The second signature for automated work

Put AI to work without losing quality control.

AI can now refund customers, send emails and change records faster than anyone can check. Cosignary lets your team spot-check that work the way a good QA department would: it stops risky actions before they happen, picks a fair random sample for people to review, and gives you proof the work was right. No AI grading AI.

The demo runs in your browser with synthetic data: audit an AI support agent, then try to cheat the audit.

AI proposesIt can explain, summarize and flag. It's never allowed to clear anything.
Code decidesSmall, readable, deterministic rules that anyone can inspect and rerun.
People signAnything irreversible needs a human signature, at a pace people can keep up with.
The problem

Software now acts on its own. Oversight hasn't caught up.

AI agents issue refunds, change orders, send emails and ship code. Two things make watching them hard.

Problem 1

AI watching AI

Today's safeguards increasingly use one model to approve another model's actions. The watcher can be fooled, can drift, and shares the failure modes of what it watches. A note in the agent's own output can talk an AI reviewer into approving it.

Problem 2

More than anyone can read

Agents produce more actions than any team can review, and the gap widens with every model generation. "A human checks everything" was never true. Sampling at random is the only honest answer, and only if nobody can rig the sample.

How it works

One trust layer for everything that acts on its own.

Cosignary sits beside your AI agents, scripts and automations. It gates what they're about to do, records what they did, and turns an endless stream of actions into a few questions a person can answer.

Your systems

Support agent
Coding agent
Automations
Staff tools

Deterministic gate

✕ blocked by policy refund.issue $400 · order total $40

Plain-JSON rules block, hold for a signature, or allow each action before it runs. Speed limits trip a circuit breaker. No model in the decision.

Verifiable random audit

HMAC(seed, 3f9a…)0.412
HMAC(seed, 81c2…)0.031 ✓
HMAC(seed, c7e0…)0.778

Every action goes on a hash-chained ledger. A seed locked in beforehand picks the sample, and anyone can recompute it afterward.

Plain-language review

Store credit: $250 for a dented box Policy: goodwill credits up to $25.
YesNoNot sure

Your own people answer one short question per card. Planted test cards measure whether they're really looking.

Works on AI agents, ordinary scripts and human workflows alike: an action is just a JSON record.

What a sample buys you

Human review that scales with confidence, not volume.

Nobody can read everything an AI produces. A random sample nobody can game, checked carefully, says something precise about all of it.

Clean checks that prove a category is below 1%
299
Review per million actions, five categories
~58h/mo
AI models in the decision path
0
Tamper attempts caught in the demo
5/ 5

The 299 holds at 95% confidence whether a category has ten thousand actions or ten million, as long as none of the checks finds a problem. Review time plans on 350 checks per category a month, so a few problems can turn up and every category still proves under 1% each quarter, at a careful two minutes per card. The demo's five attempts: editing a past action, rebuilding the chain, deleting an action, cherry-picking the sample and swapping the seed.

Real-time controls

Stop it before it runs. Sign off from your phone.

Every action passes the gate first (human-in-the-loop, designed for agent volume). Most run straight away. The risky ones wait for a person's signature, and a burst of small ones trips a circuit breaker, because per-action caps can be beaten by volume.

An agent actionrefund · email · deploy · delete
allow
Within policyRuns immediately and goes on the ledger. The gate adds milliseconds, not minutes.
hold
Needs a signatureA signer's phone buzzes with the card. Approve or deny, signed with their passkey. No answer by the deadline means deny.
block
Over a rule, or the breaker is trippedThe action never runs. A tripped breaker pauses the whole category until a person resets it.

Speed limits

"More than 20 refunds or $2,000 in 10 minutes." Counted across actions, so 300 small refunds can't slip past a per-refund cap.

Signed approvals

The ledger proves who approved, not just that someone clicked. Flip a signed "deny" to "approve" and verification fails.

No answer is never a yes

Held actions expire to deny. Irreversible actions run only at a pace people can actually sign for.

Alerts worth waking for

Now: approvals, tripped breakers, tamper. Daily digest: categories above target, missed planted tests. Webhooks reach Slack, Teams, SMS and pagers.

One audit period

Fair before it starts. Checkable after it ends.

The election-audit pattern (a seed committed in advance, a selection anyone can recompute) applied to automated work.

  1. Auditor

    Commit

    Publish only the hash of a secret seed, before anything happens.

  2. System

    Record

    Every action is chained to the one before it, then checkpointed outside.

