The question that started the Herman Scheer engagement didn't come from inside the agency. It came from their clients. Brand-side marketers had started asking, in pitches and in renewals, whether the agency had AI capability. Not as a curiosity. As a procurement question, the kind that gets a box ticked on a vendor form.
When I asked, the agency's leadership gave me the honest answer: no. Not no talent — this is a Los Angeles branding shop whose packaging you have walked past in every supermarket in the country. What they didn't have was anything to point at. No shared method, no quality bar for AI-assisted work, nothing a client could inspect. And the fastest-looking fix on the table, the one that gets pitched to agencies constantly, was to buy a batch of course seats, run the team through them, and put a line in the credentials deck.
We didn't do that. Instead we trained the whole team on their live work, and before the training block was finished people were putting client-ready image, copy, and video into active CPG launches. Estimated year-one production savings: $4.5 million. What made that possible wasn't the curriculum. It was that nobody could finish the program without shipping something real.
What a certificate actually certifies
I want to be careful here, because I'm not against structured courses and I use them myself. Every's team just ran their whole staff through Anthropic's new four-course certification and wrote it up honestly. Their verdict was that it works as "foundational vocabulary" — common definitions for skills, MCP, APIs — rather than as a guide to doing the work. That's a fair and useful thing for a program to be. Shared vocabulary matters; half the confusion in a creative department is four people using the word "agent" to mean four different things.
The failure isn't the course. It's the substitution. A certificate certifies attendance and recall. Creative work is neither. Nobody has ever hired an agency because its art directors could define a diffusion model. They hire because the work is good, it arrives on time, and it survives legal.
The substitution is tempting for a specific reason: a certificate is legible to procurement. It's a line item. It answers the client's question in a form the client's form can accept. That's exactly why creative leaders reach for it, and exactly why it doesn't move anything. You have optimized for the audit, not the output.
The gap between passing and shipping
Here's the pattern I see in week one of nearly every engagement. Ask a room how many people have done some AI training and most hands go up. Ask how many have shipped a client deliverable where AI did meaningful work in the middle of the process, and the number collapses. Not to zero — usually to two people. The same two people who would have figured it out anyway.
That gap is the whole problem, and it isn't a knowledge gap. The people in the room can explain what a model does. What they can't tell you is which step of their job it belongs in, what a good output looks like against their own standard, and who signs off when it's wrong. Those three answers don't come from a course, because a course doesn't know your line review, your retailer specs, or your client's tolerance for risk. They come from doing the work with the tool inside the real process, with the real constraints, while somebody who has done it before is standing there.
The Every review flagged a small detail that makes the point better than any argument I could build: one of the courses still teaches against a model that's no longer available in the API. That's not a knock on the program — every static curriculum decays. It's a description of the category. Anything general enough to certify at scale is, by construction, behind your actual production reality. Techniques age slower than tools, which is why I teach the technique and let the team pick the tool.
What I ask for instead
When a client asks whether your team has AI capability, four things answer better than a credential, and all four are things you can show.
Work. Named jobs that shipped, with the part AI did stated plainly. Not a demo. A deliverable a client paid for and approved.
A method. Written down, in your language, describing how a job moves through your shop now — where AI enters, where a human decides, where it stops. One page beats a curriculum.
A quality bar. The standard AI-assisted work has to clear before it leaves the building, and the name of the person who enforces it. Most agencies already have this for everything else. Extending it is a smaller job than they expect.
Coverage. How many people can do this, not which two. This is the one that separates teams that got faster from teams that got a bottleneck with a new job title. Across the engagements I've run, estimated combined client savings sits above $8M, and none of it came from concentrating the capability in an enthusiast.
None of those require a vendor. They require a team that has done the work under supervision and written down what they learned. That artifact — your method, in your voice, describing your line — is the thing clients actually respond to in a room, because it proves you've thought about their risk, not just your own upskilling.
The honest version of the client's question
What the brand-side marketer is really asking is: if I hand you this brief, will the work be good, will it come back faster, and will anything about how you made it embarrass me later?
A certificate answers none of that. It answers a different question — did your people attend — which nobody was asking. So keep the courses. Use them for vocabulary, so your team stops talking past each other. Then go install the capability where it actually lives, which is in the middle of a real job, with real stakes, with everyone in the room and not just the two who were going to be fine either way.
That's what I'd want to see if I were buying. It's what I'd want to have if I were selling.