01 · Popular is not portableThe prompt everyone copied

A global AI community channel is busy. The most-liked post contains a finance prompt that appears to save time, and practitioners in several regions copy it. The originating team reports quarterly management accounts. Another team uses the prompt for a statutory period with different definitions. The text is fluent, the calculations look plausible and the reporting period is wrong.

The channel transmitted an artefact but not the practice that made it usable. It omitted the work event, source definition, decision consequence, data boundary, model version, human authority, verification step and conditions where the prompt must not travel. Popularity increased reach without increasing evidence. A useful local shortcut became an enterprise error because context was treated as packaging rather than part of the knowledge.

A community of practice is a group learning through sustained engagement in a shared domain and practice. It is not merely a membership list, job family, repository or forum. An intelligence network adds a more demanding enterprise function: local experience becomes a structured signal; peers and specialists test it; a versioned registry preserves its status; formal owners receive evidence they can act on; and stale or harmful guidance is retired.

02 · Seven connected capabilitiesThe Practice Intelligence Network

The Practice Intelligence Network is this article's first synthesis. Local practice nodes sit close to recurring work, exceptions, users and affected parties. They notice weak signals that central teams rarely see early: an interface workaround, a population-specific failure, an ambiguous policy, an overloaded review step or an integration that changes how people exercise judgement.

Boundary stewards connect domains, product, platform, data, risk and policy. Their job is translation and routing, not approval. Signal intake captures cases, failures, near misses, evaluations, questions and reusable artefacts. Evidence review tests the source, outcome, consequence and transfer boundary. A practice registry stores versioned cases, tests, runbooks, owners, status and expiry. Promotion and retirement paths change the artefact's state. Owner feedback routes send repeated evidence to people with authority to change products, platforms, policies, contracts, workforce design or investment.

The network separates five things that organisations often collapse. The community is the relationship through which practitioners learn. Practice is the situated work and judgement they share. The network connects different knowledge pools. The registry makes reviewed artefacts discoverable and versioned. Formal authority remains with named owners. A steward can carry evidence to a policy owner but cannot quietly turn community consensus into policy.

Value appears when these elements complete a circuit. A local node detects a recurring citation failure. A structured signal allows another node to reproduce it. Review establishes a bounded pattern. The registry makes the pattern visible with exclusions and expiry. The product owner changes provenance display. The community receives the disposition and tests the new behaviour. Without the return path, contribution becomes extraction and the same signal recurs.

• Local practice nodes: situated detection and help close to work.
• Boundary stewards: translation and bridging without shadow authority.
• Signal intake: structured cases, failures, questions and artefacts.
• Evidence review: challenge of mechanism, outcome and limits.
• Practice registry: searchable state, version, ownership and expiry.
• Promotion and retirement: explicit movement, restriction or deprecation.
• Owner feedback routes: product, policy, control and investment decisions.

03 · A signal must survive inspectionThe Practice Signal Protocol

The Practice Signal Protocol is this article's second synthesis. It starts with the work event, decision and desired outcome; role, domain, users and affected population; AI system, version, data, context, tool and authority. It records the observed condition, trigger or failure, the evidence and uncertainty, the intervention and accountable decision-maker, and the downstream result or unresolved consequence.

The second half protects transfer. It states boundary conditions, contexts where the lesson should not travel, a reusable artefact or test, and confidentiality, privacy, security and legal classification. It names the reviewer, challenger, formal owner and next route. Freshness triggers, expiry and a superseding record prevent an old prompt or runbook from appearing current after the system changes.

Minimum structure does not mean every chat question needs a form. Conversation remains open and fast. The protocol becomes mandatory when a claim is promoted, reused across a boundary, escalated into formal change or applied to controlled work. That threshold preserves voluntary help while placing structure around consequential knowledge.

04 · Evidence changes the artefact's stateThe Practice Evidence Ladder

The Practice Evidence Ladder is this article's third synthesis. An anecdote is a potentially useful observation with too little context or evidence for reuse. A structured case completes enough of the protocol for peers to inspect the reasoning. A provisional pattern links similar cases into a possible mechanism and records counterexamples. A validated practice survives representative tests or repeated operating evidence in its named context.

A supported practice has a product, platform, policy or professional owner who maintains the artefact, controls, training and support. A reference pattern has evidence for adaptation across more than one context, along with conformance tests and known exclusions. Deprecated or prohibited is not the bottom of a maturity ranking. It is an active safety state that warns users, names the reason and redirects them to a current alternative.

