A month after we trained our own people on AI, fewer of them were using the developer tool we taught. By seat count, the training failed. By output, it worked. Accepted code changes rose by half, and each one cost about a quarter less.
Most organizations track the first number. Few track the second. If you have already bought AI seats, that difference decides whether they pay off.
We have seen this before. For years, we have helped organizations rescue Jira and Confluence rollouts that were bought, provisioned, and never adopted. AI is repeating the pattern, only faster.
We deploy AI internally before we recommend it to clients, so we measured our own rollout.
People with a license and no training built things they never used. We audited the shared Claude Projects (workspaces for recurring work) across the firm. Most had no instructions or reference material. About 40 percent had never been used at all.
People with training worked differently. After a two-day immersion in late August, attendees' days working with AI grew roughly twice as fast as those of colleagues with the same license.
The blank page. Without a model for how AI fits into real work, people treat it as a smarter search engine. They ask questions instead of delegating work. The value of an AI assistant comes from handing it a meaningful piece of a task, and that is a learned skill.
Tool confusion. Modern AI platforms offer chats, projects, reusable skills, connections to business systems, and agents that complete work on their own. Each fits a different kind of task. Untrained users default to whatever they found first, and the wrong choice produces work that is stranded, duplicated, or abandoned.
Clutter instead of compounding. When everyone improvises, nothing becomes repeatable. Good prompts live in one person's chat history. Useful workspaces go stale. The organization pays for the same discovery over and over.
Effective AI training does not teach buttons. Vendors already publish excellent material on features. What people lack is a working method: how to start a piece of work with AI, how to give it the context it needs, when to check its output, and how to turn a one-time win into something the whole team reuses.
Training also shows people which tool fits their role. Our user count dropped because some attendees learned that developer tools didn't fit their jobs and moved to tools that did. A dashboard would score that as a loss. It was the training doing its job.
A small group doing real work with AI is worth more than a large group that logged in once. Headcount can't tell the difference.
Large enterprises have IT departments, change management teams, and internal champions to absorb a rollout. Small and mid-sized businesses usually do not. The person who signs the agreement often configures the tool, sets the policies, fields the questions, and still runs their part of the business.
For many small businesses, the license purchase is the entire plan. Every unused seat is a larger share of the budget. Every stalled pilot is harder to restart.
That is why we built Claude Jump Start through Anthropic's Claude for Small Business program. We start with the client's actual work, not the tool. Together, we configure Claude around their policies, build the first workflows their team will use every week, and teach a method people carry to the next task. Thirty days later, we sit down together and measure what changed.
If the answers are uncertain, more licenses won't fix it. Enablement will.
If you’ve bought AI seats and aren’t sure they’re paying off, we’d love to talk. Our Claude Jump Start program, built through Anthropic’s Claude for Small Business program, starts with your team’s actual work, builds the first workflows they’ll use every week, and measures what changed 30 days in.
This blog was written by Brian Nye, VP of Strategic Initiatives at Praecipio.