PureSEM Blog

Managing AI Agents: 6 Lessons for Marketing Teams | PureSEM

Written by Keith Holloway | Nov 18, 2026, 9:35:48 PM

Your team of four is a team of 40. Or at least it should be.

If you're an individual contributor this is what work looks like for you now. You've got AI assistants everywhere. Maybe you're just talking to ChatGPT. Maybe you've got a dozen agents running. And if you manage people, everybody you manage is like that too. So you went from managing a team to managing a team of teams.

That's what Dave Shanley (Content Camel) and I talked about on Build Report #4, our fourth webinar together. It wasn't super technical. It was more about management, and how we all need to grow up our leadership and management skills, because that's where we're at today.

Imagine 100 people showed up tomorrow and started working for you. What would you do?

Most of us aren't ready for them. That's where the fatigue and the pushback come from, and the output that isn't what you wanted.

Even if you don't know it, your partners, your vendors and your software tools are all running agents right now. You may be reviewing work you don't even realize was done by an agent. If you work with PureSEM, we might be running 40, 50, 60 agents on your work. So how do you review all that? How do you make sure it's lined up with your strategy?

These agents are super smart. They're going to produce so much information, and it's probably going to be 85% on point and 15% off target, and somebody has to review all of it. They also don't know how to escalate. People have the life skill to make a good judgment call about when to bring something up to the boss and when to make the call on their own.

Agents don't. You have to teach them.

Watch the full session with Dave below, or skip ahead to the six lessons:



Why Prompt Packs and Named Agents with no Context Fall Short

We're bombarded with people telling us we can solve everything with AI if we just use their prompt set or their skill set.

I saw one on LinkedIn the other day with a 10-by-10 matrix, 100 go-to-market systems, and all you had to do was reply to get it. What is that, really? It's 100 prompts. Same with the go-to-market brain, and the agents with names and job titles and nothing behind them.

A prompt is the very tip of the iceberg.

It's just what you're trying to do right now. What your prompt can do depends on the skills under it, and your skills are your processes. You have to break those down and say, okay, how do we do things here? Then load them up when you need them. Under all that is the data and knowledge an experienced person on your team would have, or would know where to get. Your meetings, your CRM, your sales materials. Your agents don't have that by default. You have to pull it in, and it's different for every task. No one vendor has all of your context.

You have to start where you are. Our system, VERA, started over a year ago as a spreadsheet of the things we wanted to do and the mapped processes that already existed. Then it was a set of instructions that got to 15,000 words and started to break.

Now it's part of PureSEM and has 18+ categories of skills and processes and instructions on what to use when, with each one containing up to 15 tools it can call. Everything is connected to the puresem app which has dozens of data stores full of context. 

You don't need that on day one. You build it up bit by bit, and these six lessons are how.

The Six Lessons

  1. Start with the outcome. Don't tell AI how to do things, ask it for what you want to achieve.
  2. Put the troll where the troll needs to be, a gate that says you can't pass unless this is correct.
  3. Write the goal down at every level and push it from the top all the way down, through your people and your agents.
  4. Figure out what good looks like, write it down, and put it where the whole team can get at it.
  5. On the projects that can't fail, figure out up front what's going to go wrong.
  6. Find the thing slowing you down the most and get all the way to ground on it, instead of going wide.

1. Start With the Outcome and Let AI Design the Plan

A lot of people are prompting straight to the answer: hey AI, do this thing this way, write me 10 headlines.

That's where you get way more failure points and work that isn't what you expected. What you want to do is tell it the outcome you're after, ask it to design the plan and the experiment, and tweak from there. In our experience, your expertise mixed with the AI's design and execution gets you somewhere much better than either one on its own.

Dave wanted better open rates on an outbound email campaign. The AI already knew his ICP and what he was pitching, so he didn't ask for subject lines. He told it, more or less, we'll live or die by the subject line, so iterate on ideas and have a stand-in for the prospect judge them. It went off and built a mini lab. It set up a virtual inbox for a normal Tuesday, played the prospect in three moods (rushed, skeptical, curious), buried his subject lines among the internal email and the usual vendor pitches, and watched which ones got opened, skipped or deleted. Dave never asked for a lab. It came up with that because he asked about the outcome.

My version of this is to ask AI to go research what other firms are doing and the best way to get the outcome I want, then make a plan. Usually I have a pretty good idea what's coming back. Sometimes it comes back with something I never thought of, and we do it that way.

