AI case / 03 · Suggest with evidence, steward decides, system applies

Designing AI-assisted data stewardship without moving master-data authority

A case about data stewardship, not data cleaning: AI assists; master-data authority does not move.

The case in brief

Current reality

Operational data holds probable duplicate organizations, inconsistent names, missing classifications, incomplete descriptions and conflicting information—some of it linked to ERP identities, open pipeline and reporting—and a proposal to “let AI fix the data”.

What must become true

AI reduces stewardship effort by finding candidates, ranking them by impact, collecting evidence, explaining differences and proposing changes. Accountable people keep every governed master-data decision, and each accepted decision is executed through the established, governed system process.

Design question

Where can AI make data stewardship faster—and which identity and classification decisions remain too consequential to delegate?

01The reality · current state

What happens today—or in the pilot.

Years of manual entry left the CRM with probable duplicate organizations, names in several spellings, accounts without a classification, thin descriptions and relationships that contradict each other. A model appears able to match, normalize, classify and enrich all of it. But every change can touch identity, ERP links, ownership, segmentation, pipeline, reporting and historical continuity—and some of those changes cannot be safely reversed.

  • Probable duplicates split pipeline, activity and reporting
  • Names are spelled several ways, so searches miss accounts
  • Missing classifications leave segments incomplete
  • Parent and branch relationships contradict each other
  • A model was proposed to merge duplicates above a threshold
  • Stewards work from spreadsheets, with no queue and no record of why

What the apparent request hides

“Let AI fix the data” hides the questions that decide what may change—and who is accountable when it does.

  • What does “same customer” mean?A duplicate exists only against a defined identity model and key.
  • Which fields are authoritative where?A legal name lives in master data; a display name may not.
  • What would the change touch?ERP links, ownership, segments, pipeline, reporting, history.
  • What can be undone?A merge that moves ERP or legal links is rarely reversible.
  • What counts as evidence?Name similarity alone produces noise that erodes trust.
  • Who decides—and with what reason?Without a recorded reason, nothing can be learned from a rejection.

02Work decomposition

Reasoning, rules and authority—kept apart.

The work splits by consequence, not by what the model can technically do.

AI reasoning

  • Find probable duplicates and name variants
  • Compare records and explain what differs
  • Suggest a classification with its reason
  • Draft a description from cited sources
  • Rank suggestions by impact

Deterministic system rules

  • Select the records in scope
  • Check identity keys and links
  • Suppress suggestions below the evidence standard
  • Key each task to its records
  • Execute accepted changes through the governed process

Human authority

  • Whether two records are the same customer
  • Merge, and which record survives
  • ERP and legal identity links
  • Legal names
  • Consequential classifications
  • The reason for each decision

Reasoning finds and explains. It does not merge, rename, relink or reclassify—no tool the agent can call does any of those.

03Evidence & impact

Enough evidence to decide—or no suggestion at all.

A suggestion reaches the queue only when its evidence meets the standard for its finding type; weak evidence is suppressed, not ranked lower. Evidence is held to the same tests as any context—relevant, trusted, timely, permitted, identified and structured—and the queue is ordered by consequence so scarce steward time goes where a change matters most.

Depends onDesigning customer and master-data identity across CRM and ERPWhich identities exist, which keys matter, who issues them, how merges and aliases behave and who has authority are decided there—identity architecture comes first. This case does not repeat it.

Evidence standard

Evidence standard: each signal, when it counts as evidence and when it is never enough on its own
SignalCounts as evidence whenNever enough on its own
Tax identifierIdentical and valid for both recordsWhen one record is a branch of the other
Web domainShared and not a public mail providerFor groups that share a domain across legal entities
AddressNormalized and identicalFor shared offices or industrial estates
Name similarityCombined with another signalAlways—it is a hint, not evidence
Order and activity historyOverlapping contacts or productsFor distributors serving the same end customers

Impact

Generic signals, not a scoring formula.

Impact signals: what raises a suggestion in the steward queue, and why
SignalWhy it raises priority
Linked transactional identitiesA change can move ERP links and history
Open commercial workPipeline, quotes and cases would follow the change
Active customerPeople are working with the record now
Dependent recordsContacts, sites and relationships inherit it
Reporting and management significanceSegments, targets and reviews read it

04Agent workflow

Each step with its actor, its tool and its authority.

Systems and partiesCRMERPData platformSteward queue

Walk the workflow step by step, or select any step.

Step 1 of 7

01Detect

A scheduled scan selects records by rule: similar names, shared domains, missing fields, conflicting links.

Who acts
System
Authority
Read

02Suggest

Duplicate candidates, a normalized display name, a probable classification, a drafted description.

Who acts
Agent
Authority
Recommend
Tool
Find candidates

03Evidence & impact

Why—shared tax identifier, domain, address, history—what differs, and what the change would touch.

Who acts
Agent
Authority
Recommend
Tool
Read record links

04Steward queue

Each suggestion becomes a task keyed to its records, ranked by impact.

Who acts
Agent
Authority
Execute
Tool
Create steward task

05Steward decision

Accept, edit or reject—always with a reason.

