Key takeaways

  1. The organisation declares the objective, authority and policy. The system derives facts and candidate actions.
  2. Every production decision needs a named owner for inputs, rules, promotion, rollback and review.
  3. A result is not operational until another person can reproduce it and route failures to the right function.

A capable model plus a prompt is not an operating model. Production begins when the organisation states what the system may decide, which facts it may derive, who owns each boundary and how the result can be challenged, changed and reproduced.

An operating contract, not a longer prompt

Many production failures begin as category errors. A system is asked to infer a policy that nobody approved. A field team treats a missing input as a product defect. A generated record is presented as customer evidence. A successful response code is accepted as proof that a result arrived.

A written operating contract separates these categories before the first production decision. It does not need legal language. It needs explicit answers to three questions: what must the organisation declare, what may the system derive, and who controls the lifecycle of the result.

Operating contract for Applied AIThree columns separate what the organisation declares, what the system derives and what a named owner controls. THE OPERATING CONTRACT ORGANISATION DECLARESObjectiveAuthorityPolicyRisk limitsApproval gatesThese cannot be invented. SYSTEM DERIVESCurrent stateRelationshipsCandidate actionsExpected effectsTrace to recordsEach derivation is reproducible. NAMED OWNER CONTROLSInput qualityRule changesPromotionRollbackRegression reviewControl follows authority. MONITOR · CHALLENGE · CHANGE · REPRODUCEThe contract remains visible after launch
Figure 1. The system can derive mathematics from records. It must not invent the objective, authority or policy.

The declare and derive boundary

Derivation is appropriate when a fact already exists in the records and the transformation is deterministic. The current state of an order, a relationship between two records or a candidate action under an approved rule can be derived and traced.

A rule is different. Guessing a fact changes what the system knows. Guessing a rule changes what it decides. Objectives, authority, policy, risk limits and approval gates therefore belong to the organisation.

This boundary also makes absence visible. Without a declared cost, a system should not quietly rank options by cost. Without a declared route, it should not invent a movement. “Cannot decide until this is declared” is a production control, not a model weakness.

Derive the facts. Declare the rules. Name the owner.

Ownership is a production control

Ownership is often written after the technology is built. That is too late. The owner of the source records is needed on day one because every downstream decision moves at the speed of that person’s answers.

The organisation also owns the lifecycle of its rules. Promotion, effective dating, rollback and regression are not implementation details. They determine which decision policy is active and how quickly the enterprise can recover when a change behaves differently from the test.

Data accountability follows control. Whoever can move the data is accountable for where it ends up. Retention follows a written policy. If no policy exists, it is defined before launch rather than discovered during an incident.

Reproducibility is the definition of done

A pipeline is not production ready because it ran once. The test is whether someone else can run it again from what was recorded, without asking the author to remember the missing steps.

That means mappings live in configuration rather than renamed files. Transformations live in code rather than hand edits. Source versions and run times are captured. The team can prove that the result being reviewed came from the run started after the latest change.

Hand edits are especially dangerous because they disappear from the next refresh. A correct result produced by an invisible intervention is not a repeatable result.

Route the failure to the right function

Production Applied AI depends on accurate escalation. A missing input is a scope conversation. Undocumented behaviour belongs with the service owners. A capability that genuinely does not exist is a product gap and should arrive with the business impact attached. A defect belongs with engineering and a reproducible case.

Mixing these categories wastes time and damages trust. The field should never turn “I could not find it in the documentation” into “the product cannot do it”. It should also never hide a missing capability inside a custom workaround and call the operating model complete.

Failure routing in production Applied AIFour failure types route to scope, service, product or engineering. ROUTE THE ACTUAL FAILURE MISSING INPUTScopeDecide whether theinput is required. NOT EXPLAINEDServiceConfirm the actualbehaviour. CAPABILITY GAPProductState the businessimpact. REPRODUCIBLE DEFECTEngineeringSend the smallestcase that fails. One symptom can belong to four different owners. Diagnosis comes before escalation.
Figure 2. A production team moves faster when each failure reaches the function that can actually resolve it.

Safe handling is part of the operating model

Before analysis begins, the team establishes whether the data is real, generated or transformed. Generated data read as if it were real becomes false evidence the moment it enters a decision or an executive update.

Safety should be mechanical where possible. Source control can block data files, names, identifiers and keys before they are committed. A transformation that changes the meaning of a date or a status can require explicit agreement. These controls catch habits before they become disclosures.

An output derived from protected records is still protected. Turning the source into a chart, a forecast or a recommendation does not make it public.

Production is a maintained loop

The operating contract does not disappear after launch. The team monitors the result, challenges it, changes the declared rules when the business changes, reruns the method and proves which records produced the next answer.

This is why Applied AI in production is not mainly a model selection problem. It is an operating design problem with a model inside it.

Kaushal Dave

Kaushal Dave

Kaushal Dave leads Applied AI from strategy to production for global enterprises. His work spans Forward Deployed Engineering, supply chain, global retail pricing and economic decisions.

ProfileLinkedIn