Part one · The routing failureThe message arrived; the evidence did not

A global AI programme reports that every employee received the launch message. On a night shift, operators discover a recurring exception that creates duplicate customer records. They tell a local champion. The champion posts in a community channel and raises it at a monthly meeting. Product owns the fix, operations owns the consequence, a platform team controls the integration and a budget owner sits elsewhere. No route joins them. Three months later, the programme still celebrates reach while people repair records by hand.

The failure is not communication volume. Information and authority travel through different structures. Local teams see workflow friction, shadow practices, accessibility gaps, unsafe workarounds and customer effects. Executives control capital and priorities. Product, platform, data, risk, workforce and assurance each hold part of the response. A cascade can send a decided message downwards. It cannot reliably carry an ambiguous weak signal upwards, translate it across disciplines, protect its source, find the smallest competent authority and return proof that the condition changed.

A change network is a designed operating network for that journey. It senses local conditions, translates meaning, routes evidence to decision owners, mobilises support and control, preserves challenge, verifies closure and reconnects as the organisation changes. It does not replace line accountability. Its unit is a consequential change signal travelling from observation to qualified decision and verified closure without distortion, orphaning, retaliation or overload.

Four familiar substitutes fail the same test. A champion list gives names without routes or authority. A committee creates a destination but can turn every issue into a queue. An organisation chart shows reporting lines, not whether a protected concern reaches a product or budget decision. A community of practice can exchange knowledge, as Paper 138 develops, but cannot approve capital, accept business consequence or independently assure its own design. Paper 143 owns the missing operating layer: evidence-bearing ties between these structures.

Part two · The operating graphTen node roles surround one outcome

The Change Network Topology is this article's synthesis of ten node types. Nodes are hats with defined mandates, not permanent titles. One person may carry compatible hats in a small firm, but each docket records which hat was used. Delivery and independent assurance cannot be the same hat for the same decision.

Executive outcome sponsors resolve enterprise priority, capital, conflict and material exposure. Business outcome owners own a service, decision population, benefit, harm and operating consequence. Workflow stewards know actual tasks, handoffs, exceptions, load and local control. Boundary translators preserve meaning among domain, product, data, engineering, operations, risk, workforce and affected people.

Product and platform enablers can change tools, integration, access, data, interface, reliability and support. Capability and workforce partners can change learning, staffing, roles, incentives, consultation and transition support. Risk and control partners interpret obligations, design controls, aggregate exposure and challenge unresolved risk. Independent assurance nodes test governance and evidence without owning delivery or the control design they assess.

Worker and local voice nodes surface work reality, professional standards, job quality, accessibility, shadow use and unintended effects. Affected-party and recourse nodes bring customer, citizen, supplier or community feedback, complaint, appeal, remedy and distribution into the operating graph. Representation is not satisfied by one convenient person claiming to speak for a varied population.

Each material tie specifies what may travel, expected response time, destination authority, confidentiality, fallback and closure evidence. A sensing tie carries an observed pattern, not an assigned motive. Translation ties make it legible across disciplines. Enabling ties mobilise resources or implementation help. Authority ties reach someone who can decide. Challenge ties preserve dissent and independent testing. Learning ties return outcome evidence and changed practice. Appearing together on a distribution list proves none of them.

Part two · The operating graphCoverage is a route, not a representative

The Change Coverage Map tests outcomes and decision populations; workflow stages and exceptions; units, entities and geographies; location, shift, channel and employment arrangement; professions and levels; models, platforms and vendors; consequence and recourse; and lifecycle from exploration to retirement.

Putting one employee in every departmental box is not coverage. A real signal from that box needs a safe route to a person able to change the condition and a way to receive a resolution. A contractor on a night shift may be present in a staff directory yet lack access to the reporting system, translator and office-hours authority. A customer complaint may reach service operations but never the model owner. A vendor release may reach product while bypassing workflow stewards whose learned operating boundary is now stale.

Part three · From observation to actionRoute a signal without stripping its meaning

The Signal Routing Protocol is this article's synthesis of six stages: observe, qualify, protect, route, decide and act, then close and learn. A signal is healthy only when it reaches the smallest authority able to act and escalates when correlated or common-mode exposure exceeds that level.

