← Millwright Digital Framework
The AI Marketing Ops Audit

Where AI actually creates leverage — and where it's theater.

This is the framework I use in a paid engagement, published in full. Run it yourself.

The One Rule

Do not start with tools. Every failed AI rollout I've seen started with someone picking software and then hunting for a job for it to do. Start with where the hours go. The tool is the last decision, not the first.

01Inventory

Find where the hours actually go

Before evaluating a single tool, build an honest inventory of recurring marketing work. Two weeks of time tracking is ideal; a whiteboard session with the team gets you 80% of the way there.

For every recurring task, capture five things:

  1. Hours per month across everyone who touches it
  2. Who does it and what they're paid — an hour of your analytics lead is not an hour of your intern
  3. How variable it is — same steps every time, or does each instance require judgment?
  4. What it feeds — does anything downstream break if it's late or wrong?
  5. Cost of an error — a typo in a subject line versus a wrong number in a board deck

The tasks that surprise people most often: campaign QA, reconciling numbers across ad platforms, reformatting the same content for four channels, briefing agencies, and building the monthly report nobody reads.

02Scoring

Score each task on two axes

Only two things determine whether a task is a good AI candidate. Score each 1–5.

Axis A — Leverage

  • How many hours per month does it consume?
  • How repeatable is the process?
  • Is the input already structured or accessible in a system?
  • Does it happen often enough for improvement to compound?

Axis B — Tolerance for being wrong

  • If the output is wrong, does a human see it before a customer does?
  • Is "wrong" obvious, or is it silently wrong?
  • Is there compliance, legal, or brand-safety exposure?
  • Can you verify the output faster than producing it yourself?

That last question is the one people skip, and it's the one that kills projects. If checking the output takes as long as doing the work, you have not automated anything. You've moved the work and added a review step.

03Sorting

Sort into four buckets

AUGMENT AI drafts, a human approves. Most content lives here. AUTOMATE NOW Build it. Your roadmap's first three items. LEAVE ALONE Effort exceeds return. Where projects die publicly. QUESTION IT Ask why you do it at all. Deleting beats automating. HIGH LOW LEVERAGE LOW HIGH TOLERANCE FOR BEING WRONG

Only two axes matter. Everything else is a distraction.

BucketProfileWhat to do
Automate nowHigh leverage, high error toleranceBuild it. These are your roadmap's first three items.
AugmentHigh leverage, low error toleranceAI drafts, a human approves. Most content and outbound lives here — and should stay here.
Leave aloneLow leverage, low error toleranceDon't touch it. Effort will exceed return, and it's where AI projects die publicly.
Question itLow leverage, high error toleranceAsk why you're doing it at all. A surprising amount of recurring marketing work fails this test.

Deleting work beats automating it. In most audits, at least one recurring task turns out to have no downstream consumer at all.

04Plumbing

Check the plumbing before you build

This is where most mid-market AI efforts quietly fail. The model is rarely the constraint; the data around it is.

  • Is your CRM data clean enough to feed anything? If lead source is inconsistent, AI lead scoring produces confident nonsense.
  • Can the systems talk? An automation that needs a human to copy between two tools has relocated the manual step, not removed it.
  • Is there a single source of truth for performance data? If three dashboards disagree today, AI makes them disagree faster.
  • Who owns it when it breaks? An automation with no owner is an outage waiting for quarter-end.
  • What does it cost at real volume? Per-call pricing that's trivial in testing becomes a line item at scale. Model it.
05Measurement

Instrument it, or you'll never know

Before building anything, write down the number it should move and where that number lives today. If you can't name it, you're not ready to build.

  • Baseline first. Hours spent, cycle time, output volume, error rate — measured before you change anything.
  • One metric per automation. "Time from brief to first draft," not "marketing efficiency."
  • Set a kill date. If it hasn't moved the number in 60 days, turn it off. Unused automations accumulate cost, risk, and confusion.
06Traps

The five traps

  1. Buying the platform first. The tool is the last decision. Start with the task inventory.
  2. Automating visible work instead of expensive work. Content generation is what everyone demos. Reporting and QA are where the hours actually are.
  3. Skipping the verification math. If review takes as long as doing it, you've added a step.
  4. No owner. Every automation needs a name attached and a documented failure mode.
  5. Publishing unreviewed output. Anything a customer sees keeps a human in the loop. That's a brand decision, not a technical one.
07Outcome

What good looks like after 90 days

  • Three to five automations live, each tied to a number you baselined
  • A documented owner and failure mode for each
  • At least one recurring task eliminated rather than automated
  • A short, honest list of what you tried that didn't work
  • Your team spending measurably more time on judgment and less on assembly

That's the whole framework. It's the same one I run in a paid engagement — the difference there is that I do the work, argue with your team about the scoring, and build what comes out of it.

See how that works