Dispatches
Governance — control framework

Autonomy you can audit.


The question every serious buyer asks is: what happens when the agent gets something wrong, and who is accountable? We answer it before the first call. Below is the control register.

Five controls. Always thrown.

Attribution, access, no-training, human-in-the-loop, versioned procedure — enforced in the platform, not promised in a PDF. Work the door if you like; the controls don’t come off.

GOV-01Attribution & reversibility
GOV-02Access governance
GOV-03No training on data
GOV-04Human-in-the-loop
GOV-05Versioned procedure
Bronze-and-navy vault door with a Greek meander ring and eight thrown bolts
Register vault SEALED — GOV-01…05 ENFORCED
Control register GOV-01 — GOV-05 Reviewed quarterly
GOV-01Attribution & reversibility
Every action attributed. Every action reversible.

Every change — human or agent — is attributed, and undo/redo applies to everything the agent touches. This is the documented reason our most regulated customers went the distance.

09:41 Athena edited B4 undo
09:42 D. Rowe approved
09:44 Athena inserted row undo
09:47 D. Rowe reverted v11
GOV-01Attributed activity log
GOV-02Access governance
Inside your access model, not around it.

Agents inherit the same permissions as the people they work for. They can never see or touch what their operator cannot — and access is enforced by your identity provider, not ours.

Finance — General Ledger granted
Finance — Payroll denied
HR — Records denied
Scope inherits operator
GOV-02Permissions panel
GOV-03No training on data
No training on customer data.

Contractual and standard on every deployment. Your data is used to do your work — never to train a model, ours or anyone else’s.

§ 7.3 — Data use Provider shall not use Customer Data to train, fine-tune, or improve any model, and shall process Customer Data solely to deliver the Services to Customer.
GOV-03Contract clause
GOV-04Human-in-the-loop
Co-pilot and autonomous, on your terms.

Move between co-pilot and autonomous modes with review and promotion flows in between. Both workforces work in the same governed artifacts, with full visibility into who — human or agent — did what.

1. Agent drafts co-pilot
2. Human reviews required
3. Promote to autonomous approved
4. Monitor & revert always on
GOV-04Review & promotion flow
GOV-05Versioned procedure
Procedures as governed institutional memory.

AOPs are versioned, auditable, and roll-back-able. When a procedure changes, you see who changed it, when, and why — and you can restore any prior version instantly.

v14 · 07-06 Added region branch
v13 · 06-29 Reverted threshold ← rollback
v12 · 06-21 New approval step
v11 · 06-14 Initial publish
GOV-05Version history & undo
The evidence discipline

Every figure traceable to source.

Athena cites before it states. Each number in a delivered artifact ties to the document it came from — the same vouching discipline an audit team would apply to us.

Workpaper W-4 — every figure, vouched
Revenue, Q2 $4,102,300
S-1 — Bank confirmation №117
Inventory adj. −$86,410
S-2 — Sub-ledger export, 06-30
Freight accrual $156,010
S-3 — Carrier invoice INV-4473
Trace — 3 / 3 figures tied to source Method: cite, then state

It learns the institution — and the institution can audit what it learned.

Workflows codified as versioned procedures. Every version inspectable, every action reversible. The artificial employee gets better at your firm the longer it works there — on the record.