone senior spends a day collecting numbers and someone more senior spends an evening checking them, because a wrong revenue figure in a board deck is not a typo, it is a credibility problem.

The failure mode of using a chatbot for this is well known: the draft reads fluently and gets one number wrong, and nobody catches it because nobody is assigned to catch it. The problem is not drafting. It is governance — making review and sign-off a structural step instead of a favor.

How it works with an agent fleet

Fleet's workflow engine builds the report the same way every month, so no one rebuilds it by hand: a draft step reads your actual source documents (the corpus), an independent review step checks every claim against those same sources, and an approval gate pauses the run until a named human signs off on the exact document.

Trigger: schedule — 9am on the 1st, monthly

Step Kind What it does
draft Document “Write the monthly board update. Every number must come from the corpus.” Reads metrics/**/*.md, okrs/*.md.
fact-check Review “Check every claim in the draft against the corpus. Flag anything unsupported.”
ceo-signoff Approval gate Pauses the run for a human sign-off; records the approver, timestamp, and note.
send Publish Shares the approved document at a link.

You build this once on the dashboard canvas; your self-hosted Fleet worker pulls the definition and runs it against your repository.

If the reviewer flags an unsupported claim — “draft says revenue grew 40%, the source says 25%” — the run parks at the gate and the approver sees the flagged claims next to the document itself. Approving resumes the run; rejecting stops it. The decision, the approver, and the timestamp are recorded with the artifact.

The fleet pattern

Draft → independent fact-check review → human approval gate → publish. The reviewer is a separate step with its own prompt, not the author grading its own work. The approval email deep-links to a page showing the exact document and the reviewer's flags, so sign-off means someone actually read it.

Guardrails that matter here

  • The approval gate pauses the run until a human decides — there is no path to publish that skips the gate
  • Every decision is recorded with verdict, approver, and timestamp; artifacts are content-addressed and versioned, so you can prove which document was approved
  • Document steps run with all agent tools disabled, so the draft is a document, not an agent session with side effects
  • Runs execute on your own Fleet worker — source documents never leave your infrastructure

Who this is for

Founders and executives who produce a recurring investor or board update from sources that live in (or can be committed to) a repository. The value scales with how expensive a wrong number is.

Frequently asked questions

What stops the AI from inventing a number?

Two structural layers. The draft step's corpus is the only context it gets, and a separate review step checks each claim against those sources and flags anything unsupported. A flagged run parks at the approval gate with the flags displayed — it cannot proceed to publish until a human looks at exactly what was questioned.

Related use cases

AI Workflows for Weekly Status Reports →AI Workflows for Compliance Evidence →AI Workflows for Recurring Data Reports →

Run your first agent fleet

One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.

See how it worksInstall Fleet