How do communities of practice sustain AI capability?

Formal AI learning ends, but the questions do not. A tool changes, someone finds a useful approach, another person discovers a limitation, and a team encounters a case that the original learning did not cover.

Without a trusted way to share this experience, people solve the same problems separately. Advice can spread informally without evidence, while less-confident colleagues may remain outside the conversation.

A community of practice can turn local experience into continuing learning. It requires more than opening a discussion channel.

AI learning continues between formal events

Practical capability develops through use, evaluation and reflection in changing work. Formal modules can establish foundations and provide safe practice, but they cannot anticipate every future tool, task or professional exception.

Colleagues become an important source of learning. Someone can show how they checked an output, explain why an attempted use failed or ask others to test an emerging method. Different roles may reveal consequences that the original user did not see.

This peer learning is valuable when it remains connected to evidence and professional standards. Otherwise, a confident anecdote can become an unofficial rule, a copied prompt can lose its context, and one successful example can be mistaken for a reliable practice.

Channels and events do not automatically create a community

A chat channel is useful for access and quick questions. Show-and-tell events can make work visible and generate interest. Newsletters can curate changes in tools and policy.

These mechanisms become a community of practice when members repeatedly improve a shared domain of work. The focus shifts from consuming updates to examining practice: what was attempted, under which conditions, what evidence was used, what changed and what remains uncertain.

An unstructured channel may favour frequent or confident contributors. Important questions can disappear in a message stream, and responses may lack ownership. A series of presentations can also remain passive if members do not work on problems together.

The community needs a clear purpose and recurring activities that require participation.

Build the community around shared practice

Define a domain narrow enough to create useful common ground. “AI at work” may be too broad. “Responsible AI-assisted customer operations” or “AI in product discovery” gives members more recognisable decisions and evidence.

Useful recurring activities include:

• Reviewing a fictional, sanitised or approved case from start to decision.
• Comparing how members checked similar outputs.
• Examining a failed attempt and identifying what it teaches.
• Testing a proposed practice against different roles or scenarios.
• Inviting a policy, risk or domain owner to resolve a defined question.
• Curating a reviewed checklist, example or decision note.

A facilitator maintains the purpose, invites different voices and distinguishes open questions from established guidance. A sponsor protects participation time and connects community findings to organisational owners. Members contribute professional experience rather than waiting for a central expert to supply every answer.

The community should capture reusable learning. A short practice note can state the task, conditions, evidence, limitations, owner and review date. This is more useful than a tip detached from its context.

Keep shared learning trustworthy and inclusive

Peer experience is not automatically authoritative. Require sources for important claims and show the status of shared material. A draft idea, tested team practice and approved organisational rule should not look identical.

Questions about security, privacy, legal obligations or professional standards need named escalation routes. The community can identify and frame these questions, but should not invent policy. Confidential work must be sanitised or replaced with safe examples before discussion.

Participation needs deliberate support. Vary meeting formats, use small-group work and provide asynchronous ways to contribute. New learners may ask questions that expose hidden assumptions. Experienced members should show their reasoning and failures, not only polished successes.

Measure whether the community improves practice rather than counting posts. Evidence might include reviewed artefacts adopted by teams, recurring problems resolved, members who apply a method appropriately, or policy questions clarified. These indicators still require cautious interpretation.

Communities also need renewal. Review the domain, membership and useful outputs as work changes. Close or reshape a community that no longer has a live practice purpose. Sustained capability comes from continued relevant participation, not the permanent existence of a channel.

Example

An insurance community meets monthly to examine one fictional or sanitised AI-assisted work example. In one session, members compare how an underwriting and claims team checked summaries generated from similar source packs.

The facilitator asks each team to show the original task, corrections and final decision. A risk representative clarifies a policy question, while an unresolved data issue is assigned to the appropriate owner. The group publishes a short reviewed practice note with its limits and review date.

Useful local experience becomes shared learning, while uncertain advice is escalated rather than repeated as fact.

FAQs

• Is a Teams or Slack channel enough for a community of practice?

  A channel can support access, questions and asynchronous participation, but it needs a clear practice purpose, recurring collaborative activities, facilitation and ownership. Without these, it is more likely to function as a news feed or informal help channel.

• Who should lead an AI community of practice?

  A facilitator can coordinate participation and learning activities, while a sponsor protects time and connects findings to organisational decisions. Members supply practice expertise, and named domain, policy or risk contacts resolve authoritative questions.

• How can a community avoid spreading incorrect AI advice?

  Require sources, preserve context, label the status of shared material and use review owners and dates. Separate personal experience from approved guidance, and escalate legal, security, privacy or professional questions to the appropriate authority.