Your AI rollout isn't failing because of the AI.
I've been watching enterprise software rollouts since 1995. The AI version follows a script I know by heart: big kickoff, licenses for everyone, a mandate from the board, a dashboard tracking "adoption." Ninety days later the usage graph sags and someone schedules a meeting about "driving engagement."
Here's what you already have:
- Enterprise AI licenses — Copilot, ChatGPT, Gemini, take your pick
- A board that wants an AI story by Q4
- Three power users who didn't need permission anyway
- An adoption dashboard nobody wants to screenshot
And here's what you were sold: a workforce that would spontaneously reinvent how it works.
The problem: you're betting on a behavioral shift
People do not change how they work because you bought a tool. They never have. MIT researchers looked at hundreds of enterprise GenAI pilots this year and found the overwhelming majority produced no measurable P&L impact. The failed ones share a shape — they asked people to work differently before the AI had earned its place.
A rapid behavioral shift is the most expensive thing you can ask of an organization. Every changed workflow makes your experts beginners again. Beginners are slow, resentful, and creative about routing around the new thing. That resistance isn't a character flaw — it's your company's immune system doing its job.
AI that changes the workflow won't see ROI for a very long time. Maybe ever, because it usually gets quietly abandoned first.
The shift: stop rolling out AI to people. Inject it into workflows.
The teams getting real ROI aren't transforming anything. They're doing something much less glamorous:
Find the bottleneck in the workflow that exists today. Put AI at that exact spot. Leave everything else alone.
Small AI is effective AI — when it removes a bottleneck in the current workflow. Nobody has to be retrained. Nobody's job description changes. The work just stops piling up where it used to pile up.
At Q1Media we didn't hand the team a chatbot and a prayer. We put AI inside the operating system they already live in — at the specific steps where work sat overnight waiting on a human. Same inputs, same outputs, same muscle memory. The middle just got fast.
The how
I. Map one workflow end to end. Pick one team. Trace one unit of work — a campaign, an invoice, a support ticket — from the moment it arrives to the moment it's done. Write down every handoff.
II. Find where work waits. The bottleneck is rarely where people are working hard. It's where work sits — the approval queue, the copy-paste between systems, the "I'll get to it Monday" step. Ask one question: where does work sleep overnight?
III. Inject AI at that exact point — and nowhere else. The step should accept the same inputs and produce the same outputs it always has. Upstream doesn't change. Downstream doesn't change. If your rollout plan includes the word "retraining," you've picked the wrong spot.
IV. Measure cycle time, not adoption. Adoption is a vanity metric for tools people were forced to open. Cycle time is the truth. If the unit of work moves through the pipe faster this month than last month, the AI is working — whether or not anyone "engaged" with it. Then find the next bottleneck and do it again.
That's the whole playbook. It's boring. It compounds.
What I'm watching
- MIT's enterprise GenAI research — the "95% of pilots fail" finding that made the rounds. Read past the headline: the winners bought focused tools for specific workflow steps instead of broad platforms.
- Claude Code — the clearest example I use daily of AI injected into an existing workflow (the terminal) instead of a new destination app.
- Agent steps inside old-school automation — Zapier, n8n, and friends quietly adding AI nodes to existing pipelines. That's the injection pattern productized.
- The quiet death of the "AI transformation office" — watch how many get reorged into ops teams this year. The org chart is admitting the thesis.
The bottom line
Enterprise AI ROI isn't hiding in a bigger model or a better platform. It's hiding in your current workflows, at the spots where work waits.
Inject there. Measure cycle time. Repeat.
Do that for four quarters and you haven't run an "AI transformation" — you're just running a faster company. At that point, you are not adopting AI. You are compounding it.
What I'm building in my own 'lab'
TextMyAgent is this exact thesis, productized. I didn't build another app for you to adopt — your personal AI agent lives inside the workflow you already run a hundred times a day: text messages. It watches your email and runs your calendar from there. No new behavior, no new login. $99/year, 14 days free, no card to start.
And the Essentialist platform keeps powering this newsletter and a dozen other distribution plays — an agent injected into the publishing workflow, not a replacement for it.
One Last Thing Does your enterprise have an AI roundtable where leadership from different departments sit in a room and discuss AI workflows and AI strategies, where AI is having a good impact and where it's not working?
This can be a fantastic strategy to help your team get aligned. If you are interested in having a third-party AI-forward CTO sitting at that table, I can make myself available. Just reply to this email, and we'll strike up a conversation.
— I read every reply.
Best,
Keith
P.S. Let's connect on LinkedIn. https://www.linkedin.com/in/eddleman