It's the same way you'd run a team of smart people. You don't walk in and say, team, this is what I want, and this is exactly how I want it done. You say, let's talk about the best way to get there. If you lead a team, ask yourself whether your people are doing that with AI, or just coming at it with the answer.

2. Put a Gate After Every Step That Matters

Dave and I see this all the time. You have AI write a 10-step process. You ask it to follow the process. You look at the output and say, wait a minute, step three explicitly tells you to do this. And it cheerfully says, oh yeah, you're right, I missed that, my bad, let me do it again. It was in the instructions. It just went so fast it ignored it.

So you put the troll on the bridge. A gate is a yes-or-no checkpoint: you can't go to the next step unless the last one is right. I've found two ways to do it:

  • Check-back steps. Before step three starts, go back and check that step two meets these criteria.
  • Phases. Do steps one to three, wait for the output, look at it, then give the next set of instructions.

I use the same idea on my team's work. Because of AI, somebody might send me 10 pages of documents and spreadsheets. I'll ask AI to check it against our standards, which are saved and which it understands. It'll come back with something like, this is 80% there, but it missed these two or three things, and I send that back. When we first started with AI, I was reviewing so much stuff, with so many mistakes. That review fatigue has come way down.

Keep all of this where the whole team can get at it. Shared Google Docs are fine to start. The next level up is a team account (ChatGPT and Claude both have one), where you can share projects and checklists.

3. Write the Goal Down at Every Level

Dave brought this up about a month ago, and it's become mission critical for me. Every task is full of micro decisions, made by people and now by agents. If whoever's making them doesn't know the company's goals this quarter, the team's goals and who we're serving, all those little decisions drift away from where you want to go, and you're surprised at the end.

How many meetings have you been in where you're the leader, you're deep in the detail, and you walk it back and say, hang on a sec, let's go back to basics. What are we trying to accomplish here? What's the objective? That's such a useful question. I don't think we're asking it enough when we build agents.

I caught myself on this one. We had a whole agent process, going back and forth, to make sure our content analysis for a client was as good as it could be. And you know what was missing? Anywhere that said, here's what this client is trying to achieve, here's the customer they're trying to reach, and the goal is to get traffic and persuade people to fill out a form and become a customer. It's implicit. I'm sure the AI understands that's the point. But nowhere in that agent process is it called out.

It doesn't have to be complicated. A line or two of North Star stuff, passed down to the lowest-level task: remember, this is for this ICP, this end customer, and this is what we ultimately want. It can't hurt, and it nudges everything a little in the right direction. Everything's moving so much faster now that any misalignment gets massively multiplied.

4. Define What Good Looks Like, Then Save It

My main interface to work has become AI. I'm not working in spreadsheets or documents or software anymore, because AI is connected to all of it and it's so much better at using the tools. Say I'm working on a keyword strategy. I'll spend an hour or two telling it where it's wrong, where it's not good enough, why it needs to do something differently. Then at the end I say, okay, unpack everything we just did and write it into a skill.

Our keyword skill is on version 3.8 now. The next time I've got 10,000 keywords for a client that need to be organized and grouped by topic and subtopic, I ask Vera to get these keywords up to standards, and it goes off and boom, it just does it. Stuff that took days or weeks before now happens in a couple of keystrokes. It's shocking.

Dave has a line I love: rework is waste.

If you're fixing things downstream after they're done, your process is broken. So take what you had to fix and push it back upstream into the instructions. Do that three, four, five times, and the output gets incredible.

Dave also doesn't like the word taste, and he's convinced me. Taste sounds like something you either have or you don't.

But they call it acquired taste for a reason. AI makes taste calls all the time. It'll tell you it likes this better than that. The problem is that its context is too narrow, so the taste is wrong. You have to teach it what good looks like to you and to your organization.

The bonus is putting it where how your

If this feels slow, that's the same old delegation trap. The first three times, it is faster to do it yourself. But if you have to do that work a hundred times, the other 97 are pure gain.

5. Run a Pre-Mortem on the Big Projects

Dave has been doing pre-mortems with his teams for years. Everybody knows the post-mortem. The campaign goes live, it doesn't hit expectations, and afterward you figure out what you could have done better. A pre-mortem pulls some of that to the very beginning. You've got your plan, you're sending down your objectives, you've got your gates. Now take a pause and ask, what's going to go wrong with this? Or say, we launched this next month and it completely failed. What happened?