Movement is not only upward. A pattern may remain local because transfer value is low. A legal difference may send it sideways into adaptation. A model update may return a supported practice to provisional review. A severe incident can prohibit an artefact immediately. Popularity, senior sponsorship or vendor marketing never advances the state by itself.

05 · Carry the mechanism, not the surfaceThe Context-Preserving Transfer Test

The Context-Preserving Transfer Test is this article's fourth synthesis. It asks whether the target work event is materially the same; consequence, population and authority are equivalent; data meaning, provenance, access and freshness are comparable; and model, prompt, tool and integration remain inside the tested boundary. It then tests whether the target team has the assumed skill, review capacity and recovery path.

The receiving team must be able to reproduce the evidence and observe the outcome. It must identify which local constraint requires adaptation, and it must carry counterexamples, excluded cases and stop conditions with the pattern. A claims exception taxonomy may transfer conceptually across regions while the legal decision authority does not. The target can adapt the taxonomy and rerun evidence; it cannot inherit permission from the source team.

Transfer failure is informative. If a pattern works only with one expert, the missing mechanism may be expertise location. If it fails when the interface changes, the practice may have depended on a hidden affordance. If it creates review overload, capacity was part of the original intervention. These are not reasons to flatten the pattern further. They are reasons to make its dependency visible.

06 · Trust needs range without a queueThe Network Topology Rule

The Network Topology Rule is this article's fifth synthesis. Dense local ties support trust, situated learning and fast help inside a practice. Bridging ties provide access to different knowledge pools. Boundary stewards translate language and assumptions across those ties. Formal owners receive signals that require product, policy, risk, funding or contractual action. The registry supports discovery and memory, but does not replace trusted relationships.

A single steward is a fragile bridge. Every question queues behind one person's calendar, translation becomes unchallengeable and the steward quietly accumulates approval power. The repair is not a larger central expert team by default. Add multiple bridges, make protocols and owner interfaces legible, build expertise location into the registry and let routine exchanges move directly between nodes. Stewards focus on difficult boundaries and disputed meaning.

Network health also depends on who is absent. Independent voices, frontline workers, smaller locations, contractors, non-native speakers, disabled practitioners and affected-party perspectives may see different failure conditions. Dense ties among senior enthusiasts can create trust and still form an echo chamber. Range matters only when different evidence can influence state and formal decisions.

07 · Participation has operating conditionsThe Participation and Protection Contract

The Participation and Protection Contract is this article's sixth synthesis. Contribution time is recognised and managers support it. Voluntary peer learning remains voluntary, while required duties such as incident escalation or controlled training are explicit. Acceptable data, redaction and confidentiality rules are clear. Contributors retain attribution and know how their artefacts may be reused.

Respectful challenge and anti-retaliation protect dissent without turning consensus into evidence. Customer, employee, security-sensitive or legally privileged material stays inside approved routes. A redacted lesson may enter the community while the restricted record remains in the incident, security or legal system. Community tools never become a convenient way around access control.

Access is designed across disability, language, time zone, location, shift and worker status. A global meeting in one language during office hours is not global participation. Asynchronous cases, translation, accessible formats and paid contribution time alter who can supply and challenge evidence. Moderation and appeal prevent status capture. Vendor participation includes conflict disclosure and limits on promotion.

Formal decision rights remain outside community popularity. The contract names ownership, service time and expected response for escalated signals. When frontline knowledge exposes a platform defect, contributors receive attribution and time to support reproduction, while the platform owner must accept, reject or schedule the change with reasons. This closes the extraction loop.

08 · Measure movement and changeThe Practice Intelligence Scoreboard

The Practice Intelligence Scoreboard is this article's seventh synthesis. It is a balanced evidence set rather than one score. It observes time from signal to reviewed disposition; repeated problems and duplicate solutions; the share of registry artefacts with owner, evidence, boundary and expiry; and reproduction or conformance success in another context.

It also tracks use that produces verified workflow or outcome improvement, signals converted into product, platform, policy or control change, and stale or contradicted artefacts found before harm. Contributor diversity reveals whether evidence comes from the whole practice. Expert dependency and steward load expose topology weakness. Unresolved requests show whether the network is accepting more signals than it can review.

