Skip to content
Back

Method

Delivery system for AI-assisted client work

The controls around an AI-assisted workflow for client documents: figures you can trace, a person approving anything a client sees, and lessons from live work reviewed before they change the method.

The problem

Client work has to be consistent and safe to show, but one specialist's judgment is hard to repeat across many clients.

Constraints

  • Client data and decisions must never move from one account to another.
  • Every figure in a client document has to trace back to a source and a period.
  • Nothing reaches a client without a person's yes.
  • The work has to carry on when the AI tool is down.

My part

I designed the operating layer around an AI-assisted workflow for client documents: the rules, the checks and the fallback. I designed it on my own. The professional rules in it build on what experienced colleagues have learned, and on their feedback.

Decisions

  • Three layers, kept apart

    Shared professional standards, my own preferences and one client's decisions each live in their own place, so one client's choice never becomes everyone's rule.

  • Labelled by audience

    Every draft says who it's for, so internal notes stay internal.

  • Improve on a schedule

    The method changes on a cadence, not on impulse. Corrections wait for a scheduled review.

  • A written fallback

    A written method describes how the work continues when the AI tool is unavailable.

One client document, step by step

  1. Export

    Data comes in as exports from several tools, and exports are messy.

  2. Check

    Every figure gets a source and a period. A file that can't be read is reported, never skipped.

  3. Reconcile

    Where two tools disagree, it is settled before anything is written.

  4. Draft

    A first draft is written with AI assistance, within the shared standards and that client's own decisions, and labelled by audience.

  5. Approve

    I review every draft. Nothing reaches the client or a shared folder without my clear yes.

  6. Learn

    Corrections are sorted as personal, one account or general. General ones wait for the scheduled review.

Built for real client work. No client material, tool names or internal methods are shown here.

Workflow designQuality controlData validationHuman reviewProcess documentation

Want the details behind this? Email me at filip@heinel.dev.