I spent a year rolling generative AI into a marketing organization at a $50M manufacturer. Four brands, a $1.5M budget, four people, three agency partners. Then I spent the months after that using the same tools to build and launch a software product by myself.

The second thing taught me more than the first. Here is what stuck.

Adoption Is Not a Tools Problem

We cut product content production time by 38%. That number gets attention, so people assume the hard part was picking the right tool. It wasn't. The hard part was that a spec sheet writer who has done the job for eleven years does not want a machine involved, and no amount of demo enthusiasm changes that.

What changed it was prompt libraries. Not training sessions, not a policy memo. Actual saved prompts, written for the specific job that specific person does, sitting where they already work. The message stopped being "learn AI" and became "here is your task, already half done."

I had seen this before without recognizing it. At a global relocation firm I took Salesforce adoption from 52% to 87% across 140 users. Same lesson. People adopt tools that remove a step they hate. They resist tools that add a step they don't understand.

The Output Problem Nobody Warns You About

Generative AI does not produce bad writing. It produces average writing, at volume, instantly. That is a worse problem, because average writing passes review.

Three months in, all four brands started sounding like the same company. Clean copy. Correct copy. Indistinguishable copy. We were erasing the brand differences I had spent the previous year building.

The fix was a voice profile per brand: vocabulary, sentence rhythm, phrases we never use, and three samples of writing that sounds right. Then a review gate asking one question. Does this sound like us, or does it sound like a language model? If a draft could belong to any competitor in the category, it failed, no matter how clean it read.

Then I Built Something

Sales reps I have worked with for twenty-seven years all have the same problem. Their CRM was built for a sales team with a manager, an ops person, and a desk. They are one person in a truck.

My avatar has a name. Manual Tracker Mike, 47, independent commercial insurance broker, tech comfort about a three out of ten. He tried HubSpot. He tried Salesforce. He quit both inside of ten minutes and went back to a spreadsheet and a legal pad, and he is faintly embarrassed about it. He should not be. Those products are not built for him.

So I built Chumley. A ridiculously simple sales CRM. Dashboard, contacts, pipeline, lead cards, calendar. Fourteen-day trial, no credit card, about fourteen dollars a month for a solo rep.

I built the application with Claude Code. I did the branding, the identity, the website, the positioning, the paid search, and the SEO plan. One person, every layer. The full breakdown of how it came together is in my portfolio.

The Positioning Discipline

The obvious move was to attack Less Annoying CRM, the incumbent in simple. The obvious move was wrong. You cannot win "we are simpler" as a claim. It is unprovable and it sounds defensive.

The actual difference is physical. Less Annoying CRM is a simple CRM for a desk. Chumley is a simple CRM for a phone. Mike is not at a desk. He is in a parking lot between appointments with forty seconds to log what just happened.

That is a positioning rule, not a tagline, and it decides everything downstream. What gets built first. What the screenshots show. Which keywords get bid on. Whether a feature request gets built or declined.

AI did not give me that. AI helped me test it against forty-three keywords, competitor pages, and a stack of review data in an afternoon instead of a month. The judgment was mine. The speed was the tool's.

What Actually Transferred

AI collapses the distance between idea and artifact. The thing that used to take a team and a quarter takes one person and a week. That is real and it is not hype.

It does not collapse the distance between idea and good idea. Every hour the tools gave me back went into deciding what to build, who for, and what to refuse. The bottleneck moved. It did not disappear.

Judgment is the scarce input now. A model will write you a hundred headlines. Knowing that Mike is embarrassed about his spreadsheet, and that the headline has to make him feel less embarrassed rather than more, is not in the training data. That came from twenty-seven years of sitting across from people like him.

Accountability cannot be delegated to a tool. Someone has to own the output. That was true of the marketing org, and it's the same judgment I bring to every AI marketing consultant engagement now: the code running in production is somebody's responsibility, not the model's.

Where This Leaves Me

The marketing leaders I know are split. Half are waiting for AI to be proven. Half are using it to produce more of the same work, faster.

The interesting position is the third one. Use it to do things you could not previously do at all. I could not have built and launched a software product alone two years ago. I did not learn to code. I learned to direct, review, and take responsibility for work I could not have produced by hand.

That is the skill. Not prompting. Directing. It's the same thing I do inside a fractional CMO engagement, just aimed at a product instead of a campaign.

  • Adoption is not a tools problem: people adopt what removes a step they hate, and resist what adds a step they don't understand
  • Generative AI's real risk is average output at volume, not bad output: the fix is a documented brand voice and a "does this sound like us" review gate
  • AI shrinks the distance between idea and artifact, not between idea and good idea: judgment is still the scarce input
  • Built and launched Chumley, a simple sales CRM for solo reps, solo: branding, positioning, website, and paid search, using Claude Code for the build