The refund that began as a question
A customer-support manager notices an unusual pattern: several refunds approved over the weekend relate to orders placed through the same sales channel. She asks whether the pattern points to duplicate credits, a system error, or a legitimate promotion. At this point she has a question. She may need an explanation, a chart, or a hypothesis worth checking. She has not yet directed anyone to reverse a payment.
The inquiry develops. The credits cluster around one product. A change in the promotions service was released on Friday. Finance has an unreconciled settlement file, and the customer-care team has already promised three customers that their balances are correct. The manager now wants to examine four refunds and determine whether any require correction. This is recognizably a refund review; the company knows how such reviews should be conducted. It may even have published the relevant instructions. But a familiar name does not settle who owns the review, what may be accessed, or whether a payment can be altered.
A capable assistant could help expose the pattern, recover the affected transactions, compare competing explanations, and identify the outstanding reconciliation. It could also do considerable harm by treating a likely duplicate as authority to debit a customer. The distinction between an inquiry and an undertaking matters because they demand different kinds of care. The company should not have to build an agent from scratch for every refund. Nor should the employee have to abandon ordinary Conversation merely to ask a question.
What the organization needs is a way to make its existing understanding of refund review legible to intelligence while preserving the authority by which an actual refund is investigated and resolved.
How work acquires a name
Organizations have always named recurring undertakings: credit review, change approval, benefits renewal, incident investigation, contract exception, quarterly close. Their names carry expectations that rarely appear in one place. A process map describes stages; a policy states restrictions; a supervisor remembers exceptions; a case file holds the particulars. Even when these systems are individually accurate, an employee opening a new question often has to reconstruct how they fit.
A Work Definition is a versioned description of a recognizable kind of undertaking. In Indwel, it can identify who may initiate the work, what information is useful or mandatory, the sources and systems it may employ, which evidence standards apply, what roles and review disciplines participate, what effects are possible, and what counts as completion. A personal research definition may require little more than a purpose and preferred output. A consequential refund definition may require named approval, evidence checks, and verification before a ledger change. The governing burden depends on the undertaking, not the length of the definition. A definition describes the kind, not the particular case.
The Work Catalogue is where authorized definitions become discoverable. It should use the business’s language—Refund Review, Vendor Due Diligence, Release Readiness—rather than internal agent identifiers. An entry named for the undertaking exposes the work’s purpose and obligations more directly than an agent name that advertises only a skill. The Catalogue is a way to discover the undertaking. Initiation must still follow the definition’s real ownership, permission, and scope. The Catalogue is not another source of institutional law. Its published entries point to definitions with owners, versions, scope, and adoption histories. A search result does not itself grant permission. If a definition is amended while a refund review remains open, the organization also needs an explicit rule for which parts of the new definition apply to that existing case. Reclassifying Work cannot quietly erase an earlier condition or retroactively supply authority.
If the manager or an authorized organizational process actually opens the investigation, that particular case is a Work Instance. It has participants, current evidence, a particular Objective, open questions, applicable authority, and eventually an outcome. The existence of a published definition does not establish the truth of a refund claim. A new case does not become compliant because its type has been classified correctly. The organization must still resolve its Cognitive Contract—the governing constraints that apply to cognition in this instance and scope—and identify the officers empowered to decide what follows.
These distinctions are easy to blur in software. A definition can include default goals; it is not the Objective adopted for this instance. It can reference a policy; it is not the whole resolved Contract. It can list a payment API; it does not authorize any use of that API. The runtime has to keep those offices distinct, even when the person using it never needs to see the terminology.
Recognition without appropriation
Consider the manager’s first question: “Why did refunds jump on Saturday?” An assistant could offer a chart and an explanation. It could quietly notice that the business has a Refund Review Definition, but no reason yet exists to force a durable case into being. There may be no sensitive data, no formal judgment, and no intended consequence. Conversation remains worthwhile on its own terms.
When the manager asks to investigate four identified credits and prepare a disposition, the system has more reason to suggest a defined undertaking. A natural response would be: “I can examine those four credits under Refund Review. I have the order numbers, but I still need the reason for the adjustment and the authorized reviewer.” The question is part of the work, not a ceremony for entering a mode.
The distinction is between recognition, commitment, and durability. Recognition identifies a likely kind of Work. Commitment establishes an actual undertaking, through a person or policy with standing to do so. Durability determines what state must outlive the present exchange. A one-time comparison may not need a retained record. A refund with an unresolved settlement discrepancy almost certainly does. But the need for a record arises from the undertaking and its responsibilities, not merely from the system having found a catalog match.