I learned this one the hard way. We're building a benchmark report on the AI visibility of companies that sell government technology. I set up a project and gave it my objectives: a methodology rigorous enough to survive scrutiny, statistically significant, defensible. It came back with a plan to pull 300 companies and compare them all. I thought it would take a couple of weeks. It's been months. Every week I felt like it would be done next week, and the last plan I asked for said 10 more weeks.

So I finally went to a different project and asked, why is this taking so long? It keeps growing, it keeps getting more complicated, I keep filing more issues. What is going on? And it told me one of my biggest problems was the premise I started with. I wanted it as rigorous as possible, and the more we dug in, the more work it took to live up to that. I'd basically built a moving target into the project. We reined it way back, and it's a two-week job again. If I'd asked what could go wrong at the start, I probably would have seen it.

The other flavor is the devil's advocate. I was making some pretty significant architecture decisions on our software, and I had a plan I liked. But it was such a big decision that I stepped back and said, I want you to completely play devil's advocate. You're a senior software architect. Tell me why this is a bad plan and what you'd do instead. Maybe there's a much simpler way. It came back with basically a completely different plan. Then we played those off each other and came up with something that just felt so much better.

Dave's nerdy, tactical tip is to do that review in a fresh session. It's simpler than it sounds. Download the plan from the first chat, upload it to a new one, and say, hey, I created this plan, tell me what's wrong with it. It's amazing what it finds.

Don't do it on everything, though. You know the guy on your team whose perspective you value so much when you're doing something big? If he showed up on every tiny, low-risk change and said, hey, how about we do this completely differently, you'd be like, no, we just need to do this.

6. Fix One Bottleneck at a Time

This is the one I struggle with. I've got something like 27 projects on the go, and everything seems important. I'll get it down to, okay, these are the five most important things this quarter, we're only working on those. But the other projects don't go away. They rear their ugly heads every day. Hey, what about me? Oh yeah, you're important too, let me give you a little attention. And then you've got 27 things half done. Even our five quarterly rocks were too many for me. Last quarter I owned four of the five, and we missed. Now I own two.

Here's what works for us:

Twice this year, when things started getting stressful, we held an open team meeting and asked, what are your top five biggest obstacles to getting your work done? Everybody went around. We grouped them together and sorted until we ended up with the top three problems in the company.

Then we took the first one. This is the biggest problem across the whole team, so what can we do to solve it and automate it? And we worked that until it was done.

Then the next one, and the next one. And you know what? For the next two months, everything's great. Then it starts to creep back in, you've got a new bottleneck, and you do it again.

When you get one all the way to done, you feel it. We send every client a weekly update on what's happening in their SEO and paid search. Now AI analyzes all the data, figures out what's going on, puts the email together and sends it out. We don't even think about it anymore. We just read the email and comment on it, or bring it up in a meeting. It's hard to imagine that six or eight months ago that didn't exist. The monthly reports are next.

Dave's team did the same with the Content Camel newsletter. If no single person owns it, nobody owns it, and it doesn't go out on the schedule you want. So they built a system that pulls in industry news and their own point of view and works it all the way through their writing engine, right up to the final edit and hitting play.

How to Stay in Charge of Your Agents

Dave told a story about Frank Lloyd Wright. One of his clients supposedly said that first you hire him, then you realize you're working together, and by the end you're working for him.

I felt that one viscerally. You're working with AI, and you ask it to do one thing. It comes back and says, okay, I did that one thing, but there are these two other things you should probably be thinking about. So you try those two other things, and it gives you two more for each of them. Now you're drowning in work, and you go, how did I get here?

Every Team Member Is Now a Team Lead

If you're a non-technical leader and this feels a bit scary, here's the good news: most of what we covered is just good management. Forget about AI.

Dave said something on our prep call that stuck with me: you need to fold this in as a strategic advantage for your team. It's a threat and an opportunity at the same time. If you don't do it, you'll fall behind. But we're at a point where if you start doing it now, you can be ahead of the game.

The tech is new. The process isn't. Working with a bunch of humans papers over a lot of the cracks, because things move slow enough that you don't notice how bad it is. With agents, everything moves fast enough that you notice right away. There are no individual contributors anymore. We're all managers now, so we all have to get good at management.

See how your buyers find you in AI search. We'll give you a free AI visibility assessment. It shows whether you or your competitors come up when buyers ask ChatGPT, Perplexity, Claude, Gemini and Copilot, how often, and what to do about it.

Request your free assessment.