Observe records behaviour, population, context, frequency and consequence without inventing motive. Qualify distinguishes information, a question, support demand, adoption friction, risk, incident, worker concern, affected-party harm or enterprise pattern. Protect applies confidentiality, privacy, anti-retaliation, conflicts and legal privilege where qualified advice says it applies. Protection is not anonymity promised beyond the system's capability.

Route selects translation, enabling, authority and challenge ties with a response deadline and fallback. Decide and act preserves options, evidence, uncertainty, owner, authority, resource, control and expiry. Close and learn verifies that the condition changed, communicates proportionately to the origin and affected population, updates practice and reopens the case when evidence disagrees. Assignment is not closure, and a dashboard is not a route.

Qualify urgency without centralising everything

A mature route distinguishes severity from organisational height. A routine access defect may affect many people yet remain best handled by a staffed platform service. A rare local exception may require a workflow owner, not the executive sponsor. Enterprise escalation becomes necessary when related cases reveal aggregate exposure, a common vendor dependency, cross-entity inconsistency, an obligation beyond local authority or a decision whose consequence the unit cannot absorb. This rule protects senior attention while preventing local fragmentation.

Consider a multinational support service whose regional teams see slightly different failures after the same retrieval update. Each region can correct current customer records and preserve service. Linked version evidence reveals the common source, so the product owner suspends the affected retrieval path globally while regional outcome owners manage remedy. Risk challenges the scope and assurance samples the repair. The route combines local competence with enterprise action without sending every ordinary support request to a central AI office.

Part three · From observation to actionGive every active node a mandate card

A Network Mandate Card records the hat and represented population; outcomes and signals in scope; authority to decide, change, support, challenge, escalate or stop; system access; confidentiality; allocated time and budget; capability; conflicts; response service; succession and expiry.

Mandates turn invisible change labour into operating capacity. A translator expected to assemble cases, maintain relationships and prepare decisions needs time relieved from another workload. A worker representative needs protection and access that do not depend on the sponsor's goodwill. A risk partner needs authority and response time proportionate to the signals received. A successor must be able to continue the route when the trusted individual moves, takes leave or becomes conflicted.

Cards also restrict nodes. A champion may sense and translate but not accept a business outcome or approve a risk exception. A product lead may fix a control but cannot independently assure that fix. An executive sponsor may resolve enterprise trade-offs but should not absorb routine local decisions. Clear limits stop the network becoming a shadow hierarchy.

Part four · Integrity under pressureTest load, succession and independence together

The Network Load and Independence Test examines inbound volume and severity, translation work, decision delay, backlog age, represented boundaries, concentration, allocated time, emotional burden, authority-to-work ratio, delivery-control-assurance conflict, substitutes, self-censorship and repeated unresolved escalation.

The most trusted boundary spanner is often the network's hidden single point of failure. Colleagues send poorly formed issues because this person knows how to translate them. Leaders see a responsive informal route and add no capacity. The translator absorbs preparation, emotional support and follow-up while still doing a full-time job. When that person leaves, the supposed network disappears. Even before departure, overload encourages filtering and delays that no dashboard attributes to network design.

Independence requires more than a second name. Dense ties can support fast coordination and still reproduce one assumption. Product and risk may share incentives, data and leadership pressure. Assurance cannot own the design it later assesses. For a small business, independence may come from a qualified external reviewer rather than an elaborate internal structure. Compatible hats can be combined, but delivery and independent judgment for the same material decision remain separated and recorded.

Part four · Integrity under pressureGive each decision a docket and a return path

The Change Decision Docket provides organisational memory without exposing private worker data beyond legitimate need. It connects the originating signal, protected source treatment, affected outcome and workflow, evidence and rivals, route and translations, accountable owner, cross-functional input, options, decision, resources, controls, communication, expiry and closure evidence.

The docket relates signals without prematurely merging them. A small problem in six units may point to one vendor change. Local owners can resolve individual effects, while linked docket identifiers allow risk or product nodes to see aggregate exposure and escalate it. The protected source need not travel with the pattern. Conversely, a common symptom may have distinct causes; the rival-explanation field prevents a central team from imposing one diagnosis.

Unresolved dissent stays visible. A decision can proceed within authority while assurance records why evidence remains incomplete and what would reopen the case. The originator receives a proportionate rationale and the evidence used to declare closure. If the frontline team sees no change, the docket reopens. The action owner cannot mark “done” merely because a ticket has been assigned.