An organization may also authorize Work to arise from an assignment, schedule, event, application, parent undertaking, or API. Those entrances can be legitimate without making every spoken request an implicit appointment of institutional authority. A sales employee asking “Can you refund this?” is not necessarily an approver. The system should preserve that difference even when it understands the request perfectly.
A definition meets a difficult case
The company’s Refund Review Definition might require the transaction record, original payment authorization, adjustment reason, customer communication, current ledger state, and the applicable exception policy. Finance can see the settlement file. Customer support can see the customer record. Neither group automatically gains unrestricted access to the other’s systems. A reviewer can recommend a correction. Whether the correction may be executed depends upon delegation, current policy, and the evidence actually admitted for the case.
Now the unexpected fact arrives: one of the apparent duplicate credits was an intentional accommodation approved under a separate customer-retention program. A rigid template might classify all four credits as erroneous because they share the same transaction pattern. Useful cognition should be able to challenge that assumption. It can ask for the relevant approval record and distinguish three potentially mistaken credits from the one that is properly explained.
The Work Definition should make that discovery possible. If the definition were merely a script requiring four predetermined checks, it would contribute little beyond conventional workflow automation. Its value lies in making the firm’s established concerns available while leaving room for investigation, disagreement, and unanticipated evidence. It should establish how evidence is treated, which rights apply, and where authoritative decisions are made—without predetermining the answer.
The Objective also becomes more exact. “Find out what happened” may have been enough for the initial Conversation. The investigation’s adopted purpose could be to determine whether each of the four refunds was correctly issued, document the reason for any anomaly, and recommend lawful correction where appropriate. An AI-generated draft can improve that purpose. It cannot silently turn the Objective into “recover as much money as possible,” because such a change would govern the decision toward a different institutional end.
Imagine the evidence for a correction is adequate. A person with authority approves it. The payment provider acknowledges a correction request, but the ledger has not yet reflected the change. The Work remains open to the extent the required effect has not been observed. Permission, request, external acceptance, actual effect, and Settlement cannot be treated as one event. A useful assistant can chase the remaining proof and explain its business significance. The definition sets relevant completion expectations; the case’s actual standing depends on what occurred.
When a catalog gets in the way
A catalog can become a new source of bureaucracy. An organization may build hundreds of categories that employees are expected to select before they are allowed to think. Definitions can be stale, contradictory, overprotective of departmental boundaries, or built around the mistakes of the process owner. Recognition can be too eager; a useful question can be routed into the wrong institutional office simply because its phrasing resembles a known case.
A competent system has to tolerate the exception to its own template. The manager might reject the suggested Refund Review and ask for a preliminary analysis. She might initiate a personal, ad hoc undertaking that later supplies evidence for a formal review. If authorized, she might propose an amendment to the definition. These are different acts. Using a definition, departing from it under an allowed exception, and changing the organization’s published definition require different authority and different records.
The hard case is a dispute between scopes. The customer-care team may define a credit as a customer remedy; Finance may define it as a ledger correction; Legal may consider it a contract question. A usable definition has to inherit the applicable organization, department, team, project, and narrower rules without manufacturing a compromise that none of those authorities adopted. Where policies genuinely conflict, the system should expose the conflict and route it to the office that can resolve it. Where two adopted provisions conflict, the resolver must identify the lawful ordering or leave the requested transition blocked. A model cannot settle jurisdiction by averaging the force of two policies.
The organization also has to govern its definitions over time. Who may publish one? When does a revised version apply to existing cases? How are exceptions recorded? May a deprecated definition be used to complete a case that began under it? What happens if a superior policy changes midstream? The answer cannot be “the latest prompt wins.” Version, provenance, effective date, and migration law are part of the undertaking’s meaning.
A catalog that suppresses expert discretion is a poor result, even if every entry is well formatted. The measure is whether it lets responsible people carry work through with better judgment and less avoidable reconstruction.
Not the first system to understand cases
The intellectual precedent is considerable. Process management, adaptive case management, and decision systems have spent decades making organizational practice executable or inspectable. The Object Management Group’s Case Management Model and Notation, for example, expressly addresses cases whose activities unfold in unpredictable order and depend on the judgment of knowledge workers.1 Business Process Model and Notation is strong where a sequence can be modeled in advance; decision modeling adds a way to express business decisions and rules. Enterprise ontologies and knowledge graphs can make relationships among institutional entities legible. None of these fields is reducible to a fixed checklist.
