Fighting for the Inches: How I Added AI Agents to My Marketing Team

I started out as a marketing AI skeptic. Every CEO is telling their CMO to add agents to their team and do it quickly so they can tell the board what they are doing with AI.

I sat through the demos. Every one of them showed me a model writing marketing copy, and I kept thinking the same thing: writing copy has never been my bottleneck. I have been doing this for over thirty years. I can produce a landing page, a whitepaper, or a launch narrative faster than I can explain what I want to somebody else and wait for it to come back.

So when I started using AI at BlastWave, I expected to be underwhelmed.

What changed my mind was not the writing. It was the first time a system read HubSpot, GA4, Search Console, Semrush, LinkedIn, and our call recordings in a single pass and handed me one page that said: “Here is what moved this week, and here is what it means”.

I had been assembling that page by hand, badly, for two years.

The Big Picture: AI in marketing

I run marketing with a Creative Director and a part-time Marketing Contractor. The scarce resource on a team that size is not hours, and it is not output. It is attention. There are nine systems holding pieces of the answer to “is this working,” and each is happy to give you a number in isolation, but is absolutely no help at all with what it means relative to the other eight. So you do what I did: check the two or three you trust most, make a decision, and privately hope the other eight aren’t saying something important.

Aggregation is the unlock. Not generation. The moment I stopped hand-assembling the picture, I got the thing a small team never has: the ability to look at the whole board at once and pick the next move on purpose, rather than reacting to whichever fire was loudest.

That sounds like a small change. It is not. It is the difference between running a marketing program and reacting to requests.

Skills are the multipliers, not agents

The part I got wrong at the start, and I think most people getting into this do the same, is that I thought the work was about building agents.

An agent is a role. A competitive intelligence analyst, a pipeline analyst, a technical content writer. Giving a model a role, a scope, and a list of tools gets you somewhere real, but it plateaus faster than you expect. A well-briefed generalist still makes generalist decisions, and if not guided correctly, it does not remember what you told it last week or how to get better.

The step change came from writing skills, which are procedures. Not who the agent is, but exactly how a specific recurring job gets done at BlastWave, not just generically speaking.

Our monthly competitive review is a skill. It knows the watchlist; it knows that a competitor move is only material if it bears on protection-first secure access rather than on monitoring; it knows what tier a change has to hit before it is worth my attention, and it knows to come to me for approval before anything becomes a battlecard. Our compliance work is a skill that operates at the clause level across IEC 62443, NERC CIP, TSA directives, CMMC, and NIS2, because the difference between a useful compliance mapping and a liability lies in whether it maps to the actual clause or to the general vibe of the standard. Our conference submissions are a skill. Our monthly path-to-lead audit is a skill.

This is the guidance I want to give all CMOs investing in AI.

Writing those skills forced me to write down how BlastWave actually works, and most of it had never been written down anywhere. It lived in my head, which meant it could not scale, could not be delegated, and would leave the building with me.

The exercise is not prompt engineering. It is documenting your own judgment, and it is worth doing even if you never point a model at the result. The most important thing is to iterate on your skills each time you run, so you keep getting better and better.

Train and verify your AI, don’t just trust it

I would love to tell you how fast the numbers went up once I did this. But that really isn’t the story (yet). Instead, I want to tell you about the many months my dashboards were confidently and elaborately wrong before I began my AI journey.

One of the first things I did when I began setting up automation was to mandate that every run include a counterpoint, question the results, and suggest ways to make each automation better and more accurate.

While creating my overall marketing dashboard, I built a validation step that runs before the monthly inbound journey report and determines whether the data can be trusted. In its first run, it failed miserably. GA4 was reporting more new users than sessions, which is arithmetically impossible, and it had been reporting month-over-month traffic flat within a few percent for half a year, which real traffic does not do. I trusted GA4, so I never really questioned the data. However, since I connected multiple data sources for this report, another data point used for comparison was Semrush. Semrush reported a very different number of organic visits than GA4, which was more realistic.

A journey report built on the GA4 data would have looked completely normal. Every chart would have rendered. Every conversion rate would have computed cleanly, against a denominator with almost no humans in it, and I would have made real budget decisions on it. The report would not have been wrong in any way I could have seen.

That is the actual risk in this whole category, and almost nobody selling you an AI marketing stack will say it out loud. Aggregation does not make bad data safer. It makes it more dangerous, because it launders it into a clean chart with a confident summary on top. The output is smooth either way. Smoothness is not a signal.

So the first thing worth building is not the report. It is the gate that decides whether the report is valid and realistic, and what might show a contrary view. If your system cannot take the position that you should not trust this month’s numbers, and here is why, it is not a reporting system. It is a machine for generating confidence.

Fix the skill, not the output

Two more misses taught me the operating rule I would hand to anybody starting an AI stack in marketing.

I had another case where the data was completely wrong in the opposite direction. One report came back claiming that barely 7% of our contact records had an entry URL, which reads like a total attribution collapse. It was not. When properly segmented by how each record was created, contacts that came in through a form had an entry URL 99% of the time. The rest were from an imported prospect list and records created by sales, people who never touched a form and correctly have no entry page. The alarm was real; the interpretation was garbage; the difference was a single segmentation step that required context on how BlastWave operates.

Another example: for a few weeks, one of our weekly checks kept flagging qualified leads with no associated deal as a data defect. They were not defects. BlastWave does not attach a deal until a record reaches Opportunity stage, which is a deliberate choice about how our funnel works. The check was applying a generic B2B assumption to a company that does not run that way.

The tempting fix is to correct the report. Delete the false flags, move on, ten minutes.

The correct fix is to change the skill because the failure was the model reverting to its default assumptions, and it would do so again on the next fifty runs. Correcting the report takes ten minutes and buys you nothing. Correcting the skill takes ten minutes and saves you the next fifty. Once I did that, the analysis of the success or failure of the lead flow changed drastically.

I will not pretend this was cheap. The guardrails were most of the work. Standing up agents and skills took days, and building the parts that check them, ground every claim against verified source documentation, and refuse to let anything reach a customer without a person approving it has taken months and is not finished. But this is where you make the biggest long-term gains.

Validation and continuous improvement win long term

An agent can watch data all day and draft anything it likes, and a human still reviews, approves, and pushes. That is not a limitation I am working around. It is the design, and it is the only reason I am comfortable running this at all. I generally let something run for at least a month with multiple improvement loops before I start to trust the data and the process.

For my OT readers, these agents are non-human identities with credentials into our CRM, our analytics, and our Slack, which is the same thing you are about to be asked to approve on a plant floor. I would rather learn what they get wrong somewhere with a blast radius that’s an embarrassing chart.

Why this is critical at a startup

At a startup, every day you seek to make the biggest impact on the business that you can. My CEO, Tom Sego, loves The Bear and the sign in one of the episodes that says “Every Second Counts”, because he understands that with scarce resources at a startup, every second you save can be used to add more value to the company.

AI has given me more seconds every day. But the bigger impact is the one Tony D’Amato (played by Al Pacino) is talking about in that locker room speech in Any Given Sunday, when he tells his team that life, like football, is “this game of inches,” and that the inches are everywhere around us if you are willing to claw for them. You do not win by one enormous play. You win by adding up the inches until they are the difference between winning and losing.

Over the past year, I have been gaining those inches every day. I believe that it will make the difference between winning and losing, and every time I see the impact that marketing has on the business, I work just that much harder.