Audit before you automate.

§ I — Cover concept

The context behind the article.

Journal 042

5 min

Image direction

Two stacks of paper on a worn wooden desk — one tall, one a single sheet — late afternoon light

AI

5 min

Article

An association almost spent $180K automating a workflow that ate twenty-four hours per year, while an 8,680-hour problem sat untouched. The fifteen-minute audit that caught it is the cheapest piece of work in any AI engagement.

Why it belongs in the journal

This entry exists to make the operating logic visible: not just the system we would build, but the constraint, tradeoff, or failure mode that forced the architecture to matter in the first place.

§ II — Article

The wrong workflow

An association came to us with a $180,000 vendor quote for an AI-powered event-management platform. They were sure event registration was their biggest workflow problem.

It was not. Event registration was 1.1 hours per event across 22 events — about 24 hours of staff time per year. Their biggest problem was sitting two floors away, in the certification team.

Certification processing was eating 6.2 hours per application, across 1,400 applications a year. That is 8,680 hours — roughly 4.2 full-time equivalents — spent on manual document verification that nobody had ever clocked.

They almost spent $180K solving a 24-hour problem while an 8,680-hour problem sat untouched.

The audit that would have caught it

The question that surfaced this took fifteen minutes:

"What is your cost per member interaction, across all workflows, today?"

Silence. Nobody had measured it. That silence is the audit. Everything after is bookkeeping.

The framework we now run with every prospect, before any AI scoping work, scores five dimensions on a 1–5 scale:

Volume. How many times does this workflow run per year?
Unit cost. How many staff hours does one instance consume?
Variance. Does the work change shape every time, or is it the same shape with different fields?
Decision content. How much of the work is judgment vs. transcription?
Reversibility. What happens if the AI gets it wrong?

A workflow scoring 18+ out of 25 is a candidate for automation. Below 11, automation will cost more than it saves — the dollars and the trust both. Between 11 and 17, it is an instrumented pilot, not a platform purchase.

Why this matters more in AI than in classic automation

For classic RPA, the bias toward the wrong workflow was expensive but bounded. The vendor quoted a fixed price, the wrong workflow got automated, the savings did not materialize, the contract ended. Painful, recoverable.

AI vendors do not work that way. They sell platforms with year-long minimums, integration commitments, and a maintenance burden that grows over time. MIT CISR's "GenAI Divide" study found that 60% of companies evaluated enterprise AI tools and only 5% reached production. The other 55% paid for evaluation, integration, and contract exit. The bill arrives whether or not value does.

What we now require

We will not scope an AI engagement until the audit is on the table. Not because we are precious about process — because the audit is the cheapest piece of work in the engagement, and it is the only piece that can save the client from spending the rest of the budget on the wrong thing.

Fifteen minutes. Five questions. One number that nobody had measured. That is the gate.

§ III — Reading note

What the article is really about.

Operating tension

An association almost spent $180K automating a workflow that ate twenty-four hours per year, while an 8,680-hour problem sat untouched. The fifteen-minute audit that caught it is the cheapest piece of work in any AI engagement. In practice, the hard part is usually not implementation syntax but aligning delivery, controls, and operator trust so the thing can survive contact with a real team.

Kleio view

We treat these articles as public design memos: short, opinionated, and anchored in systems that have to be bought, operated, and defended long after launch week.

§ III — Continue reading

Three adjacent articles.

Cloud Computing

Modernization is a pipeline, not a rewrite.

We killed a $400K rewrite in week two and replaced it with a three-tier modernization pipeline. Eleven modules extracted, 19,000 lines of dead code retired, drift down 62% — without a single big-bang cutover.

Agentic AI

MCP servers are the orchestration layer, not the demo.

Most AI tooling lives one paste away from the systems it could help with. Model Context Protocol servers close that gap without giving up the security posture an enterprise can defend.

Agentic AI

Subagents are a team-management pattern.

One prompt refactored a payments service in twenty-two minutes by dispatching three subagents in parallel. The interesting part is not the speed. It is that the pattern matches how strong engineering leads have always worked.

Back to journal