BEFORE

Interlock sits before execution, not after it

Before an AI agent acts,
Interlock decides whether it should.

Interlock intercepts consequential agent actions before execution, applies policy, and records exactly what was authorised.

Try demo

Try demo

See how it works

See how it works

What Interlock does

Three things Interlock does
before an action happens.

Interlock Overview: share of actions completed without a human, reviewer capacity, interception gap, and outcome after execution.

See what ran without a human

The share of actions that completed inside policy, what reached a person, and where a decision is still waiting.

Interlock Decision queue: a refund of $4,200 to 12 customers awaiting a human decision, with agent, policy, scope and reversibility.

Decide what policy cannot

The card a reviewer sees: agent, policy, scope, reversibility, and the deadline to answer it.

Interlock Audit log: immutable records of consequential decisions with time, agent, action, risk, decision, reviewer and rule.

Prove what was authorised

Every consequential decision, written once, with the evidence that was on screen at that moment.

The product

One system.
Every decision it touches.

Overview

What the system did without a human, what reached one, and where a decision is still waiting for an answer.

Decision queue
Audit log
Rules
Interlock Overview: share of actions completed without a human, reviewer capacity, interception gap, and outcome after execution.
Interlock Decision queue: a refund of $4,200 to 12 customers awaiting a human decision, with agent, policy, scope and reversibility.
Interlock Audit log: immutable records of consequential decisions with time, agent, action, risk, decision, reviewer and rule.
Interlock Rules: draft rules running in shadow mode with their agreement counts and promotion status.
  • Overview

    What the system did without a human, what reached one, and where a decision is still waiting for an answer.

    Interlock Overview: share of actions completed without a human, reviewer capacity, interception gap, and outcome after execution.
  • Decision queue

    The card a reviewer actually sees: agent, policy, scope and reversibility, and the four ways to answer it.

    Interlock Decision queue: a refund of $4,200 to 12 customers awaiting a human decision, with agent, policy, scope and reversibility.
  • Audit log

    Every consequential decision with the evidence available at the moment it was made. Written once, never rewritten.

    Interlock Audit log: immutable records of consequential decisions with time, agent, action, risk, decision, reviewer and rule.
  • Rules

    Draft rules run in shadow and record what they would have decided. They change nothing until someone promotes them.

    Interlock Rules: draft rules running in shadow mode with their agreement counts and promotion status.

How it works

From intent to
authorised action.

Step 1
Intercept

Intercept

The agent’s tool call is held before it reaches the outside world.

Step 2
Evaluate

Evaluate

Policy, scope, decision class and reversibility are resolved for this exact request.

Step 3
Decide and record

Decide and record

Allowed, routed to a named human, or refused — then written once, and never rewritten.

Governance

Everything a decision
needs to be answerable.

Decision class

A money reviewer cannot decide a data action. Misrouted requests are handed over, never declined silently.

Amount limits

Above the limit the request goes to the duty lead, separate from the rule threshold.

Escalation

Named, not the next shift. An escalation without a name is not an escalation.

Handover

Open requests cross shifts with their timers intact. Nothing is closed by the change of shift.

Rule ownership

Whoever promotes a rule becomes its author, not the reviewers whose decisions produced it.

Immutable records

Written once, retained seven years, never rewritten by a later decision.

One policy. Every agent.

Try demo

Try demo

Try demo

Try demo

The missing layer

Agents can decide what to do. That is not the same as being allowed to do it.

Agents can decide what to do. That is not the same as being allowed to do it.

Without Interlock

The tool call goes straight to the world

A human sees it after the money has left

No record of what was shown at the time

Irreversible actions look like any other

Failure lets the agent keep going

No one is named as answerable

Interlock

The call is held before it executes

A named human decides while it can be stopped

The card shown at that moment is recorded

Irreversible actions need stronger confirmation

Failure fails closed, with a stated reason

Every decision carries a name behind it

In the product

What the system did
and what it does not claim.

