You bought the seats. You ran the kickoff. Six weeks later, the honest summary from your team is some version of: it's fine, I guess, but it's not really saving me time.

That's a frustrating place to be, because the obvious next move — switch tools — almost never fixes it. Teams that swap ChatGPT for Copilot, or Copilot for Claude, usually land in exactly the same spot two months later.

The tool is rarely the variable. What varies is whether the people using it have built the specific skill the tool requires. Here are the five diagnoses we see most often on small teams, how to tell which one is yours, and what actually changes the outcome.

Diagnosis 1: The Output Is Mediocre Because the Input Is Vague

Symptom: People say the answers are "generic," "obvious," or "something I'd have to rewrite anyway."

Nine times out of ten, the prompt was one sentence with no context, no audience, no examples, and no constraints. The model returned exactly the average of everything it had ever read, because that's the only thing an averaged question can produce.

This isn't a knowledge gap about AI. It's a skill gap in a very narrow, very learnable thing: how to give a model enough context to be useful.

The fix: Teach one pattern — task, context, audience, constraints, example of good — and have each person apply it to the three tasks they do most. Skill shows up within a day. This is the highest-return hour of learning available to most teams right now.

Diagnosis 2: They're Using It for the Wrong Tasks

Symptom: People tried it on their hardest, most judgment-heavy work, got burned, and concluded the tool is unreliable.

AI is strongest on first drafts, summarization, restructuring, comparison, and volume work. It's weakest exactly where your team's expertise lives — the judgment calls that make them worth employing. If the first thing someone tried was the judgment call, the trial failed for a good reason.

The fix: Map the tasks, not the tools. For each role, name the three tasks where a fast, imperfect first pass is genuinely valuable. Start there. Nobody's opinion of AI is formed by a feature list; it's formed by their first three attempts.

Diagnosis 3: Nobody Knows What "Good" Looks Like

Symptom: Individuals can't tell whether their own AI output is strong or weak, so they either over-trust it or discard all of it.

Without a reference point, people can't self-correct. The power user on your team has an internal bar for what a good result looks like and iterates until they hit it. Everyone else accepts the first response and judges the tool by it.

The fix: Make good work visible. A shared library of real prompts and real outputs from your own team, labeled by role, does more than any course. Skill spreads laterally on small teams faster than it spreads top-down.

Diagnosis 4: It Never Made It Into the Workflow

Symptom: Usage spikes the week after training and decays to near-zero by week three.

If using the tool requires opening a separate tab, remembering it exists, and deciding it's worth trying, it will lose to the habit that's already there. This is the most common failure mode and the one least related to the software.

The fix: Tie the skill to a recurring moment, not to a general intention. "Before every discovery call" beats "when it makes sense." One anchored habit per role, then add the second only once the first is automatic.

Diagnosis 5: People Are Quietly Unsure What's Allowed

Symptom: Adoption is uneven in a specific pattern — client-facing and regulated roles use it least.

When nobody has said out loud what data is safe to put in a prompt, careful employees resolve the ambiguity by not using the tool. From the outside this reads as resistance. It's caution, and it's the correct instinct in the absence of guidance.

The fix: Write down, in one page, what can and can't go into a prompt, and what has to be verified before it leaves the building. Teams with a clear rule use AI more, not less, because the anxiety cost drops to zero.

How to Tell Which One You Have

Ask three people in different roles the same two questions:

  1. Show me the last thing you used AI for. (If they can't produce one, it's Diagnosis 4.)
  2. Show me the actual prompt. (If it's one vague line, it's Diagnosis 1. If it's a judgment call they should have made themselves, it's Diagnosis 2.)

Then ask whether they know what data they're allowed to paste in. Hesitation is Diagnosis 5.

That's a 20-minute exercise, and it will tell you more than any usage dashboard, because usage counts don't distinguish between a team that's learning and a team that's clicking.

The Underlying Point

Access is not fluency, and neither is enthusiasm. The teams where AI visibly pays off aren't the ones that picked the best tool — they're the ones that treated it as a skill to be built role by role, with a starting point, a standard for good, and a moment in the week where it actually gets used.

That's learnable. Most of it is learnable in a few focused hours per person.

If you want a starting point that isn't guesswork, run a skill assessment first and let the results tell you which of the five diagnoses you're actually dealing with — the answer is often different for each role on the same team.

Run a free AI skill assessment for your team → results in under 15 minutes.

Related reading: Why Your Employees Aren't Learning AI · The Learning Gap Between Your Best and Worst AI User · How to Use AI at Work: A Small Business Guide · Pricing · Explore free courses

LinkedIn repurpose (founder account)

"The AI tools aren't working for my team."

I hear this a lot. And almost every time, the next move people are considering is: switch tools.

It's almost never the tool.

Five things it usually is instead:

  1. The prompts are one vague line, so the output is the average of everything ever written.
  2. They tried it first on their hardest judgment call — the one place AI is weakest — and concluded it's unreliable.
  3. Nobody knows what a good result looks like, so they accept the first response and judge the tool by it.
  4. It never got anchored to a recurring moment in the week, so it lost to the habit that was already there.
  5. Nobody told them what data is safe to paste in, so the careful people quietly opted out.

Here's the 20-minute diagnostic: ask three people in different roles to show you the last thing they used AI for, and the actual prompt they used.

If they can't produce one, it's #4. If the prompt is one vague line, it's #1. If they hesitate about what they're allowed to paste in, it's #5.

Usage dashboards can't tell the difference between a team that's learning and a team that's clicking. Two questions can.