Teach the agent how work actually runs.

A concept for Edra’s founding product designer: the gap between the process on paper and the process in practice, made visible, reviewable, and safe to hand to an agent.

Open the process lab
Vendor invoice approvalAccounts payable · Sample process
The agent has only read the SOP. Learn from real work to see how this process runs.
Documented stepLearned stepRisk controlLine thickness shows how often a path is takenSwipe the map sideways to see every step
ReviewTeach the agent

Select a step on the map to see the evidence behind it and decide what the agent should do.

AgentINV-20417
Readiness 50%Learn from real work first

High-risk controls always keep a person in the loop. Everything else is yours to promote.

    Three design bets, and how I would test them

    Enterprise AI fails when people cannot tell what the agent learned or why it acts. These are the bets behind this concept.

    01

    Show the gap, not the model.

    Process owners trust what they can check. The first screen compares the SOP with what the agent saw, instead of showing a confidence score they cannot verify.

    Test: how long until an ops lead makes their first correction.
    02

    Autonomy is earned, step by step.

    Every step moves from observe to suggest to ask first to automatic, with an audit trail. Risk controls never go fully automatic, and the interface says why.

    Test: share of steps promoted without a rollback.
    03

    Design for the person who owns it.

    They know every exception and have never written them down. The product asks short questions with evidence beside them, and turns each answer into a rule.

    Test: can a new owner explain the agent’s behavior after ten minutes.
    PatternConfidence

    Confidence always sits next to its evidence. Anything below 80% asks a person, and uses amber, never an alarm.

    PatternAutonomy dial

    Week 1Shadow a process owner

    Sit with someone who runs this process daily and write down every exception they mention.

    Week 1Replay twenty real cases

    Run real cases through the agent and mark each place a person would disagree.

    Week 2Watch a first correction

    The first time someone fixes the agent tells me more than any survey.

    Week 2Ship the smallest loop

    One review step, in production, with the engineers, so we learn from real use.

    Let’s teach AI how companies really work.

    I design and build, I am comfortable with open-ended problems, and I am based in New York, ready to work in person with the team.

    leonwprimary@gmail.comView portfolio: leonwayne.com

    Independent concept, not affiliated with Edra. The process, evidence, and numbers are simulated sample data.