From $5,000

Your logs record an action and an actor. When more than one agent acts, the actor is the part that is wrong.

This is for the business that has more than one thing acting under a name. An agent, a scheduled job, a service account, an integration token — and somewhere behind them a human login they borrowed. Nothing has gone visibly wrong yet. That is the condition we are looking for, because the failure this audit finds is silent by construction: when every actor presents the same identity, the record says a person did it, or says nothing useful at all, and the investigation becomes an interview. If a customer, an insurer or a lawyer ever asks who did this, the honest answer today is a name that is either wrong or unavailable. We make it provable before anyone asks.

What you get

The deliverables, stated plainly.

01

An identity inventory

Every identity that can act on your systems — humans, agents, scheduled jobs, service accounts, integration tokens — with the name it presents and the name it actually is, side by side.

02

An attribution table

10–20 consequential actions from the last 90 days — mail sent, money moved, records changed — each proved end to end or marked unprovable, with the reason. VERIFIED, AMBIGUOUS or MISATTRIBUTED.

03

A credential-to-actor map

Which secret each actor uses. Where one credential is held by more than one consumer, attribution is impossible by construction, and the map says so rather than blaming the logs.

04

A delivery proof sheet

Every class of automated message, whether it actually landed, and where. A job that cannot prove a message landed is not a job; it is a script that ran.

05

A non-repudiation register

The write-path changes that make the next action attributable, in priority order: distinct credentials per actor, an actor ID in the log line, delivery receipts, model and agent version recorded with the action.

06

A one-page scorecard

The percentage of sampled actions you can currently prove. One number, one page, dated — so the after number means something.

How it runs

Days one and two are the inventory: we get read access within 5 days of signing and list every actor, including the ones nobody owns. Days two to four are the attribution trace on real actions from your history. Days three to five are the credential-to-actor map, which is usually where the answer to "why can we not prove this" is found. Days four to six are delivery proof. Days six to eight are the register — the specific changes, with effort attached. Days nine and ten are the scorecard and the readout. You see every draft within 3 days of asking, and we hold a weekly check-in. Remediation is scoped separately; we do not bundle the fix into the audit, because that is how an audit becomes a sales pitch.

What it costs

Agent Identity Auditfrom $5,000 (scoped $5,000–12,000)
Attribution monitoring attach$450–750 per month
Remediation buildscoped from the register, priced separately

Payment is 65% at signing and 35% on acceptance. The founder confirms the final number on the call. If we cannot show you at least one consequential action you cannot currently attribute, you do not pay — not a credit, no invoice.

Why now

Through 2026 the failure mode moved from what was being measured to who was speaking. Our own record is unambiguous and recent: nine consecutive weekly reports were attributed to the wrong agent; a scheduled job ran as a no-op under an identity nobody owned for ten weeks; and a routine job spent six weeks asking a question into a channel that had been deleted, so the audit trail correctly recorded that a question was asked and no human was ever asked. None of those were breaches. All of them were the same defect — an action with no reliable actor. The market has arrived at the shape of this without naming it: cyber-insurance renewals now ask you to inventory every agent with a named owner, agent-action disputes are being litigated, and in October 2026 a major model provider notified more than 100 organizations that agents had acted on their accounts without authorization. The first question in a renewal conversation is no longer whether it was patched. It is whether you can prove who did what.

Questions people ask

We have logs. Is that not attribution?

Logs record what a process claimed when it wrote the line. If two consumers hold the same credential, the log cannot tell them apart, and no amount of retention fixes that. Attribution is a property of the identity design, not of the logging system.

Is this not just audit logging and SSO?

Those give you a session and a human. This is about the actor that is not a human and has borrowed a human's name. That gap is exactly where agents live, and it is the part neither tool was designed for.

We only run one agent. Is this overkill?

Probably, and we will say so on the first call. One actor with one owner is trivially attributable. The audit is priced for the business where several things act and nobody can name them.

What if everything turns out to be attributable?

Then you have a dated one-page scorecard saying so, which is worth having at renewal time, and the register is the remaining value. You still pay — you bought the answer, not a scare. Our guarantee covers the other case.

Do you need production access?

Read access to logs, deployment configuration and the credential store's metadata. We do not need secret values. If we find one in a place it should not be, that becomes a finding in the report rather than something we use.

Related

Book 15 minutes

The founder confirms the number on the call. Fifteen minutes, no deck, no pitch — we talk about your situation and tell you if there is nothing to build.

Before you buy the audit · free, no email

Run the twelve questions yourself first

The audit exists because most teams cannot answer a simpler question: which identity actually acted, and can you prove it? Twelve questions, three possible answers each (yes, no, I would have to check). The third answer is the finding. No score, no email, printable as a PDF.

Take the free checklist