Try demo

Try demo

  • When Interlock cannot answer, agents are stopped rather than let through.

    Fail-closed

    the default when policy is unavailable

  • Interception was unavailable for seven minutes. Two requests were refused, and both refusals were recorded.

    Interception down

    failure changes the system state

  • It is not called an escalation when there is no one to escalate to. It becomes a refusal, named honestly.

    No competent reviewer

    an escalation needs a name

  • Decided elsewhere, withdrawn, or expired fail-closed. Links outlive requests by design.

    Request no longer exists

    a link is not a promise

  • Money that has left cannot be pulled back. An export or a plan change is undone inside its window.

    Irreversible is not severe

    it marks the action, not the amount

  • A draft rule records what it would have decided, and changes nothing until someone promotes it.

    Shadow mode

    a rule proves itself first

  • Execution results never return to the log, so Interlock does not claim the action succeeded.

    Unknown by design

    outcome after execution

  • When Interlock cannot answer, agents are stopped rather than let through.

    Fail-closed

    the default when policy is unavailable

  • Interception was unavailable for seven minutes. Two requests were refused, and both refusals were recorded.

    Interception down

    failure changes the system state

  • It is not called an escalation when there is no one to escalate to. It becomes a refusal, named honestly.

    No competent reviewer

    an escalation needs a name

  • Decided elsewhere, withdrawn, or expired fail-closed. Links outlive requests by design.

    Request no longer exists

    a link is not a promise

  • Money that has left cannot be pulled back. An export or a plan change is undone inside its window.

    Irreversible is not severe

    it marks the action, not the amount

  • A draft rule records what it would have decided, and changes nothing until someone promotes it.

    Shadow mode

    a rule proves itself first

  • Execution results never return to the log, so Interlock does not claim the action succeeded.

    Unknown by design

    outcome after execution

  • When Interlock cannot answer, agents are stopped rather than let through.

    Fail-closed

    the default when policy is unavailable

  • Interception was unavailable for seven minutes. Two requests were refused, and both refusals were recorded.

    Interception down

    failure changes the system state

  • It is not called an escalation when there is no one to escalate to. It becomes a refusal, named honestly.

    No competent reviewer

    an escalation needs a name

  • Decided elsewhere, withdrawn, or expired fail-closed. Links outlive requests by design.

    Request no longer exists

    a link is not a promise

  • Money that has left cannot be pulled back. An export or a plan change is undone inside its window.

    Irreversible is not severe

    it marks the action, not the amount

  • A draft rule records what it would have decided, and changes nothing until someone promotes it.

    Shadow mode

    a rule proves itself first

  • Execution results never return to the log, so Interlock does not claim the action succeeded.

    Unknown by design

    outcome after execution

0.0%

0.0%

Completed without a human

0.0%

0.0%

Completed without a human

0.0%

0.0%

Reached a human

0.0%

0.0%

Reached a human

0%

0%

Reviewer capacity

0%

0%

Reviewer capacity

What gets intercepted

Works on the actions agents already take.

Interlock sits on the tool call itself, so any action an agent can take is an action that policy can hold.

Try demo

Try demo

Try demo

Try demo

  • issue_refund

  • export_records

  • delete_records

  • update_plan

  • pii_export

  • issue_credit

  • change_owner

  • rotate_keys

  • issue_refund

  • export_records

  • delete_records

  • update_plan

  • pii_export

  • issue_credit

  • change_owner

  • rotate_keys

  • issue_refund

  • export_records

Shadow rules

A rule proves itself
before it changes anything.

refund_under_500

91

91

% agreement in shadow

Derived from 23 decisions by 4 reviewers.

Try demo

Try demo

21 of 23 decisions agreed

Runs in shadow, changes nothing

Needs a second condition first

No owner assigned yet

Not promoted

plan_change

READY TO DECIDE

100

100

% agreement in shadow

Outside the billing window. Derived from 18 decisions by 2 reviewers.

