Autonomous, not reckless: Safe, Ask me, Never

A policy engine, not a black box

“Autonomous” makes people nervous, reasonably. Software that changes what your brand says without asking sounds like a risk, not a feature. VISIBLE's answer to that is a policy engine: every proposed change is sorted into one of three tiers before anything happens, and the tier — not a person's mood, not urgency, not how confident the system feels — decides what happens next.

The point of naming the tiers plainly is that a customer shouldn't have to trust a vague promise of “safety.” They should be able to see, for any single change, which of three categories it falls into and why — and know in advance what each category is and isn't allowed to do.

Three modes, one default

The tiers decide what kind of change is allowed through; the operating mode decides how much is allowed through at all. There are three: Autopilot, where safe, high-confidence changes execute on their own; Approval Mode, where every change is prepared and shown to you first; and Monitor Only, where VISIBLE measures and diagnoses but publishes nothing. Approval Mode is the default for every new customer unless they explicitly choose full Autopilot.

• Autopilot — safe, high-confidence changes execute automatically. For teams that want maximum automation.
• Approval Mode — every change is prepared and shown to you before it goes anywhere. The default.
• Monitor Only — measurement and diagnosis, with no changes published at all. For regulated or tightly controlled organisations.

Starting everyone in Approval Mode is a deliberate choice, not a default that happened to fall out of the engineering. A customer who has never watched the system work has no basis yet for trusting it with full automation, and being asked to trust it on day one — before seeing a single approved change — is a worse experience than being shown the evidence first and choosing to loosen the leash once it's earned.

What approval actually looks like

Every proposed action carries the same four things: the reason it was proposed, the evidence behind it, a preview of the exact change, and an approve-or-reject decision — with a roll-back option where the connected system supports it. Take the same worked example VISIBLE shows on its own product screens: a fictional coffee roaster with four items queued at once.

In that worked example, correcting a stale shipping claim is Safe and already done; a page comparing the brand to its biggest rival is Ask me and sitting in the queue; a claim the brand couldn't back up — being “Portland's number one roaster” — is Never, and stays refused regardless of how it's asked for. That last case is the one that actually earns the trust: a system that will happily publish anything you approve isn't safe, it's just obedient.

We will not publish a claim we cannot back up — even if you ask us to.

VISIBLE

Rollback and integrations

None of this works without a real connection to the systems it's changing — a supported CMS, a search console, an analytics account. Every connection is permission-based: you decide what VISIBLE can measure, prepare or execute, and a capability is never claimed unless it's actually supported in production. Where the connected system allows it, an approved change can also be rolled back — the same one-click control in reverse.

A record of everything, not just the pending queue

The three tiers only mean something if there's a record to check them against. Every autonomous action leaves an audit trail — what was proposed, what tier it fell into, who approved or rejected it and when, and what it was supposed to measure afterward. Workspace access is role-based, so who can approve a change is a deliberate setting, not an accident of who happens to be logged in. Credentials for connected systems are stored encrypted, and workspace data stays isolated from any other customer's.

None of this is presented as a certification, because none of it is one. VISIBLE does not claim SOC 2, ISO 27001, GDPR certification, a penetration-test status, or an uptime SLA — those are specific, auditable claims, and making one without having actually obtained it would undercut the entire point of a product built on not overstating what it can prove. What's listed are the controls that exist today: permission-based integrations, an audit trail, approval history, and export and deletion controls, so a customer's brand data stays theirs to take with them.

Thirty days, then an active choice

The Intelligence Audit includes 30 days of Autopilot Early Access, starting once your baseline is measured and you activate it — not on the day you pay. At launch, that period doesn't renew automatically: at day 30 you're asked to actively continue at $99 a month, with the renewal date and price shown plainly at checkout rather than buried in it.

Across those 30 days, the product is meant to earn the continuation rather than assume it. The early days surface the first opportunities and approvals; the middle stretch shows what competitors and citations are doing while you watch; and the final week builds toward a single, honest summary rather than a marketing moment.

If nothing measurably improved during a period, the monthly outcome report says so — it lists what was learned, what's still being tested and what's queued next, not invented progress.

Verification is just asking again

Every change is written down as an experiment before anything is touched: what it was meant to move, and the baseline beforehand. Then the same Money Prompts get asked again, and the result is reported honestly as one of four states — improved, inconclusive, no movement, or regressed. Two of those four are bad news, and they're shown with the same weight as the good one, because a tool that only ever reports success isn't measuring anything.

That's the whole safety model: nothing changes how your brand sounds without your say-so, nothing unproven gets published no matter who asks, and every claim of improvement is checked by asking the exact same question again.

It's a slower, more accountable version of “autonomous” than the word usually implies — closer to a very diligent junior colleague who shows their working every time than to a system quietly rewriting your brand overnight. That's a deliberate trade: less dramatic than full, unsupervised automation, and considerably easier to actually trust. See how the score is built for what those repeated questions are actually measuring, or read the full method end to end.