AOS8Book a demo
Governance · human in the loop, where it matters

Govern decisions, not just access.

Controlling what an AI agent can reach stops it from touching the wrong system. It doesn't stop it from approving the wrong payment. AOS decides who must approve each decision your agents prepare, by role, amount and risk, and keeps a record you can check without us.

Access vs decisions

Two kinds of control. Most agent platforms stop at the first.

Access controls are necessary, and AOS has them too. Decision controls are what let a finance or operations leader hand real work to AI agents.

QuestionAccess controlDecision control in AOS
Who is this agent?An identity per agentAn agent in a declared seat, with a level and a report card
What can it reach?Systems, data and tools it may callThe commands its seat lists, under a grant with a ceiling and an end date, plus the data classes your policy lets any AI provider see
Who approves its work?Not coveredComputed for each decision by role, amount and risk. Never the agent that did the work, and never an agent that has not earned its vote on that workflow
Can the same person prepare and approve?Not coveredYour choice, by risk class, for the company, a platform or one product
Can you try it before it counts?Not coveredA preview of the version itself, on a twin of your work, with nothing leaving it
What if it goes wrong?A switch to stop the agentA rehearsed way back for every change, any agent's task stopped on the spot, and one click that pauses the improvement loop while your workflows keep running
Can you prove what happened?Activity logsA tamper-evident record of who decided what, which you can verify outside AOS
Every decision, explained

A decision arrives as a brief, not a bare button.

The same brief serves every kind of decision: a payment, a new rule, a promotion for an agent, an install, a change to the workflow itself.

  1. The question and the options, in plain words.
  2. Why it came to you: the rule and the number behind it.
  3. Who may decide, and who may not, with the reason.
  4. What the agents proposed, with how sure they are and how often they've been right.
  5. Whether it was previewed, by whom, what replay found and how it differs from Live.
  6. What it reaches: the products it touches, at the highest risk class among them.
  7. What happens if it waits: the deadline and who it goes to next.
  8. The way back, or a plain statement that there isn't one.
  9. Approve in two steps, with a fresh sign-in for high-risk decisions.
Pay carrier invoice INV-4471High risk
Question
Pay $12,400.00 to Red Rock Freight for load LD-8851?
Why you
$12,400 is above the billing team's limit of $10,000.
Who decides
The controller. An operations manager may take it, except the one who recorded this match.
Proposed
Pallas matched it to LD-8851 · right 47 of 50 at this level
If it waits
goes to the operations manager at 3:30 pm
Way back
None once the payment leaves. You'll sign in again before it does.
Example brief. Names and figures are illustrative.
Your policy

You set the rules. AOS applies them to every decision.

Approval is configuration. Each company sets its own policy, for the whole company, for a platform or for one product. The engine works out who may decide; no screen and no agent can work around it.

The default

One person, or two different agents

A decision takes one person, or two different agents voting the same way from the decision's verifying seat, on work neither had a part in. Out of the box, a person may approve their own work.

Default: one person · or two agents, below the floor
Separation of duties

A different person when it matters

Ask for a different person by risk class. The setting is off by default, and you turn it on where you want it: for the whole company, or just for one product.

Freight Ops, high risk: a different person approves
The human floor

From which risk a person must decide

Below the floor, two different agents may agree on routine items. At or above it, only a person can. The floor is high risk by default; set it lower and people decide more.

Default: high risk and above, a person decides
Limits and cover

How much, and when someone is away

Each role has its limits, and anything above goes to the next person who may decide it. Give a colleague a role until a date. Nothing waits silently: an item nobody can decide shows every manager how to fix it.

Billing clerk up to $10,000 · J. Ortiz covers the controller until Friday
The verifying seat

A verifier knows the work

Every decision names a verifying seat. An agent's vote counts only from that seat, and only once it has earned it on that workflow, by voting in preview and matching what your people decided. You place an agent in a seat, or take it out, with a sentence that a person signs.

Default bar: 100 preview votes, 9 in 10 matching · never under 30 at 85%
Two real opinions

Two votes must differ in substance

Two agents decide only when they share no playbook and, unless your policy says otherwise, no AI provider. With no such pair, a person decides.

Same playbook or same provider: counts as one · a person decides

A setting can tighten. It can never loosen below the promises. Four rules hold whatever a company sets: an agent never approves its own work, an agent's vote counts only where it has earned it, the promises are not weaker in any company's hands, and a company's own layer on a product may tighten a guard and never loosen one. Anything that pays, refunds, signs or certifies is a person's decision.

LowStanding policy may decide. Routine work inside limits a person already set.
MediumGoes to a decision. One person, or two different agents with a proven record, below your floor.
HighA person decides, after a preview someone has walked. Payments, refunds, signatures and anything at or above your floor. The walked preview is a setting that starts at high risk.
Preview and Live

Try it before it counts. What you preview is what you sign.

