Conversational Governance

Rules the workflow enforces — the AI can't skip them.

Governance sits inside the workflow, not bolted on after the fact. The AI drives the conversation. At approval steps it pauses and waits for your yes. Every step is logged.

Conversational Governance

What this pillar covers

  • 1

    Conversation steps. Approval steps. The workflow decides which is which.

    Conversation step — the AI drives, you respond naturally as the work progresses.

  • 2

    The AI cannot advance past an approval step without you.

    Rules live in the workflow — not in a system prompt the AI can reason around.

  • 3

    Seven kinds of checkpoints. All enforced by the engine.

    Requires human approval — the conversation stops until a human says go.

7
Gate types enforced structurally
100%
Audit trail coverage
0
Manual compliance docs needed
Two kinds of steps

Conversation steps. Approval steps. The workflow decides which is which.

Every step in a workflow is either a conversation step or an approval step. The AI knows the difference because the workflow defines it — not because you prompt for it.

  • Conversation step — the AI drives, you respond naturally as the work progresses.
  • Approval step — the AI pauses and waits for your explicit yes before advancing.
  • The workflow engine enforces the distinction. There is no prompt trick that skips an approval.
Structural enforcement

The AI cannot advance past an approval step without you.

Approval steps are not reminders or suggestions. They are halts. The conversation stops at the step, the workflow waits, and the AI has no path forward until a human says go.

  • Rules live in the workflow — not in a system prompt the AI can reason around.
  • Checkpoints are structural. The conversation cannot route around them.
  • You decide where approvals go. Full autonomy between them.
Seven gate types

Seven kinds of checkpoints. All enforced by the engine.

Every gate type the workflow supports is structural. The AI does not choose whether to honor a gate — the engine does. These are the seven, each doing one job well.

  • Requires human approval — the conversation stops until a human says go.
  • Child tasks must finish first — a parent step waits for every sub-task.
  • All phases complete — a plan cannot close until every phase is verified.
  • Task must be the right status — steps only run at the status the workflow expects.
  • Task must be the right type — certain steps only apply to certain kinds of work.
  • Wait for dependencies — a step waits until its listed dependencies are done.
  • Sequential child order — children execute in the order the workflow defines, never in parallel.
Audit trail by design

Every step. Every change. Every decision. Logged automatically.

The audit trail is not a feature you turn on. It is the shape of the platform. By the time a compliance team asks for the record, it is already there — who did what, when, and the before-and-after values.

  • Field-level diffs on every change — what was, what is, who changed it.
  • Every AI action logged with the AI that performed it.
  • "Did we follow our own process?" becomes a dashboard query, not a quarterly fire drill.
Audit Trail
Employee Onboarding #1420 events
Audit trail loading...
Gate catalogue

Pick the gate that matches the work.

Grid of the seven gate types — rendered as compact cards. Each gate is a primitive the workflow engine understands natively.

  • Human approval — explicit yes required
  • Child completion — parent waits for children
  • Phase completion — plan waits for phases
  • Status precondition — step requires a specific status
  • Type precondition — step requires a specific task type
  • Dependency wait — step blocks until dependencies close
  • Sequential order — children execute in declared order
See it in motion

See governance in action

Watch a sales proposal move forward autonomously, then pause at the approval step the workflow defines — the AI cannot advance until a human says go.

1
2
3
4
5
Claude
Watch the workflow in action
0 / 5

Gate-enforced workflow vs prompt-based hopes

Rules in a prompt

  • The AI can reason around the rule, or simply forget it.
  • "Require human approval" means hoping the AI pauses.
  • Audit happens after the fact — maybe.
  • "Did we follow our own process?" is a quarterly fire drill.

Rules in the workflow

  • Gates are structural — the engine enforces them, not the AI.
  • Approval steps are halts. No conversation advances without a human yes.
  • Every step, every change, every decision logged as the work happens.
  • "Did we follow our own process?" is a dashboard query.
Questions

Frequently asked questions

Monitoring watches conversations fail. Conversational Governance prevents failure. The rules sit in the workflow — the AI cannot argue with them, cannot skip them, cannot reason its way around them. This is structure, not observation.

No. Approval steps are enforced by the workflow engine, not the AI. There is no prompt trick that advances past a gate. The conversation halts at the step and does not move until a human says go.

Yes. Audit trails are built by design, not after the fact. Every structured conversation produces the record regulators need — who did what, when, the before-and-after values, and the checkpoints a human approved. Compliance is a byproduct of how the platform works.

Yes. You author the workflow. You place approval gates wherever the work demands human judgment — legal review, budget sign-off, customer-facing copy — and leave the rest to the AI. Autonomy between gates. Humans where it matters.

It is already there. Every change on every task is logged with who, what, when, and the before-and-after values. You do not build an audit trail in ConvOps. The system builds it as the work happens.

Only where they should. Between gates, the AI drives full speed. At a gate, the conversation waits for a human yes. You decide the placement — every approval is a choice, not a default.
ConvOps

Rules the workflow enforces — the AI can't skip them.

Governance sits inside the workflow, not bolted on after the fact. The AI drives the conversation. At approval steps it pauses and waits for your yes. Every step is logged.