Posts, members, meetings, downloads and prompt submissions may describe activity, but they do not establish value without an evidence path. The scoreboard should be read as a set of tensions. Faster disposition can be harmful if review becomes superficial. More cross-context reuse can be harmful if exclusions disappear. Higher contribution can coexist with inequitable unpaid work. Each signal needs a counter-signal that exposes distortion.

09 · Ten worked casesRoute, review, transfer, own or retire

Each case maps the seven network components, completes the signal protocol, assigns an evidence state, runs the transfer test, applies protection, selects scoreboard evidence and names the formal owner. Some signals remain local; some become supported practice; some require restricted escalation or retirement.

10 · Executable instrumentPractice Intelligence Network Designer

The designer is this article's eighth synthesis. It tests the practice, signal completeness, evidence state, transfer boundary, topology, confidentiality, participation, formal routes and freshness. It returns one of five actions, never a universal community score. Popularity, membership and post volume cannot promote a practice.

11 · Strong rival and failure boundaryKeep conversation light and consequential knowledge governed

The strongest rival argues that communities of practice are valuable precisely because they are informal. Schemas, review gates, metrics and owners can destroy voluntary participation, slow help and turn social learning into knowledge-management bureaucracy. Communities can also entrench orthodoxy, exclude newcomers, leak sensitive information and compete with formal product, policy and support channels.

The answer is minimum viable structure around consequential knowledge. Everyday conversation stays relational and fast. The protocol begins when someone wants to promote, reuse, escalate or apply a claim to controlled work. Local groups remain local when transfer value is low. Stewards connect rather than approve. Formal owners decide. Sensitive evidence follows approved routes. The registry remembers status and expiry but never substitutes for people practising together.

Close or merge a community that has no live practice, active practitioners, reviewed artefacts or owner feedback path. Restrict a signal when confidentiality cannot be protected. Keep a pattern local when the transfer test fails. Deprecate an artefact when version, policy, evidence or harm invalidates it. Escalate when the required change sits with product, platform, policy, risk or funding authority. Do not allow participation metrics, vendor status or seniority to override these boundaries.

The changed leadership decision is simple: stop asking how many people joined the AI community. Ask which situated signal crossed a boundary with its context intact, changed a formal system and remained safe to find. That is when a community becomes an intelligence network rather than another channel.

Evidence and limitsSources checked 3 September 2026

Wenger's book is foundational social-learning theory, not an AI-era effectiveness study. Wenger-Trayner guidance is author-practitioner material. Brown and Duguid, Carlile, Argote and Ingram, Reagans and McEvily, Moreland and Myaskovsky, and Edmondson provide bounded organisational, network, experimental and team evidence; their populations and constructs do not establish universal community outcomes. UK, NIST, OECD and ILO sources provide current capability, lifecycle and worker context. NBER field studies concern particular tasks and settings.

The Practice Intelligence Network, Signal Protocol, Evidence Ladder, Context-Preserving Transfer Test, Network Topology Rule, Participation and Protection Contract, Scoreboard and interactive designer are this article's synthesis. Their service times, states, transfer gates and protections require local validation.

1. Wenger, Communities of Practice: Learning, Meaning, and Identity.
2. Wenger-Trayner and Wenger-Trayner, Introduction to Communities of Practice.
3. Brown and Duguid, Organizational Learning and Communities-of-Practice.
4. Carlile, A Pragmatic View of Knowledge and Boundaries.
5. Argote and Ingram, Knowledge Transfer.
6. Reagans and McEvily, Network Structure and Knowledge Transfer.
7. Moreland and Myaskovsky, Exploring the Performance Benefits of Group Training.
8. Edmondson, Psychological Safety and Learning Behavior in Work Teams.
9. UK Government, AI Skills Tools Package.
10. UK Government, Skills for AI.
11. UK Government, Artificial Intelligence Playbook.
12. NIST, Artificial Intelligence Risk Management Framework.
13. NIST AIRC, AI RMF Core.
14. NIST AIRC, AI RMF Playbook.
15. OECD, Artificial Intelligence and the Changing Demand for Skills.
16. ILO, Social Dialogue Critical to Ensure Positive Outcomes of AI in the Workplace.
17. Dell'Acqua et al., The Cybernetic Teammate.
18. Brynjolfsson, Li and Raymond, Generative AI at Work.