This article is based a Q&A with Mitel VP of Marketing Operations & Customer Lifecycle Marketing Andrew Black and Greg Kihlström for The Agile Brand Guide’s Expert Mode series.
The standard account of AI in marketing is a labor story. Drafting gets cheaper, teams produce more, and sooner or later somebody asks the headcount question. Most of the conversation at the CMO level is still about output.
Andrew Black, VP of Marketing Operations & Customer Lifecycle Marketing at Mitel, thinks that’s the less interesting half. His team runs on a rule: you can’t walk into a meeting and say Claude put this together. If your name is on the work, you own all of it, accuracy included. What he argues from there is that AI didn’t mainly change what gets produced. It changed what a review is for, and which people on a marketing team are worth betting on.
Excel Never Decided Anything
The usual rebuttal to a rule like Black’s is that it won’t age well. Nobody credits Excel on a financial model or names the agency that wrote a draft. Once AI assistance is just how work gets made, insisting on ownership starts to look like a transitional courtesy. Black doesn’t accept the analogy.
Nobody says Excel built this model because Excel never decided anything. A person frames the problem and the assumptions, and then the tool does the math. That counterargument only holds if AI is that kind of tool. It’s not. It makes assumptions and decisions nobody made or asked for, inside work that reads as finished.
The distinction does more work than it first appears to. A spreadsheet executes assumptions a person supplied. A model supplies its own, inside prose that reads finished. For us as marketing leaders that’s a different shape of risk, because the failure mode isn’t a visible error. It’s a quiet one that clears review precisely because nothing about it looks wrong.
His second point is about blast radius. One person with a spreadsheet produces one bad model in one file. One badly built skill produces the same wrong claim on the website, in a campaign and in a byline, all at once, and keeps producing it until somebody fixes the skill. Which is why he thinks the rule hardens rather than relaxes as models improve: “The rule gets more necessary as the models get better, not less.” Read that way, human-in-the-loop stops being a compliance checkbox.
The Error Isn’t in the Article
Asked what his team checks now that they weren’t checking three years ago, Black didn’t describe a longer checklist. He described a different first question. A review used to be about the piece in front of you. Now the reviewer decides whether the problem is systemic.
We had AI surface a platform we only sell in Europe inside an article written for US enterprise buyers. Correcting the article would have been easy, but it was the wrong fix. There is an underlying skill driving that output, and if the regional designation was missed when the skill was built, it gets missed everywhere downstream. Web, campaign, thought leadership, all of it.
Fixing the article takes a minute. Fixing the skill takes a conversation with marketing operations, and it protects everything that skill touches afterward. Before AI, someone on the content team would have checked the product pages and asked product marketing which examples suited a US enterprise audience. That conversation still has to happen. What changed is the assumption people bring into it — “People assume the output is already correct because our skill-based operating system produced it.”
For a CMO the translation is a change to what a reviewer is accountable for. If your process ends when the draft is correct, you’re buying the same catch over and over, by different people, forever. No fix outlives the person who made it. A reviewer needs somewhere to send a systemic error and someone who owns the skill on the receiving end, which is a marketing ops staffing question more than an editorial one.
The Supervision Tax Is Optional
Optimizely’s June study, fielded by Savanta across 2,003 marketing leaders in seven countries, found 76% of marketers spending at least three hours a week editing, fact-checking or correcting AI output. Only 4% said AI saves them time at every stage. Put to Black as a challenge — the gain isn’t real, the work just moved from writing to reviewing — he gave away more ground than expected.
I’d concede that in one case. If you’re using AI as a crutch (generate, skim, and ship) then yes, you moved the work from writing to reviewing and you should expect exactly what that research found.
What separates his case from the survey data is what happens to the rework. Generate-skim-ship treats every correction as a cost. Black treats it as an input, improving the system that produced the draft so the next cycle starts further along. Whether that compounding actually happens at most companies is fair to doubt, and the Optimizely numbers suggest it usually doesn’t.
Where he won’t concede is anywhere the inputs are structured and the process is formulaic. Demand generation is his example. Pulling target accounts, reading them against CRM notes and intent signals, then prioritizing the outreach — that used to be weeks of analyst work, plus BDR hours burned on the wrong accounts every day a campaign was live. “There’s no supervision tax that eats a gain of that size once you get the machine working.” He puts the result at 12.5% of his function’s budget freed in the first half of this year. That’s his number, not an audited one. The useful part is the question it implies, which is narrower than is AI saving us time. It’s whether the work eating your review hours is work nobody ever structured the inputs for.
Judgment Is Knowing What Not to Build
On hiring, Black gives the obvious answer first and says so. AI fluency became a requirement for his team in Q4 of last year, alongside named operating principles and an answer engine optimization goal that runs across the whole marketing organization. Then he sets it aside. Skill is trainable. What he’s sorting for is will, and lately something narrower than that.
Do you know what is not worth doing? I look at what someone declined. When the cost of producing something collapses, everything starts to look cheap enough to try, and saying no gets harder than saying yes. Nobody should be standing up a CRM over a weekend because they saw a demo on LinkedIn.
That’s a performance criterion almost nobody has written down. We evaluate people on what they shipped. Nobody gets credit for the initiative they talked the business out of, and when production costs fall far enough, the number of things that look worth trying grows faster than the capacity to finish any of them well. Black’s version is a gate: every request names the goal it serves and what it’s worth before work starts.
The other half is enablement, and it’s where the recovered time is supposed to go. He describes a device promotion that used to eat most of a week in assembly, leaving whatever was left for getting agreement. The analysis now takes a fraction of that, and the expectation isn’t that the promotion ships sooner. It’s that the difference goes into adoption across other functions. “Writing a good prompt doesn’t make you a supply chain VP.” Same reasoning on Mitel’s first in-house AI application: data governance, IT, legal and executive leadership came in before anything got built, which is an AI governance board in everything but name. It cost weeks. It’s why the builds since have had sponsors instead of roadblocks.
Black’s rule started as a line about what you can say in a meeting, but the test he’d hand a CMO has nothing to do with attribution. Read the last three pieces of thought leadership your team produced with AI. Differentiated, or vanilla? If it’s vanilla, the review process isn’t your problem. The system is, and usually nobody owns it. Most of these systems get trained on what the company has already said, which makes them very good at producing more of your own past. “It’s hard to build the future when your AI is looking in the rearview mirror.”
There’s a faster version of that test, and it’s worth stealing. Black’s CMO had him screenshare the operating system agent he keeps for his function, then put the last two questions he’d asked Black directly to the agent. One answer landed. One didn’t, because the agent didn’t have that context yet. Five minutes, and it told them something a process audit wouldn’t have. Go run that one this week.
