The short answer.

Quality stays up when AI prepares the work because the check is designed into the workflow, not left to whoever happens to read the draft. The workflow tests its own draft against a written standard and marks where each fact came from; one named reviewer then compares the draft with its sources using a short checklist, and sends it back for a reason from a fixed list. Each send-back changes the workflow, so the quality of the work rises month by month instead of depending on how careful someone was that day.

The principle underneath is simple: a person decides; a workflow prepares. People and AI on one team: who does what sets out that split for every piece of work. This piece is about the moment where the two meet: the review, and how to make it quick, consistent and hard to skip.

Write the standard before the first run.

A workflow’s quality standard is a written page that says what a finished piece of work must contain, what it must never contain, and which sources it may draw from. Without that page, every reviewer checks against their own taste, and two reviewers pass different drafts. With it, the workflow and the reviewer check against the same thing.

  • Must contain.The sections, figures and answers every finished piece has, in order.
  • Must never contain.Anything the company does not say: a commitment only a principal makes, a name not cleared for use, a figure with no source.
  • Approved sources.The files, exports and past documents the workflow may take facts from, and nothing else.
  • One good example.A finished piece the reviewer would sign today.

The standard is written by the person who reviews the work today, because that person already holds it in their head. How to hand a piece of work to AI covers writing down the rest of the work around it.

The workflow checks its own draft first.

A well-built workflow runs its own check before a person sees the draft. After preparing the work, the workflow reads its draft against the written standard and lists what it found: each required section present or missing, each fact with the source it came from, and each place where it was unsure. That list goes on top of the draft.

The self-check does not replace the reviewer. It points the reviewer at the places worth reading. A reviewer handed a clean draft reads it end to end and still misses things. A reviewer handed a draft that says “two figures have no source, one requirement is not answered” goes straight to those three places first. We build this step into every workflow we develop with a company, in its own ChatGPT, Claude, Copilot or Gemini.

What the checker looks at, and against what.

The reviewer checks four things, and each one is checked against something written down, never against memory.

  1. Facts, against the sources.Every name, date, figure and reference in the draft matches the file or export it came from. The reviewer opens the source, not just the draft.
  2. Completeness, against the standard.Every section the standard requires is there, and every question the request asked is answered.
  3. Limits, against the never list.Nothing in the draft commits the company to something only a person may commit to.
  4. Fit, against the request.The draft answers this client, this project, this month, and not the last one the workflow saw.

The judgment calls stay with the reviewer: whether to send, what to promise, what to charge. The workflow leaves those places marked and blank.

What sends the work back.

A draft goes back when it fails one of the four checks, and the reviewer records which one. A fixed list of send-back reasons turns a reviewer’s irritation into something the workflow’s owner can act on.

  • Wrong fact.A fact does not match its source.
  • No source.A fact appears with nothing behind it.
  • Missing part.A required section or answer is absent.
  • Over the line.The draft says something only a person may decide.
  • Wrong case.The draft carries over material from another client or project.

The owner of the workflow reads the send-back log and changes the workflow so the same reason stops appearing: a tighter instruction, a source added to the approved list, a new line in the self-check. The reviewer fixes the draft; the owner fixes the workflow. The log comes to the monthly AI review, where “is it useful” gets an answer from the record instead of from opinion.

A review checklist to copy.

Copy these headings onto one page for each workflow, and keep the page beside the workflow’s instructions. Fill it in with the person who reviews the work today.

  1. Workflow and owner.The workflow’s name and the one person who changes it.
  2. Reviewer.The one person who reads every draft before it leaves the team.
  3. Approved sources.The files and exports the workflow may draw from.
  4. Must contain.The required sections and answers, in order.
  5. Never without a person.What the draft leaves blank for the reviewer to decide.
  6. Self-check the workflow runs.What the workflow lists on top of every draft.
  7. Facts the reviewer opens the source for.The names, figures and dates checked every time.
  8. Send-back reasons.Wrong fact, no source, missing part, over the line, wrong case, and any your work adds.
  9. Where the log lives.One place where each send-back is recorded with its reason.
  10. When the owner reads the log.A date each month.

One engineering firm builds the check into its proposals.

Take a 40-person civil engineering firm that answers requests for proposals from public agencies and private developers, and sends each client a monthly project report.

Illustrative example.

  • The work.A proposal workflow reads the request and prepares a draft from the firm’s approved project sheets, staff résumés and standard scope language. A principal decides what to bid and signs every proposal.
  • The standard.The principal who reviews proposals today wrote it: every requirement in the request answered in the order asked, only projects from the approved list cited, only staff on the current résumé list named, and the fee and schedule commitments left blank.
  • The self-check.On top of each draft, the workflow lists each requirement with the section that answers it, each cited project with the sheet it came from, and anything it could not match.
  • The review.The principal opens the request and the project list beside the draft, fills in the fee and schedule, and either signs or sends it back with a reason.
  • The send-backs.Early drafts came back twice for “wrong case”, with scope language carried from a different kind of project. The proposal coordinator, who owns the workflow, split the scope language by project type and added a line to the self-check.

The monthly project report follows the same pattern: the workflow drafts from the schedule and hours exports, lists every date with its source, and carries over each open item from last month’s report. The project manager checks the dates against the schedule and signs it.

What is different in a company of under 200 people.

In a company of under 200 people, the reviewer is usually the person who did the work before the workflow existed, and that is an advantage. A large company sets up a separate quality function. You already have the right reviewer, who knows the standard and can write it down in one sitting.

  • One reviewer, by name.There is no queue of reviewers to share the load, so name one person and a cover for leave.
  • The owner and the reviewer are often the same person.That is fine. Keep the log anyway, so the fixes to the workflow are written down.
  • The standard is short.One page per workflow is enough, and it doubles as training for the next person in the role.

Put the check on the plan alongside the workflow it belongs to. FullyPowered Planner is where the plan lives: you set out your teams, their roles and the work between them, choose what to build first, and your colleagues add their own teams. How organizations plan their use of AI covers the whole method: plan it, develop the workflows, deploy them. For work where the checking itself is the job, see checking the work.