ai:audits
How to run an AI workflow audit at your agency.
By Loan Laux · June 2, 2026 · 5 min read
An AI workflow audit maps how work actually moves through your agency, scores each workflow for AI fit, and ends in a sequenced rollout plan with owners and checkpoints. Run internally, it takes about two weeks: one of observation, one of scoring and writing. Here's the full method, the same structure we use in client audits.
What an audit is, and is not.
The word audit suggests compliance. This is closer to a time-and-motion study with an AI lens: where do hours actually go, which of those hours are pattern work, and in what order should you change things.
It is not a tool evaluation. Tools come last, chosen to fit workflows you've already ranked. Agencies that start from a tool list end up with subscriptions in search of a problem.
The method, step by step.
01:
Choose two or three delivery paths
Not departments: paths. 'New build from lead to launch', 'retainer month', 'campaign sprint'. Follow work the way it flows, across roles.
02:
Shadow live projects for a week
Sit in the standups, watch the handoffs, read the tickets. Log where time goes and where it leaks: re-typing, reformatting, chasing context, waiting on review.
03:
Interview one person per role
Thirty minutes each, three questions: what do you re-type most, what part of the week do you dread, where does work wait on you. The dread question finds automation candidates faster than any process doc.
04:
Build the candidate list
Every recurring task you observed, one line each: who does it, how often, how long it takes, what breaks when it's wrong, what data it touches.
05:
Score each candidate
Frequency times time saved, discounted by error cost and data sensitivity, divided by implementation lift. The rubric below makes it concrete.
06:
Sequence and write the roadmap
Two workflows per quarter, each with an owner, the tool, the new process on one page, and a checkpoint date to judge whether it stuck.
The scoring rubric.
Score each candidate 1 to 5 on five axes, multiply the first two, then discount by the rest. Precision doesn't matter; separation does. You're looking for the four or five candidates that clearly beat the field.
- Frequency. Daily scores 5; quarterly scores 1. Frequency is what compounds.
- Time. Hours per occurrence, summed across everyone who touches it.
- Error tolerance. How cheap is a wrong first draft? Internal drafts score high; anything that reaches a client without review scores 1.
- Data sensitivity. Public or routine internal data scores high; credentials, health, and payment data score 1 and wait for proper controls.
- Lift, inverted. Prompt-and-template changes score 5; anything needing engineering scores lower.
What audits usually find.
Every agency believes its situation is unique. The findings mostly aren't. The usual suspects:
:
Proposal assembly
Senior hours spent copying structure from old decks and rewriting the same service descriptions.
:
Meeting re-entry
Decisions made on calls, then re-typed into tickets, summaries, and follow-up emails by hand.
:
Status reporting
PMs assembling weekly updates from information that already exists in the PM tool.
:
First-pass QA
Checklist sweeps done by expensive people, inconsistently, under deadline pressure.
:
Dev context switching
Developers reconstructing context on every return to a codebase: what was decided, where things live, why.
The mistakes that sink internal audits.
- Surveying instead of shadowing. Surveys return the org chart's view of the process. The audit's value is the gap between that and reality.
- Auditing everything. All workflows at once means finishing none. Two or three delivery paths give you more than enough to act on.
- Letting the tool decide. 'We already pay for X, where can we use it' inverts the whole exercise.
- No named owners. A roadmap without owners is a wish list. Every line needs a person and a date.
Do it yourself, or bring us in.
Everything above is doable internally if someone senior can protect two weeks of attention. The honest reasons agencies hire us instead: nobody has the two weeks, teams are more candid with an outsider, and we've seen enough agencies to benchmark yours against what similar teams changed first.
Our audit runs this method with your team and ends in the same deliverable: a sequenced roadmap you execute with the people you already have.
Frequently asked questions.
How long should an AI audit take?
Two weeks for a 5-to-50-person agency: one for observation and interviews, one for scoring and writing. Anything past three weeks is drift, not depth.
Who should run an internal AI audit?
Someone senior enough to see across roles and blunt enough to write down what they actually saw. Ops leads and founders fit. The person whose workflows score worst does not.
How often should we repeat the audit?
A light pass twice a year: re-walk one delivery path, re-score, adjust the roadmap. The full version is worth repeating when the team or the service mix changes materially.
What should the final deliverable look like?
A short written roadmap: the ranked list with scores, the top four to six workflows with owner, tool, new process, and checkpoint date, and a parking lot of what you deliberately deferred. If it doesn't fit in a few pages, it won't be used.