  3. Auditor

    Select

    Check an action when HMAC(seed, fingerprint) falls under its rate.

  4. Watchers

    Review

    Yes, No or Not sure on each card, with planted tests mixed in.

  5. Auditor

    Reveal

    Publish the seed and the planted-card key.

  6. Anyone

    Verify

    Recompute the sample. Any edit, deletion or cherry-pick fails.

Rates adjust by a published rule. If a category's problem rate comes in above its target, its audit rate doubles next period, up to 100%. A clean period halves it back toward the base rate. Anyone can recompute that from the ledger too.

Where it runs

Your data stays with you. Only fingerprints leave.

The cloud part of Cosignary is a witness, not a judge. It holds no customer data, no seeds and no decision power, so breaking into it can't rig an audit.

Your servers or cloud

everything that touches data
Gate
Ledger
Sampler
Review page
Report
Redaction
Customer data, seeds and verdicts never leave this box
fingerprints onlyhashes · counts
signed receiptskept by you

Outside witnesses

several, independent
Cosignary witnesssigned checkpoints
Bitcoinvia OpenTimestamps
Public timestamp authorityRFC 3161

Nothing worth stealing

A witness stores hashes and counts. An attacker who breaks in learns nothing about your business.

Tampering shows

Each witness signs its own chained log, and you keep the receipts. Backdating or forging a checkpoint breaks a signature.

No single point of failure

The worst a breached witness can do is go silent, which shows up as missing receipts. Anchor with several.

Read the full security model and threat model

Assurance levels

Start on your own. Bring in an outsider when someone asks.

Every report states its level plainly, so a self-reviewed period never reads as an independent one.

Level 1

Self-reviewed

Your own people review the sample. The selection is still fair, but only you vouch for it.

For small businesses and internal assurance. Free and open source.

Level 2

Self-reviewed, verifiable

The seed is revealed and the public checks pass. Anyone can confirm the sampling was fair without seeing your data.

For showing customers and partners the process is honest.

Level 3

Independently re-reviewed

An outside reviewer re-judges a seed-picked share of cards, redacted and in your environment. The stricter verdict counts, and agreement is reported.

For insurers, enterprise buyers, regulators and auditors.

Markets

Wherever automation touches money, customers or production.

The same protocol fits every industry. What changes is the policy, the checklist on each card, and who the watchers are: in each case, the people who used to do the work.

Customer support

Agent actions
Refunds, store credits, order changes, replies to customers
Watchers
Experienced support reps and team leads
Why they pay
AI support vendors need proof to close enterprise deals, and service brands need to know the agent meets their standard.

Finance operations

Agent actions
Payments, invoices, journal entries, vendor changes
Watchers
Bookkeepers, AP and AR staff, controllers
Why they pay
Money comes with built-in checks (double entry, reconciliation), and auditors already accept sampling as evidence.

Insurance claims

Agent actions
Claim approvals and denials, payouts, reserve changes
Watchers
Adjusters, examiners, claims processors
Why they pay
Regulators expect governance of AI in claim decisions, and insurers can reward audited agents across every other industry.

Coding agents

Agent actions
Deploys, deletes, migrations, config changes, commits
Watchers
QA testers, junior developers, SREs
Why they pay
One wrong command can take down production. Teams want a gate the agent can't argue with, and a record that proves it held.

Legal operations

Agent actions
Document review, clause changes, filings
Watchers
Paralegals and document reviewers
Why they pay
Courts have accepted statistically sampled document review for over a decade. The method is already familiar.

Lending and KYC

Agent actions
Approvals, declines, fraud and AML alert decisions
Watchers
Loan processors, KYC and compliance analysts
Why they pay
Fair-lending rules mean every automated decision has to be explainable on demand.
Also:Healthcare billingGovernment servicesConnected-device fleetsSpecialty retail
CredibilityCoding agents: open-source gate and verifiable audits on real repositories.
First revenueAI customer-support vendors: a verified trust report for their enterprise customers.
Big marketsFinance with an audit-firm partner; insurance claims with an insurer.
CredentialA certified-watcher credential, once real catch-rate data exists.
Who pays, and what it costs

Trust costs hours, not headcount.

Nobody buys trust in the abstract. They pay when its absence is costing them a deal, a discount or a compliance finding. And because review is sampled, the cost stays small.