The docket also prevents two symmetrical mistakes. An executive should not treat one vivid report as an enterprise pattern without checking related populations, versions and rival causes. Nor should leaders dismiss a repeated mechanism because each unit's numbers appear small. The relationship among cases carries its own evidence: common dependency, timing, error signature and affected group. Pattern qualification names both what is shared and what remains local, then assigns an owner capable of changing the shared cause.

Part four · Integrity under pressureRead seven health views without a maturity score

The Network Health Panel keeps coverage, reach, latency, closure, load, independence and voice, and learning as separate views. A dense, fast network can still be unsafe if it suppresses dissent or closes work without changing conditions.

Coverage exposes populations and workflows with no safe route. Reach tests which owners can be contacted in practice. Latency separates time to qualify, decide, act and verify. Closure distinguishes resolved, partly resolved, rejected, expired and orphaned signals by consequence. Load shows backlog, concentration, hidden work, burnout and succession risk. Independence and voice show conflicts, self-censorship, retaliation concerns and access to protected routes. Learning shows recurrence, changed standards, shared cases and topology changes.

The Topology Reconfiguration Rule applies when outcomes, workflows, organisation, models, vendors, regulation, workforce, affected populations, risk, geography or volume changes materially. Add, remove or reconnect nodes based on observed routes, not prestige or historical committee membership. A new vendor may require a direct product-change tie; expansion across time zones may require regional authority or a staffed fallback; repeated local dockets may justify a new enterprise node.

NIST's current AI RMF materials support differentiated responsibilities, relevant actor participation, affected-party feedback, monitoring and lifecycle decisions; they do not prescribe this topology. The UK AI Playbook supplies senior responsibility, multidisciplinary work, auditability and recourse context. ILO cases show worker representation influencing AI strategy and working conditions across varied institutions. OECD workplace evidence remains survey- and country-bound. ISO/IEC 42001 is a management-system standard for organisational processes and continual improvement, not proof that one network route works.

Network and implementation research contributes bounded mechanisms. Granovetter describes bridging information across social groups; Centola's controlled online health experiment shows that some behaviours need clustered reinforcement; Valente stresses that intervention choice depends on behaviour, data and context. Tushman and Scanlan study boundary spanning and information transfer. Klein and Sorra, the updated CFIR, May and Finch, Oreg and colleagues, Edmondson, and Weick and colleagues illuminate implementation climate, context, recipient reaction, routine work, speaking up and sensemaking. Their settings and constructs sharpen questions, not a ready-made enterprise AI map.

Part five · Put the network under loadTen cases that force different repairs

Each case traces the topology, coverage and six ties; applies mandate, load and independence tests; opens a docket; and names line, enabling, risk and assurance ownership.

Part five · Put the network under loadAI Change-Network Designer

The designer evaluates one consequential signal route. Set current evidence or load a named scenario. It returns exactly one of five topology decisions and exposes the uncovered population, tie contract, protected route, mandate gaps, load, independence, docket, closure path and reconfiguration trigger.

The instrument cannot prove social influence, infer coverage from membership or turn message volume into performance. Its output is a testable operating change with a responsible owner, a protected source treatment and an expiry. Run an actual signal through the proposed route before relying on the design.

The decision changedBuild for the signal that must come back

An AI transformation network succeeds when consequential evidence travels from lived work to the smallest competent authority and returns as a verified change. Everything else is supporting structure.

Begin with an outcome and a signal that matters. Draw the ten hats and only the six ties needed to move it. Test coverage with real populations, shifts, exceptions, technologies, consequences and recourse. Give each node a mandate, access, protection, time, budget and successor. Keep delivery, control and assurance distinct. Trace the protocol, open the docket, preserve dissent and refuse to close until the originating condition has changed.

This design changes the executive question. “How many champions do we have?” becomes “Which consequential signal still cannot reach authority?” “How many employees saw the message?” becomes “Which route returned closure evidence?” “Who is influential?” becomes “Which mandate has the access, capacity and legitimate authority to act?” Those questions place responsibility with the organisation that designed the change rather than the worker asked to volunteer around its gaps.

A network should stay smaller than its ambitions. Informal relationships can remain voluntary. Local decisions should remain local when consequence permits. Enterprise patterns should escalate before they become common-mode failure. Independent challenge should be structurally possible, affected people should have recourse, and the graph should change when work changes. The test is not whether information moved. It is whether a real condition changed and the evidence found its way home.