The single thing that determines whether an AI rollout sticks is not the technology. It's whether you can answer this question clearly: who is accountable for making sure it gets used?
Not who selected the tool. Not who signed the contract. Who, specifically, is checking usage numbers every week, running feedback sessions, adjusting what's not working, and escalating when the initiative is losing momentum? If that person doesn't exist — or if it's someone who's also doing a full-time job on top of the initiative — the rollout will stall.
This sounds obvious. It is not widely practiced.
After running AI implementations at organizations like ADP, Anthem, and Humana — and now at mid-market companies that are facing the same challenges with fewer resources — the pattern is consistent: the rollouts that stick have clear ownership. The ones that stall don't.
Why mid-market AI rollouts stall
Here are the five specific failure patterns we see most often. They're not about the technology.
1. No single owner
The initiative has sponsors (usually leadership) and implementers (usually IT or operations), but nobody whose job is to drive adoption. The sponsors check in quarterly. IT considers their job done after the technical setup. Nobody owns the gap between "tool is live" and "people are using it consistently."
On a large enterprise rollout, this role is called an implementation PM. At mid-market companies, it often gets assigned to whoever has the most bandwidth — which usually means whoever has the least, because bandwidth is correlated with organizational priority.
2. Generic training with no role-specific use cases
A 45-minute "here's what AI can do" session is not training. It doesn't give anyone a specific, actionable thing to do differently tomorrow. A project manager and a billing analyst need completely different use cases. Rolling out the same session to both produces the same result: polite attention, followed by inaction.
The use cases that drive sustained adoption are usually unglamorous: first drafts, meeting summaries, data formatting, report templates. They're not the ones in the vendor's marketing material. They're the ones where the tool saves 30-90 minutes a week on something the employee already finds tedious.
3. Rolling out to everyone at once
Enterprise implementations almost always phase rollouts: pilot with 10-20 users, iterate, then expand. Mid-market companies often skip the pilot because it feels like extra work, or because there's pressure to show ROI quickly across the whole organization.
The cost of skipping the pilot is that you discover what's broken — the integration issues, the use cases that don't work as expected, the teams that need more support — after you've already made the tool available to everyone. Fixing problems at scale is harder than fixing them in a controlled pilot.
4. No success definition before launch
"We want people to use AI more" is not a success metric. What does 90-day adoption look like? What's the target usage rate? Which workflows should change, and by how much? Without specific, measurable targets defined before launch, you have no way to know whether the rollout is working — and no lever to pull when it isn't.
5. Tool doesn't integrate with daily work
If using the AI tool requires switching contexts — opening a separate interface, copying content back and forth — it won't become a habit. Habits form in the flow of existing work. If the tool lives outside that flow, it gets used occasionally by enthusiasts and ignored by everyone else.
What enterprise implementations do differently
The discipline that makes large-scale rollouts work is directly applicable to mid-market companies. The tools are simpler. The process doesn't have to be.
At ADP or Humana, a rollout of this scope would have: a dedicated implementation PM who owns nothing else, a scope document agreed on before anything is built, a phased rollout plan, role-specific training built with input from the people who do the work, and adoption metrics reviewed in a weekly meeting.
None of this is exotic. Most of it is common sense. But it requires someone whose job is to enforce it — and that's the thing mid-market companies typically don't have.
The 90-day framework that works
Here's a framework for a 90-day AI rollout that holds up in practice. This assumes you've already selected a tool and have technical implementation underway.
Days 1–30: Discovery and pilot design
Identify 2-3 roles with the highest potential for meaningful AI adoption based on where they spend the most manual time. For each role, define 3 specific use cases — tasks the AI tool can meaningfully accelerate. Identify 10-20 people across those roles for a pilot. Define what success looks like at the end of the pilot: usage rate, specific workflows that should change, measurable time savings.
Also during this phase: get baseline data. Most enterprise AI tools have admin dashboards that show usage. Take a snapshot before the pilot starts. You'll need it.
Days 31–60: Pilot, iterate
Run the pilot. Conduct role-specific training sessions (90 minutes, live, with real examples from the participants' actual work). Share prompt libraries and quick-reference guides in the systems the team already uses. Check usage numbers weekly. Run a 2-week and 4-week feedback session with pilot participants. Find out what's working, what isn't, what questions keep coming up, and what's getting in the way. Adjust.
A well-run 30-day pilot answers two questions: What's actually working for these roles? And what would we do differently if we ran this again?
Days 61–90: Expand and track
Apply what you learned in the pilot to the broader rollout. Expand to additional roles and headcount. Continue tracking usage weekly. Run a 30-day check-in with expanded users. By day 90, you should have measurable adoption data — usage rates, specific workflows that have changed, early productivity signals — that you can present to leadership.
Common mistakes with numbers
A few specific things we see go wrong — with estimates of how much they cost:
Skipping the pilot: Rolling out to 200 people before validating use cases with 20 often results in a second, corrective rollout 3-6 months later. Budget 40-60% of the original rollout cost for the corrective effort.
One-time training only: Adoption from a single training session typically peaks in week 2 and reverts to baseline by week 6. A 4-6 week follow-up cadence sustains adoption through habit formation. The cadence doesn't have to be intensive — weekly usage reviews and biweekly office hours are usually enough.
Not tracking usage: Without tracking, you're managing by anecdote. Executives report "people love it" while the admin dashboard shows 18% active usage. The number is almost always worse than people think, which is why tracking matters — it creates urgency and accountability.
Before you launch: a quick checklist
- Is there a single named owner of adoption outcomes?
- Have you defined role-specific use cases for the 3 highest-priority roles?
- Do you have a pilot group identified (10-20 people)?
- Is training built around their actual work, not vendor demo content?
- Do you have a 90-day adoption target defined in measurable terms?
- Is there a weekly cadence for reviewing usage data?
- Does the tool integrate into existing workflows, or require context-switching?
If you can answer yes to all of these, you're better positioned than most mid-market AI rollouts we see. If you're not sure, start with the first one — ownership — and work outward from there.
If your rollout has already stalled: the most common recoverable cause is the use case problem — generic training with no role specificity. The fix is targeted: pick the two roles with the highest potential, define 3 specific use cases each, run 90-minute hands-on sessions, and track for 6 weeks. We've seen this pattern turn adoption from <20% to 58%+ in that timeframe.