This is the final part of a 3-part series on building an AI-powered marketing team. Part 1 covered the individual trap of AI and how to build the first shared system. Part 2 covered proving it's working and keeping it safe. This week: getting the team to actually use what you've built.
Deloitte’s 2026 State of AI in the Enterprise surveyed 3,235 senior leaders and found something that should worry anyone who’s just spent two issues building and governing a shared AI workflow: talent readiness is the lowest-scoring dimension of AI maturity of anything they measured, at just 20%. It’s also dropped two points year-over-year, even as 84% of those same organisations are increasing AI investment.
Money going in, readiness going down. Only 13% of workers say their employer offers any AI training, per SurveyMonkey. 75% of people using AI at work taught themselves.
So here’s the sequel to Parts 1 and 2.
The adoption trap is Part 1’s trap, wearing a different coat
Going back to where this series started: solo AI workflows persist because they’re easier than whatever the “official” alternative is. That problem doesn’t go away once you’ve built the team system.
SurveyMonkey’s research puts a number on it: 29% of employees admit to using AI at work without telling their manager. Call it shadow AI. If your shiny new shared workflow requires more clicks, more context-switching, or more waiting on someone else than a person’s personal ChatGPT habit, they’ll just quietly keep doing it their own way and tell you it’s “going well” when you ask.
This is the same failure mode as Part 1’s ‘individual trap’. Building the system doesn’t kill the instinct that created the problem in the first place. Only making the system genuinely less friction than the workaround does that.
Rollout needs an owner too
Part 1 named a workflow owner, accountable for accuracy and maintenance. Part 2 named a review owner, accountable for what ships. Adoption needs a third kind of owner, accountable for something different again: whether people are actually using the thing.
This role gets skipped constantly, usually because whoever built the workflow gets handed rollout by default. That’s often a mismatch. The person who’s great at prompt engineering and data plumbing isn’t necessarily the person who’s good at getting a skeptical content team to change a habit they’ve had for three years. Different skill, different job, and it’s worth naming separately even if it’s the same person wearing two hats.
A framework: the adoption checklist
Make the new way strictly easier than the old way. Not “better once you learn it.” Easier, immediately, on day one. Any workflow that adds a step before it removes three gets ignored, no matter how good the output is.
Train for real. Don’t just announce it. A Slack message with a link isn’t training. Given only 13% of workers get any formal AI training from their employer, this is exactly where most rollouts quietly die, not in a dramatic failure, just a slow drift back to old habits nobody ever corrected.
Watch usage, not deployment. “We launched it” tells you nothing. Track active usage rate and depth of use — are people doing real work in it, or opening it once to be polite and never again? Deployment is a milestone. Usage is the metric.
Name what happens if adoption stalls. Same discipline as Part 2’s kill-or-scale date, applied to people instead of revenue. If usage hasn’t hit a threshold by a set point, that’s a decision moment to retrain, redesign the workflow to remove friction, or accept it’s not going to stick.
What this looks like in practice
Back to the social-listening-to-content-angles workflow that’s run through all three issues. By Part 2, it had an owner, a shared data source, a documented prompt, a Tier 2 review step, and a quarterly ROI check tied to content conversion rate.
Part 3 adds the piece that actually determines whether any of that mattered: someone ran an actual training session, not a demo — the content team walked through it themselves. Someone’s watching whether people are opening it weekly or once and never again. And there’s a real answer to “what happens if usage is flat next quarter” that isn’t a shrug.
That’s the difference between a workflow that exists and a workflow that’s actually part of how the team works.
Reply and tell me: of the three: build, safety, adoption — which one is your team actually stuck on right now? I’m planning a follow-up issue pulling together what people are running into across all three, so tell me where it’s breaking.
The AI;DR
Elsewhere in the AIverse
Zuckerberg makes the case for AI you can actually own. In a whopping 6,500-word essay, Meta's CEO argues AI should be distributed widely rather than concentrated in a few labs, and that its potential to help people build new things outweighs the job-automation risk. The essay landed alongside Muse Glimmer, an open-weight model small enough to run locally on a Mac or PC.
Grok 4.6 lands close to the frontier, at a fraction of the price. SpaceXAI's latest matches OpenAI's GPT-5.6 Sol on the Artificial Analysis Intelligence Index and trails only Claude Opus 5 and Fable 5 — while holding per-token pricing flat from Grok 4.5, at roughly 60% below what GPT-5.6 Sol and Opus 5 charge at list price.
Microsoft strips Copilot down before building it back up. Consumer and business Copilot are merging into one app, and Group Chats, Podcasts, Deep Research, Copilot Labs, and the Mico mascot announced last October are all being cut by August 18. An internal memo reportedly said Copilot needed to "earn the right to exist." This is the clearest admission yet from a major vendor that feature-stacking an AI product doesn't automatically translate to usage.
Yea, this is the part almost every clean AI demo skips.
The workflow can work perfectly and still die because the marketer has to change five habits to use it.
I have started treating adoption as part of the build: one real job, one owner, one obvious handoff, and a result they can actually verify. Otherwise you built a technically correct orphan lol.