Security

Approval-Gated AI Communication Is a Control, Not a Feature

September 6, 2026 6 min readBy Octopus AI
Share
Approval-Gated AI Communication Is a Control, Not a Feature

Most AI in finance is pull-only. Someone opens a chat window and asks a question. Nothing happens unless a person initiates it.

The useful version is push-based. It notices a variance and goes to the budget owner for the explanation, every cycle, without being asked. That is where the time actually goes in FP&A: chasing the "why" by email.

But an AI system that contacts people on the company's behalf introduces a control question the pull-only version never had. Who authorised that message?

The gate is the control

The answer should be structural: every outbound message is approval-gated, and a named human signs off before anything sends. Not a policy statement. A path the system cannot bypass.

That single design decision is what makes autonomous finance AI defensible to an auditor. Without it, an agent that emails a cost-centre owner is an unlogged, unauthorised communication from the finance function. With it, the same action becomes a control activity with an owner, a record, and a reviewable outcome.

What evidence the gate should leave

An approval gate is only worth as much as the record it produces. Insist on all four elements stored together, as one chain:

  • The approver's identity.
  • The exact content that was sent.
  • The recipient.
  • The timestamp.

Split those across systems and you have an email trail again. Kept together, you have evidence: a reviewer can pick any message, see who released it, and see the variance it referred to.

What else belongs behind the gate

Outbound messages are the obvious case. Three more matter as much:

Configuration. Thresholds, recipient lists, message templates, cadence, assumptions, and metric definitions should be customer-configured and versioned, with the approving user recorded. If a model can quietly change what counts as a material variance, the threshold is not a control.

Publication. A rebased forecast released to a wider audience should pass a named human first, and a failed reconciliation should block publication rather than degrade gracefully into a plausible number.

Escalation. Requests outside the system's documented scope should be refused, logged with a reason, and routed to a person.

How this maps to the frameworks

Under COSO, the approval gate is a control activity, the retained record is information and communication, and the sampling of released messages is monitoring. Under SR 11-7, the relevant point is narrower: because the arithmetic behind the message runs as deterministic code rather than model output, model drift cannot change the number the approver signed off on. Under DORA, the logged, gated, read-only architecture is exactly the third-party documentation a regulated institution has to produce.

None of that requires the AI to be less autonomous. It requires the autonomy to end at the moment something leaves the system.

The practical test

Ask any vendor of agentic finance AI two questions. Can it contact a person unattended? And can you show me the approval record for the last message it sent?

Our answers, the boundaries per AFI capability, and what gets logged are published in the Octopus AI Trust Center, with the enforcement detail in our constitution and the model system cards.

Frequently asked questions

What does approval-gated AI communication mean?+

It means an AI system cannot send a message to a person until a named human reviews and signs off on it. The approval is a required step in the send path, not a policy.

What evidence should an approval gate produce?+

The approver's identity, the exact content sent, the recipient, and the timestamp, stored together as a single evidence chain that a reviewer can inspect for any message.

How does an approval gate map to COSO and SR 11-7?+

Under COSO the gate is a control activity, the retained record is information and communication, and sampling released messages is monitoring. Under SR 11-7 the relevant point is that arithmetic runs as deterministic code, so model drift cannot change an approved figure.

Analyze this article with AI

Get an instant executive brief right here, or open it in your preferred AI assistant.

Executive brief — generated in-app

Tailored for CFOs and VPs of FP&A. Takes ~5–15 seconds.

Prefer another assistant? If the prompt doesn't autofill, use the fallback below.

Deep link didn't work? .

Next step

See the full picture of AI in Finance

The comparison above is one slice. Explore the complete AI in Finance hub for the pillar view — workflows, vendor archetypes, governance, and the 12-month sequence — then book a working session on your own numbers.

Book a Demo

Tell us a little about you and we'll arrange a demo on your own finance data.

We use these details only to reply to your request.

Stay ahead in AI-powered finance

Get insights on AFI capabilities, FP&A automation, and enterprise AI delivered to your inbox.