How do we drive adoption of an AI coding tool from a third of engineering to everyone?

When AI coding tool adoption stalls at a third, that third usually adopted on their own and would have found the tool anyway. Many of the other two-thirds tried it, got a wrong answer about your codebase or lost an afternoon to setup, and went back to how they worked before. Mandates and usage leaderboards tend to backfire. The steps that work are plain. Find out who stopped and why. Push the setup out centrally so nobody configures anything. Give the tool real context about your code so the first answers are right. Let the engineers who already use it teach the rest on real tickets. Then judge the rollout by delivery outcomes such as lead time and review time.

Why AI coding tool adoption stalls at a third#

The tool is rarely the problem. Stack Overflow's 2026 survey found that 73 percent of developers who use AI coding assistants or agents use them daily. Claude Code (66 percent) and GitHub Copilot (59 percent) lead the agents (Stack Overflow, October 2026). People who get value keep using it. The same survey found that 93 percent of developers need source attribution before they trust what the AI gives them.

That trust gap is where the other two-thirds drop off. Atlassian's 2026 developer experience report found that 65 percent of developers "now regularly spend time correcting AI output". Finding information was again the top source of friction, ahead of meetings and context switching (Atlassian, October 2026). For example, an engineer asks about the billing service and gets a confident answer that ignores last quarter's migration. They have learned not to ask again. That engineer ran a fair test, and the tool failed it.

Find out who stopped, and ask them why#

A quiet user is an engineer who was active for a week or two and then stopped. They are the people to talk to, and your usage data already lists them. The Claude Code analytics dashboard shows daily active users and sessions, and Anthropic's docs suggest watching for "dips in usage that may indicate friction or issues" (Claude Code docs). GitHub sorts Copilot users into chat-first, code-first, agent-first and multi-agent cohorts. The cohorts exist "so you can see whether your organization's usage is deepening over time rather than plateauing at initial trial" (GitHub docs).

Talk to five quiet users. Ask what they tried first and what went wrong. Most answers fall into a few groups, and each has its own fix.

Remove setup and make the first answer right#

Managed settings are configuration an organization deploys that a developer's own files cannot override. Claude Code reads them from the claude.ai admin console, an MDM policy or a system file, and "nothing in your own settings files or --settings overrides a managed key" (Claude Code settings docs). A platform team can ship the same permissions, MCP servers and defaults to every laptop on day one. The pattern across several agents is in how to get context into coding agents consistently across the org.

Committed config matters for quality too. A study of 441 repositories found that agent-first repos without committed AI configuration showed about twice the rise in cognitive complexity, +53 percent against +27 percent (Denisov-Blanch et al., 2026). The authors call this observational and hypothesis-generating. Rules files cover repo-local conventions, though agents often skip them.

Most of what a skeptic needs lives outside the repo, such as why the service is built this way or what the team decided in Slack last month. A context engine is a service that reasons across code, PRs, issues, docs, Slack and incidents and returns one cited answer. Unblocked is one, delivered inside Claude Code, Cursor, Copilot, VS Code and JetBrains, and filtered to what the asking engineer may see (how the Unblocked context engine works). A skeptic who can click through to the PR behind an answer has a reason to ask a second question.

Spread working habits through people, not mandates#

Adoption moves through teammates. A study of Microsoft's early-2026 rollout of command-line coding agents found that "first use spread primarily through social networks" (Murphy-Hill et al., 2026). Pressure backfires. Atlassian found that 60 percent of developers say AI usage is treated as a productivity signal, "and 41% sometimes use it more than necessary because they feel expected to" (Atlassian, 2026). Only 17 percent say their organization has "a clearly communicated and implemented AI strategy".

So write a short plan first: what the tool is for on your team, what it is not for, and what review its output needs. Then use the people who already rely on it. Anthropic describes the Claude Code leaderboard as a way to find users who can "share prompting techniques and workflows with the team" (Claude Code docs, linked above). Ask each one to pair with two teammates on a real ticket from the current sprint. A demo on a toy repo proves nothing to a skeptic. A fix to their own flaky test does.

People move at different speeds. Some stay on autocomplete for months while others run agents in parallel. The eight levels of agentic engineering lay out that ladder. Meet each engineer one step above where they are.

What to measure once people are using it#

Atlassian's advice to leaders is to "measure the resulting outcomes, not activity" (Atlassian, linked above). Daily active users tell you the rollout reached people. Three numbers tell you it worked:

• Lead time for changes, end to end.
• Review pickup time on AI-assisted pull requests.
• The share of AI-assisted changes reworked within two weeks.

This matters more as adoption grows, because faster coding by more people piles up in review unless the rest of delivery keeps pace. That is the subject of why AI coding gains disappear at scale, and the metrics are covered in how to measure AI productivity. Recheck the quiet-user list every month. When it stops growing, and the reason people give shifts from "it was wrong" to "I haven't had time yet", the rollout is working.