Issue #011·12 Nov 2026·AI LEADERS DECODED

Your Best Account Director Is Your AI Problem

If every AI-assisted deliverable needs their approval, faster production just creates a longer queue.

The draft is done. The person who can tell you whether it’s right is in three client calls back-to-back. I keep coming back to that gap, because that’s where your AI productivity disappears.

I spoke this week with Katya Tarapovskaia, founder of Ustellar, who previously worked at Snowflake, Twilio and Mastercard. She now builds AI-powered marketing systems, and says agents execute around 80% of her work.

But the line that stuck with me wasn’t the percentage. It was what she said about the work that remains:

In the end you need a human validation, human in the loop.

Exactly. But which human? Checking what? Against rules written where?

That’s the bit I’d worry about in your agency.

Your best account director knows that one client hates the word “innovative”. They know the founder will reject anything that sounds too salesy. They remember why you stopped making that particular claim in March.

So everything goes through them. Sensible. They catch the mistakes.

Except now your team can produce drafts much faster, and that person still has the same number of hours. You haven’t removed the bottleneck. You’ve built a faster conveyor belt into it.

The thing I keep coming back to is the difference between judgement and memory.

Judgement is deciding whether a campaign is worth running. Whether the angle is strong enough. Whether the client needs challenging rather than accommodating.

Memory is remembering that this client uses “members”, not “customers”. That pricing claims need approval. That the audience already understands the category, so stop explaining it from scratch.

You’re paying senior people to do both. Then calling all of it quality control.

That hides the problem. Some of those checks need an experienced human. Others need a rule your tools and team can actually find.

I’m not arguing for unchecked output. I’m arguing that your best people shouldn’t have to keep being the only place the instructions exist.

So here’s what I’d do this week. Take one AI-assisted draft your account director corrected. Spend five minutes beside them and ask one slightly uncomfortable question: “Which of these edits could someone have got right without you, if we’d given them the rule?”

Write down their answers. Don’t redesign your delivery process. Just separate the missing instructions from the actual judgement calls.

You’ll learn whether your review queue is protecting quality or compensating for context you never captured.

In the conversation, Katya describes a client delivery comparison that makes this bottleneck hard to ignore, and explains what she measures beyond output volume. Watch it with your own approval queue in mind: https://www.youtube.com/watch?v=wCeLZ4n7LOc

Talk next week, Michael

This is why I’m building Vythos: an AI company brain that gives your tools your client rules, brand voices and ways of working. Less re-explaining, more room for the judgement you actually need your team for —

Learn more about Vythos →

The Review Split

Only in the newsletter not in the video.

Use this review split on one recently corrected AI draft. Ten minutes, with the person who approved it. Copy the fields below into a document and repeat them for three edits.

  • Original wording or decision: ___
  • Correction made: ___
  • Why it changed: ___
  • Type of correction: missing client rule / missing factual source / genuine judgement call.
  • If it was a missing rule or source: what exact instruction or reference would have prevented it? ___
  • Scope: this client only / this project only / agency-wide. Confirm with: ___
  • If it needed judgement: what decision must still come back to a human, and who owns it? ___

Before using any captured rule, have the reviewer confirm it. Then put the confirmed instructions into the next brief for that client. Keep the judgement calls as explicit review points.

On the next draft, check whether the same corrections recur. That's your test: fewer repeated edits, not merely more output.

Watch the Full Breakdown

Watch the full breakdown on YouTube

Get this every week.

Every week we translate the top AI news into clear strategy and one practical thing to actually do this week. For founders, builders, and operators. Plus a subscriber-only template, checklist, or teardown in every issue.

← Back to the archive