Try demo

Try demo

18 of 18 decisions agreed

Runs in shadow, changes nothing

Ready to decide

Awaiting a named owner

Not promoted

pii_export_under_2000

78

78

% agreement in shadow

Derived from 9 decisions by 3 reviewers.

Try demo

Try demo

7 of 9 decisions agreed

Runs in shadow, changes nothing

Not enough data yet

Two disagreements unexplained

Not promoted

refund_under_500

91

91

% agreement in shadow

Derived from 23 decisions by 4 reviewers.

Try demo

Try demo

21 of 23 decisions agreed

Runs in shadow, changes nothing

Needs a second condition first

No owner assigned yet

Not promoted

plan_change

READY TO DECIDE

100

100

% agreement in shadow

Outside the billing window. Derived from 18 decisions by 2 reviewers.

Try demo

Try demo

18 of 18 decisions agreed

Runs in shadow, changes nothing

Ready to decide

Awaiting a named owner

Not promoted

pii_export_under_2000

78

78

% agreement in shadow

Derived from 9 decisions by 3 reviewers.

Try demo

Try demo

7 of 9 decisions agreed

Runs in shadow, changes nothing

Not enough data yet

Two disagreements unexplained

Not promoted

refund_under_500

91

91

% agreement in shadow

Derived from 23 decisions by 4 reviewers.

Try demo

Try demo

21 of 23 decisions agreed

Runs in shadow, changes nothing

Needs a second condition first

No owner assigned yet

Not promoted

plan_change

READY TO DECIDE

100

100

% agreement in shadow

Outside the billing window. Derived from 18 decisions by 2 reviewers.

Try demo

Try demo

18 of 18 decisions agreed

Runs in shadow, changes nothing

Ready to decide

Awaiting a named owner

Not promoted

pii_export_under_2000

78

78

% agreement in shadow

Derived from 9 decisions by 3 reviewers.

Try demo

Try demo

7 of 9 decisions agreed

Runs in shadow, changes nothing

Not enough data yet

Two disagreements unexplained

Not promoted

A rule changes nothing until someone promotes it

FAQ

Questions.
Answered plainly.

What does Interlock actually intercept?

The agent's tool call, before it reaches the outside world. Interlock resolves policy, scope, decision class and reversibility for that exact request, then allows it, routes it to a named human, or refuses it.

What happens if Interlock goes down?

Agents are stopped, not let through. Every action outside policy is refused with interception unavailable as the stated reason, and the refusal is recorded.

Does Interlock know whether the action succeeded?

No. Interlock stands before execution. Execution results do not return to the log, so outcome after execution is reported as not observed rather than claimed as success.

Can a rule change production on its own?

No. A draft rule runs in shadow and records what it would have decided. It influences nothing until a named person promotes it.

Who is responsible for a decision?

The person who made it. Every decision carries a name, and an escalation without a name is not treated as an escalation.

How long is evidence kept?

Seven years. A record is written once and never rewritten. A later revocation is a second record, not an edit of the first.

What does Interlock actually intercept?

The agent's tool call, before it reaches the outside world. Interlock resolves policy, scope, decision class and reversibility for that exact request, then allows it, routes it to a named human, or refuses it.

What happens if Interlock goes down?

Agents are stopped, not let through. Every action outside policy is refused with interception unavailable as the stated reason, and the refusal is recorded.

Does Interlock know whether the action succeeded?

No. Interlock stands before execution. Execution results do not return to the log, so outcome after execution is reported as not observed rather than claimed as success.

Can a rule change production on its own?

No. A draft rule runs in shadow and records what it would have decided. It influences nothing until a named person promotes it.

Who is responsible for a decision?

The person who made it. Every decision carries a name, and an escalation without a name is not treated as an escalation.

How long is evidence kept?

Seven years. A record is written once and never rewritten. A later revocation is a second record, not an edit of the first.

Give AI agents authority without giving up control.

Interlock gives agentic systems a decision layer before consequential actions become real.

Try demo

Try demo

Create a free website with Framer, the website builder loved by startups, designers and agencies.