Free Audits · running one more than once

Run an Audit on a Schedule

Every audit says how often it is worth running again. This is the part that comes after: what to change about the run so the second one is worth reading, and the one rule we would not break if you wired it to an agent.

The idea in one line: the first run is the one that surprises you. Every run after it is checking whether the fixes moved the measures the first one set, which is a different job and a shorter one.

Why the second run is a different job

Each of the seven audits diagnoses from a standing start. That is correct the first time: nobody has seen a brief on this before, and the value is the gap between what you believed and what your records show. Run the same thing again next week unchanged and it will find the same gap, describe it with the same confidence, and tell you the same thing to do about it.

The Acquisition Payback Audit will report the same miscoded referrals every Monday until somebody fixes the coding. By week three, people stop opening it. What makes a repeat run worth reading is change: what moved, what is new, what was fixed and stayed fixed, and what was supposed to move and did not.

The rule we would not break

It reports. It never changes anything. No cancelled seats, no paused campaigns, no excluded locations, no edited records.

This is the part most people skip, because full automation demos better. We think it is the wrong trade at this stage, for a reason that shows up in the audits themselves: the Technology Waste Audit, run on a 24-person firm, found 22 seats on one tool it could not evaluate, because the export carried no login data. It was the biggest line in the account. The right answer was to say so and stop, and an agent with permission to cancel would have been one confident inference away from cancelling seats people were using.

The same audit found two tools on personal credit cards and recommended sanctioning them rather than shutting them off, because the people using them had a reason. No rule you can write catches that. A person reading a brief catches it in a second.

So the useful shape is an agent that reads, reconciles, and writes the brief, and a person who decides. If you later want the agent to prepare the change as well, the step to add is an action someone has to approve before it runs, and the approval is the product.

What you need before you start

The same export, on a cycle

Whatever the audit asked for the first time, pulled the same way each run. A schema change upstream is the thing most likely to quietly corrupt a trend.

Last run’s brief

Kept somewhere the next run can read. Section 7 of every brief names its measures and stamps each baseline with the date it was pulled, which is what makes them line up.

Your standing answers

The questions the audit stopped to ask you, written down once. Why two tools both exist, which work happens off-system, what counts as qualified.

A person who reads it

Named, not assumed. An unattended brief nobody opens is worse than no brief, because it looks like coverage.

Three blocks to add to the prompt

Keep the audit’s own stages exactly as they are. Append these.

1. Standing answers — so it stops re-interviewing you

Why: some stages stop and ask you something no export contains. On a schedule there is nobody there to answer, and skipping the question removes the thing that made the finding safe.

Paste: “Below are my answers to questions these stages would otherwise stop and ask. Treat each as if I had just given it to you. Use them, and do not ask again unless this run’s evidence contradicts one. If it does, say which answer, what contradicts it, and what you did instead of assuming.”

Give each answer a date. An answer about who uses which tool is worth trusting in March and worth re-checking by November.

2. What changed — so the headline is a diff

Why: without last week’s brief, every run re-argues findings nobody has had time to fix yet.

Paste: “Here is the brief from [date]. Your job this time is what changed. The headline becomes: what moved, each measure with its old value, new value, and whether it moved toward the target or away; what is new, sized the same way; what was fixed and held; and what should have moved and did not. Do not re-argue a finding that has not changed. List it in one line under ‘Still open since [date]’ and move on.”

That last sentence is what keeps a weekly brief short, and it gives you the list worth watching: what has been open the longest.

3. A quiet week — so it can tell you nothing happened

Why: a prompt written to produce a finding will produce one. Run that weekly and it will promote noise to significance, and people will learn to ignore it.

Paste: “If no measure moved beyond [threshold], no new finding clears [size], and nothing that was fixed has regressed, the correct output is four lines: the period, the measures you checked and that they held, anything still open since the last brief, and nothing else. Do not promote a small movement to a finding in order to have something to report. A quiet week is a result.”

Set the threshold from your own numbers rather than a rule of thumb. A useful default: a measure has to move more than it moved between the last two quiet runs.

How often to run each one

Running something more often than its data changes is the fastest way to teach someone to ignore it.

Recommended cadence, by audit

Operations Bottleneck: Weekly. A week is enough throughput to see a queue move. Any tighter and you are reading noise in the same eighteen projects.

Advertising Waste: Weekly. Spend accrues daily and the waste compounds, but the reconciliation needs a week’s conversions before the ladder means anything.

Process Waste: Weekly. Log a fortnight once, then re-read a week at a time. The patterns repeat faster than the process changes.

Pipeline Leak: Weekly. Deals move daily. A stall caught in week one is still recoverable; the same stall at day sixty usually isn’t.

Acquisition Payback: Monthly. Cohorts mature in months. A weekly CAC is mostly noise, and the 90- and 180-day windows barely move week to week.

Margin Leak: Monthly. A job has to be finished, and its last invoice has to have had time to be paid, before it can be scored. A month gives you enough finished jobs to see a distribution; a week gives you three jobs and noise.

Technology Waste: Quarterly, and before you set a budget. Seats and renewals don’t move weekly. The value is concentrated at renewal dates, so run it far enough ahead of the nearest one to act.

What gets more important, not less

Two habits the audits already have matter more once nobody is reading over the agent’s shoulder. It must never invent a number, never quote an industry benchmark from outside your data, and never treat silence in a place you told it goes unlogged as evidence of nothing happening. And it has to keep observed facts, inferences, and unknowns apart, including the willingness to say the evidence cannot settle something.

That second one has a failure mode specific to repetition. An unsettled thing can quietly become settled by being repeated every week in the same words. If a finding was marked as unproven the first time, it stays unproven until something actually tests it.

None of this needs a platform. It is the same audit, the same prompt, and three appended blocks, run on a cycle by whoever already pulls the export. The part worth paying for is the wiring and the standing answers, not the method.