Soundly RunSound familiar?

AI got it wrong. Then what?

The AI made a mistake. The review stalls: nobody can say which role, setting or model acted.

InsteadEach AI action, allowed or blocked, records who allowed it, under which rule and model.

The video

Video planned53 secondsHow every video is made
The script
The one point
Every AI action, allowed or blocked, leaves a receipt naming its role, owner, rule and model, so who authorised it has an answer on the record.
Why a video
The same incident review shown twice: days of rescheduled meetings and chat scrolling, then one Record screen holding the answer, and the owner’s click held live.
  1. Intro0:00-0:03sting

    Picture
    Brand mark and sound, then the heading ‘AI got it wrong. Then what?’ on the site’s background, fading straight into the guide with no black.
    On screen
    AI got it wrong. Then what?
  2. The moment0:03-0:10person

    Picture
    Guide in the usual place and framing, speaking to camera. Lower third on first appearance.
    Voice
    The AI made a mistake. The review stalls: nobody can say which role, setting or model acted. When every helper acts under one shared account, no act has an owner.
    On screen
    Guide’s first name, Soundly Run guide
  3. The cost0:10-0:17motion

    Picture
    Chat threads scroll, one shared service account on every line; question marks hover over Role, Rule, Model. Calendar pages flip as the same invite keeps reappearing: ‘Incident review (rescheduled)’. No number.
    Voice
    So the review rebuilds events from chat history, and the same meeting keeps moving.
    On screen
    Incident review (rescheduled)
  4. The mechanism0:17-0:34screen

    Picture
    Drawn Record screen, Example tag top left, context ‘Jo’s crew’. Step band prepare, review, decide. Cursor walks in order: Act flagged, invoice 4471 paid twice; Role, owner Jo; Rule, version 3, set by Jo; Model; Cause; Drafted change. Side panel: 212 allowed, 3 blocked, each with a receipt.
    Voice
    It paid the same invoice twice because nothing compared them. The receipt shows the role and its owner, Jo; the rule, version three, and who set it; and the model.
    On screen
    Example: invoice paid twice, and why.
  5. The decision0:34-0:42person and screen

    Picture
    Apply the change, Change and Pass visible; guide in picture-in-picture. Cursor on Apply the change, click held two seconds. Rows turn green: ‘Version 4: Jo confirms this vendor's invoices’, ‘Duplicate match before scheduling’.
    Voice
    Jo, the role’s owner, decides: apply the change, change it, or pass. The old receipts are kept as they were.
    On screen
    Applied by Jo. Receipts name version 4 from today.
  6. The difference0:42-0:49person

    Picture
    Guide back in the same place and framing as The moment.
    Voice
    Nobody rebuilds events from chat history any more, so who authorised it has an answer on the record. The helper acts alone only inside limits Jo set; anything else asks first. It does not watch, score or replace anyone.
    On screen
    Each AI action, allowed or blocked, records who allowed it, under which rule and model.
  7. Outro0:49-0:53sting

    Picture
    Standard end card, the same in every video: the page title, the brand mark and one small line, ‘Example screens’.
    On screen
    AI got it wrong. Then what?

On screen as proof

  • Example tag on the drawn Record screen throughout the mechanism
  • Rows naming role, owner Jo, rule version 3 and who set it, model and cause for invoice 4471
  • Side panel counting allowed and blocked acts, each with a receipt
  • Apply the change click held two seconds, then ‘Applied by Jo. Receipts name version 4 from today.’

Before and after

Before
Helper acts logged under a shared service account
After
Each act logged with its role, rule version and model
Before
Owning a helper means owning its mistakes, so accountability goes unclaimed
After
Every role has one named human owner
Before
Incident review reconstructs events from chat history
After
A read-only record, with any later edit shown beside the original, answers who allowed it

How it works

Every AI action is tied to its role’s owner, the rule it acted under, and who set it.

  1. Each role's owner decides what its AI may do alone, as a rule with a version and date.
  2. Every action, allowed or blocked, leaves a read-only receipt: the role, the rule and version, who set it, the model.
  3. When something goes wrong or an auditor asks who authorised it, the receipt answers; the owner decides.
RecordExample

Jo's crew

Invoice paid twice, and why

Act
Paid invoice 4471 twice, 12 Oct
Role
Accounts payable. Owner: Jo
Rule
Alone under 5,000 · version 3, set by Jo, 1 Oct
Model
Fast tier, 1 Oct version
Cause
Arrived by two routes; nothing compared them
Drafted change
Jo confirms this vendor's invoices; add duplicate match

Source: Record and ledger, 12 Oct

Apply the changeChangePass

Built to work in the email, calendar, chat, finance and sales tools you already use.

What it never does

  • Acts alone only inside limits a person sets; asks first for anything else.
  • Reaches only what its job needs.
  • Never watches, scores or replaces anyone.