Every product has two states. Preview is a version shown for a person to try before anyone decides. Live is the version pinned and running for your company. Governed is the path between them: the checks, the review and a decision on the record.

Anything that changes what runs

A product version, an install or an upgrade, your own layer, a lesson in an agent's playbook, a rise in a level or a dial, a change of the AI provider behind an agent.

Fed three ways

A person walks it as any role. Replay shows what it would have done on your past cases. Alongside runs it on today's work with no effect.

Nothing leaves a preview

It runs on a twin, with a budget and an end date. Its opening, walking and closing go on your record, and the decision brief says what it found.

Earned autonomy

An agent's trust is earned, measured and approved.

Every agent has a report card: how often people agreed with it, how often they changed its work, and how well its confidence matched reality. You can label its work before you see its answer, so the numbers stay honest.

When the record supports it, AOS proposes a promotion with a watched trial and a way back. A person decides. An agent never votes on its own promotion, and its vote counts only on work it had no part in. Until it has earned its vote on a workflow it casts a preview vote, which decides nothing and is compared with the person's decision.

How far an agent may act in a product is the lowest of four things: the level it has earned, the ceiling on its grant, the dial on the product, and the dial on the workflow where you set that lower. All four sit under AOS8's own limit.

GovernGoverned autonomy: runs its part of a workflow end to end, inside every rule on this page.
ApplyApplies its own routine work within your limits, with every step on the record.
AOS8's limit · rises only through approval
WatchObserves how your team does the work.
SuggestRecommends the next step; a person does it.
PrepareDrafts the work ready for a person to approve.
CheckMay agree with a second agent on routine items below your floor.
PersonAlways a person: payments, refunds, signatures, and anything above your floor.
The record

A record you can check without trusting us.

Everything that happens is a command, and every command leaves an event on your company's own record, written in the same step as the change itself. The record is sealed in batches, and each seal is witnessed by a service outside AOS, so an edit after the fact would show.

What verification proves

  • Nothing was changed or reordered afterwards
  • What was approved, and by whom
  • Who decided, and under which rule
  • When it happened, from the outside witness

What it doesn't prove on its own

  • That a rule was worked out correctly
  • That an AI model saw exactly what the record says

AOS covers those separately: it replays recorded cases in its tests, and keeps the exact inputs of every agent turn for your auditors.

Built-in protections

The boundaries no setting can remove.

Outside text stays with people

An agent's vote never counts on anything that holds text from outside: a partner's message, an emailed or uploaded document. A person decides those. Everything an agent reads is fenced and labelled, so instructions hidden in a document stay data.

Your company, sealed off

Each company has its own record, its own people and its own policy. Inside it, people see only what their roles may read. No platform, product or publisher is a way through that wall: a publisher sees its installers as counts, never as records.

Vendor access you can see

Flowwolf and AOS8 staff can't read your records. Emergency access is a grant with a reason, a start and an end. It is written to your own record, and you see every read.

Outside AI clients are seated

An outside AI client reaches a product only as tools, seated like an agent. Every call is a command under a grant, and nothing is open until you seat one.

Sensitive reads are recorded

A read of sensitive data is recorded, for people and agents alike. What an outside client reads is bounded by data class and by your rules for what may leave.

Spend you control

Every agent turn is metered in tokens and money. Budgets stop new turns when they're spent, and the screen says so and when they resume.

A seat for your auditor

A read-only role for the record, the evidence behind every agent turn, and the definition you signed, or an export they can verify on their own.

Nothing runs unsigned

Products run only as versions that were signed. Code that moves money or reaches outside is a signed capability, and only AOS8 signs those.

What a publisher signed stays signed

You change a product you installed by adding your own layer, never by editing the publisher's version. Every app, step and field shows whose it is.

Your data

Your policy decides what any AI provider sees.

Classed before it leaves

Every input to an agent is classed. Nothing goes to an AI provider outside your data policy, and anything that looks personal is treated as personal unless a person says otherwise.

Keys never shown

Provider keys are sealed and write-only. Each company's data sits under its own encryption key.

Erase, hold, export

Erase a person's data without breaking the record, place a legal hold, or export everything and verify it offline.

Where your data lives

Your company lives in one cell, chosen when it is created. One page, Where your data lives, names how the cell runs and any cloud vendor in the data path.

Always on

What runs today: processes that always run, on a database the cell runs itself. Each company's rows are walled inside the database.

On demand proposed

In design: work runs when it arrives, on a cloud vendor's services. The vendor is then in the data path, by your choice, and people are denied the tables.

Book a demoVerify the record yourself