AI vendors selling to enterprisesBuyers won't let an agent act on their systems without evidence. A trust report shortens the sale.Precedent: SOC 2. Vendors pay for the audit so they can sell.
Companies that want more autonomyAn agent stuck in "suggest only" saves little. Evidence lets it act, which is where the savings are.Oversight priced as a small share of what it unlocks.
Regulated buyersLogging and human-oversight requirements are already budget lines.EU AI Act, ISO 42001, SOC 2.
InsurersAI liability cover needs evidence of risk. Lower premiums fund the oversight.Like cyber insurers requiring MFA.

Review time calculator

Hours a month to prove every category is under a target problem rate, at 95% confidence.

1M
5
1%
120
58.3 ha month: 1,750 cards to prove every category is under 1% each quarter, allowing up to five problems found per category. Move the volume slider; above a few thousand actions, the hours don't change.
Free and open sourceThe gate, ledger, sampler and verifier. Self-review mode for anyone.
Paid when someone else needs proofVerified period reports, outside re-review, planted-test libraries, watcher certification.
Priced per agent, per periodNot per action, because cost tracks confidence, not volume.
Watchers

You don't need to understand AI to oversee it. Just your own job.

Watcher (n.): the person accountable for confirming that automated work in their field is correct, judged against recorded evidence, and for signing off on anything that can't be undone.

A fitter checks fittings, a bookkeeper checks invoices, a support lead checks refunds. The credential is a live, measured catch rate on planted tests, renewed from evidence rather than earned once in an exam, so pay can follow accuracy instead of speed.

Careers

The AI-proof job

What jobs AI will replace and what's left, the six-level watcher career ladder, roles by industry, and how watchers are measured.

Becoming a watcher

FAQ

Questions worth asking

More in the human-in-the-loop guide, the watcher guide and the security model.

What is Cosignary?

Cosignary is an oversight layer for AI agents and any software that takes actions on its own. Deterministic rules decide whether each action runs, a tamper-evident ledger records everything, verifiable random audits choose which actions people check, and planted test cards measure whether the checkers are paying attention. No AI model makes the decision.

How is Cosignary different from AI agent monitoring tools?

Most agent monitors use an AI model to judge whether another model's action is safe. Cosignary keeps AI out of the decision: rules and code decide, people sign, and the evidence (ledger, sample, signatures) can be verified by anyone. An AI may flag an action for extra review, but it can never clear one.

Isn't this just AI watching AI?

No. Selection, enforcement and verification are deterministic code, and people give the verdicts. AI can help, but only by adding scrutiny: it may flag an action for review, never clear one. An AI flagger must come from a different vendor than the agent it watches, and it's scored on planted tests like any other reviewer.

Why can't we just build this ourselves?

You can build the checker. You can't build someone else's signature on the result. A self-audit isn't independent; a homemade method has to be defended to every auditor and insurer; and benchmarks, planted-test libraries and certified reviewers only exist across many organizations. The core is open source on purpose, so building your own gains you nothing.

Can it stop an agent in real time?

Yes. Every action passes the gate before it runs: allow, hold for a signature, or block. Speed limits trip a circuit breaker that pauses a whole category, such as all refunds, until a named person resets it. Held actions that nobody answers are denied, never approved by default.

Does our customer data leave our systems?

No. Everything that touches data runs in your environment. Only fingerprints (hashes and counts) go to outside witnesses. Review cards are redacted, and outside reviewers, if you use them, work through your own secure access.

We're a small business. Do we need an outside reviewer?

No. Self-review is the default and a first-class mode: your own people, your own servers, no outsider. Turn on outside review the day an insurer, an enterprise customer or a regulator asks for evidence.

How much review time does it take?

Review grows with the confidence you want, not with volume. Plan on about 350 random checks per category each month. That proves the category is under 3% every month and under 1% every quarter, even when reviewers find the handful of problems a good agent produces; 299 clean checks prove under 1% in a month only if they find none. A million actions across five categories is about 1,750 cards a month. At a careful two minutes a card, that's roughly 60 hours: a part-time job for one reviewer, whatever the volume.

Does it work for software that isn't AI?

Yes. Any system that takes actions can log them: scripts, scheduled jobs, code commits, device-fleet updates, or people using internal tools.

Built by a founder who runs AI agents in production. Sean Morrissey built the guardrails behind Cosignary to keep AI agents from breaking his own live product, FitMyGolfClubs. Read the story.

See it work

Audit an AI support agent in five minutes.

Review a verifiable sample, catch the planted tests, try five ways to cheat the audit, then approve a held refund from a phone and trip the circuit breaker.