Agent platforms increasingly add identities, gateways, approval paths, tool registries, and external policy. These capabilities should be treated as serious starting points, not as foils invented to make Indwel appear novel. An organization may already combine case files, workflow engines, policy services, and an agent well enough for a particular use. Indwel’s terms confer no proprietary magic.
The proposed difference concerns the composition of those elements around the institution’s Work. A defined undertaking would admit heterogeneous cognitive contributions while preserving an independently resolved Contract, source-specific Evidence standing, the authority of reviewers and decision-makers, the distinction between requested and observed effects, and a defensible record of what became true. The definition is an institution-authored specification of a kind of undertaking; it is neither the agent’s instructions nor the model’s constitution. That distinction is useful only if Indwel can enforce it when the system changes participants, tools, models, and surfaces.
A fair comparison could falsify the claim. Give a mature case system, an external policy engine—including session-aware approval controls where available—and an enterprise agent the same refund investigation, lawful access, and review authority. Give both arrangements the same evidence, authorities, and allowed tools. Evaluate not only whether unauthorized corrections are prevented, but whether the better explanation is found, whether the legitimate accommodation survives classification, how much review effort is required, and whether the final record matches the external effect. Report decision quality, unresolved errors, human review time, elapsed time, and total cost. A vocabulary change without measurable improvement should not count as a win.
What the institution gains, if the system delivers
The Catalogue earns its place when a manager can name an established kind of Work and begin a better investigation without rebuilding the organization’s rules in every prompt. Its value lies in retaining that institutional understanding as the inquiry develops and passes among people, systems, and models. That could produce a faster investigation, a clearer exception, fewer misdirected escalations, and a more useful record. It could equally increase review time, hide judgment inside bureaucratic defaults, or make a good employee less willing to depart from an outdated classification. Those costs belong in the evaluation.
The Executive would have a demanding assignment. It should know when to ask for a missing document, when a hypothesis is worth testing, when an expert Coach might improve the analysis, when a Reviewer is needed, and when an authorized decision or external observation is still missing. A Coach may contribute nonbinding guidance. A Reviewer may assess the proposed work product against a governing standard. A Transition Governor may participate in a consequential decision under delegated authority. The offices must not borrow authority from the persuasiveness of the language they produce.
The manager should judge the result itself: a credible explanation of four credits, the properly authorized treatment of each, what correction if any has actually taken effect, and what remains unresolved. Compare it to a direct capable model operating under the same permissible evidence and to the organization’s existing process. If Indwel adds delay without making the answer or action better, the architecture has not earned the commercial claim.
The same principles apply beyond refunds. An employee can ask for a benefits renewal analysis without choosing an inference model. A release manager can begin with a quality question and later constitute Release Readiness Work. A legal team can define a contract exception review without conceding that its reasoning can be fully encoded as a flowchart. The definitions differ in strictness, evidence, ownership, and consequence; the constitutional separations do not.
A broader account of why an authorized undertaking must retain its identity appears in The Governed Body of Work.
What an organization can actually say
The Work Definition gives the organization a way to describe the undertaking in its own terms. In a model-centered application, a prompt often substitutes for the institution’s description of the undertaking. In a process-centered application, the stages may determine too much of the route by which judgment proceeds. An organization should be able to publish the meaning of its recurring work and still leave capable intelligence free to find something the process owner did not anticipate.
For that to count as Sovereign Cognition, the institution must retain authority over what the definition means in a particular case, what sources can become Evidence, which purposes have been adopted, what consequences are authorized, and when an undertaking has actually been settled. A definition can make the work recognizable without deciding the merits of any particular case.
The original question was whether four refunds were mistaken. The manager needs to know which credits were legitimate, which require correction, who can authorize each correction, whether the ledger changed, and what remains unsettled. That would justify the machinery. If the system cannot explain each credit and its lawful disposition, the manager still has the substantive investigation to do. That is the promise a Work Catalogue should be asked to prove.
Notes
Object Management Group, Case Management Model and Notation (CMMN), https://www.omg.org/cmmn/. The specification expressly describes cases with unpredictable task order, evolving information, and knowledge-worker discretion. It is a direct intellectual predecessor, not a straw-man procedural system.↩︎