The Authority Gap

Your AI agents send data, commit spend, and change records in your name. Your logs record that it happened. They almost never record who authorised it. That distinction stays invisible until a regulator, an auditor, or a customer asks who signed off. Ambit measures the gap from evidence you already have: locally, offline, in one command.

Most teams can tell you what their agents did. Almost none can tell you what their agents were authorised to do. The difference is not a logging configuration you got wrong — for most evidence formats it is structural, because the field that would name the authorising principal was never part of the record. Grading tells you which of the two you have, before someone else asks.

Authority evidenced

Actions your records can attribute to a named principal.

Actions permitted

Everything the agent was allowed to do and did.

The gap

Permitted actions your evidence cannot attribute to anyone.

The Four Questions

Every consequential action an agent takes answers to the same four questions. Grading asks them of your evidence.

  1. A ratified rule

    Is there a policy basis recorded for this action at all, identified and versioned — or did the action simply succeed?

  2. A named authority

    Does the record name the principal whose authority the agent acted under, or only the service account that executed it?

  3. A limit

    Is the action bounded — in amount, scope, or time — and is the bound recorded alongside the action rather than assumed?

  4. A record

    Is the decision reviewable the way a human approval would be, with integrity you can check rather than trust?

Every no is a gap. The useful finding is usually not that a field is empty — it is that the field does not exist in the format at all, so no amount of tuning will fill it.

How it runs

Sample grade

3 permitted actions: 0 attributable to a named principal, 2 policy-permitted only, 1 escalated without a resolving approval, 0 with no authority evidence at all.

Next: link the unresolved escalation to an approval record carrying a named approver.

Who acted fully fillable
Under whose authority structurally unfillable
What action, within what bound fully fillable
Which policy applied partially fillable
Why the decision went that way opaque
What data and resources it touched fully fillable
Where in the action lifecycle fully fillable
How far any of it can be verified fully fillable
Reconstruction completeness 78.6%
Ambit authority verdict structurally unfillable
A grade of a sample evidence file. The verdicts and figures are the tool's own; the property names are written out in full here. Nothing is sent anywhere.
What it reads
Evidence you already produce, as JSONL: OpenTelemetry GenAI spans, OpenInference, Langfuse, LangSmith, Weave, Microsoft Agent Governance Toolkit decision records, and Ambit's own receipts.
What it sends
Nothing. There is no network code in the grader, so there is nothing to configure and nothing to disable. No telemetry, and nothing in the evidence is ever executed.
What it costs
Nothing. Grading is a mode of the product, not a service anyone delivers and not a line item. It ships with the Ambit CLI rather than as a separate download.
The measure is not ours
  • Graded against an evidence-maturity measure published independently of Ambit, implemented as published
  • A yardstick we authored would be one we could be accused of shaping to flatter our own product
  • No maturity level is inferred — a level describes an evidence regime, which a static file does not reveal

Who Asks For This

Whoever answers for it when the autonomous system does something nobody authorised.

Operating mandate CEO / Founder In an AI-native company the agents act under your name. The grade is the authority story your customers' risk reviews keep asking for, measured rather than asserted.
Head of AI / CTO You and your platform leads own what the agents do in production, whether you signed off or not. Grading separates what was deliberately authorised from what merely works.
Risk ownership CISO Sooner or later you have to explain the control. The grade reports which properties your evidence supports and which it structurally cannot, one by one, with the reason.
CRO Delegation limits exist for humans. Grading shows whether your evidence can demonstrate that your autonomous systems respected them.
CDO Agents read data, change it, and release it. The grade reports what your records establish about the resources each action touched, and under whose authority.
Assurance CAE / Audit When you ask for the receipt, today the answer is a log — and a log is not a receipt. The grade is the evidence-backed statement of that difference.

If an agent's action would land on your desk, you want the answer before the question arrives.

What It Reports

  1. One sentence you can act on

    How many permitted actions are attributable to a named principal, how many are policy-permitted only, how many carry no authority evidence at all, and the single next step that would move the number.

  2. Eight evidence properties

    Each property your records could carry — who acted, under whose authority, within what bound, and how well any of it can be verified — reported as fully fillable, partially fillable, structurally unfillable, or opaque.

  3. The architectural reason

    Where a property is unfillable, why. A field the format never carried is a different problem from a field you left empty, and only one of the two is fixable by configuration.

  4. Two aggregates, never blended

    A reconstruction-completeness score and Ambit's own authority verdict, reported side by side. They frequently disagree, and the disagreement is the finding.

Grading measures the evidence a decision left behind. The decision itself — ALLOW, DENY, or ESCALATE, before the action executes — is what Ambit Authority does. You can drive that decision, and check the evidence record it produces, without installing anything.

Open The Gate