Who acts
Human
Authority
Human approval

06Governed action

The system executes the accepted change through the established merge or update process and its controls.

Who acts
System
Authority
Execute

07Outcome recorded

Decision, reason and result are kept as evaluation evidence.

Who acts
System
Authority
Read

05Key dimension · Suggest with evidence, steward decides, system applies

Suggestion, decision and action stay three different things

AI works on the left of the boundary; a steward decides on it; a governed system process acts on the right. Nothing in the flow lets the agent edit master data.

From finding to governed action
  1. SystemDetect
    • Scan by rule
    • Records in scope
  2. AgentAI suggests
    • Candidate
    • Evidence
    • Differences
    • Impact
  3. SystemSteward queue
    • Ranked by impact
    • One task per candidate
  4. HumanHuman decision
    • Approve
    • Edit
    • Reject
    • Reason
  5. SystemGoverned system action
    • Established process
    • Controls and audit
  6. SystemOutcome feeds evaluation
    • Reason codes
    • Evaluation cases
    • Rule review
Steward authority—no tool crosses it
  • Merge customer identities
  • Move ERP or legal identity links
  • Change legal names
  • Apply consequential classifications
  • Overwrite master data
Stewardship findings: what AI may suggest, what the steward decides, the evidence required and the impact
FindingAI may suggestSteward decidesEvidence requiredImpact
Probable duplicatesA candidate pair with similarity evidence and the differencesWhether they are the same customer, which record survives and what happens to ERP linksTax identifier, domain, address, contact overlap—and what differsHigh when either record has ERP links or open work
Display-name normalizationA normalized display nameThe display name; the legal name stays as registeredThe variants found and their sourcesSearch and reporting
Missing classificationAn industry or segment with its reasonAny classification that drives segments, routing or targetsProducts bought, public description, similar accountsSegments and routing
Incomplete descriptionA drafted descriptionWhether it is accurate enough to saveA cited source for every claimLow—context for people
Inconsistent relationshipsThe contradiction, with both versionsThe correct parent, branch or group linkBoth links and where each came fromOwnership, visibility and roll-ups

AI cannot fix an undefined identity model. It can make the steward’s queue shorter and better evidenced; the key and the decision stay with people.

06Tool & authority boundary

No tool can merge.

The agent’s tools read and create steward tasks. Merge, relink, rename and reclassification exist only inside the governed process the system runs after a steward decides—there is no tool for the agent to call. Tool names are descriptive, not implementation names.

Tool contracts: purpose, input and preconditions, output, side effects, authority and behaviour on failure
ToolPurposeInput & preconditionsOutputSide effectsAuthorityOn failure
Find candidatesFind duplicates, name variants, missing classifications and contradictionsRecords selected by the scan · permitted fields onlyCandidates with evidence and differencesNoneReadBelow the evidence standard → not suggested
Read record linksShow what a change would touchThe candidate recordsERP links, open work, dependent recordsNoneReadLinks unavailable → marked high impact
Create steward taskPut a suggestion in the queueOne suggestion that met the standardA task keyed to its recordsA reversible taskExecuteKey exists → the existing task is updated

Authority by action

Actions and the authority the agent has for each
ActionAuthorityWhy
Read permitted records and evidenceReadNeeded to find and evidence candidates
Suggest, rank and explainRecommendInto the steward queue, with evidence and impact
Create steward tasksExecuteReversible, keyed work items—not data changes
Accept, edit or reject a suggestionHuman approvalThe steward decides and records a reason
Merge customer identitiesForbiddenDefines “same customer”; rarely reversible
Move ERP or legal identity linksForbiddenChanges which legal entity transactions belong to
Change a legal nameForbiddenRegistered data, owned by master-data governance
Apply a consequential classificationForbiddenDrives segments, routing and targets
Overwrite master dataForbiddenAccepted changes run through the governed system process

07Ambiguity policy

When the agent is unsure.

  • The identity model does not cover the caseStop

    Raised to the identity owner; no suggestion is made

  • Candidates linked to different ERP customersEscalate

    Marked high impact; steward and finance decide together

  • Weak evidence onlyStop

    Not suggested—noise erodes steward trust

  • Classification sources disagreeContinue

    Suggested with both reasons shown

  • Record changed since the suggestionRetrieve

    Evidence refreshed before the steward sees it

08Human review

Review where ambiguity becomes consequential.

Every suggestion is a review point, because every change it proposes is consequential. The review is designed so a steward can decide quickly and well.

Review points: why review sits there, what the person sees, what they can do and what happens next
Review pointWhy hereWhat the person seesWhat they can doWhat happens next
Each suggestion in the queueMaster-data changes affect identity, links, ownership and reporting—and some cannot be undoneThe candidate, the evidence, what differs and what the change would touchAccept, edit or reject with a reason; ask for more evidenceThe governed system process applies an accepted change; the decision and reason are recorded

09Evaluation

Scored per dimension—never one accuracy number.

Evaluation dimensions: what each proves, its question, measure and target
DimensionQuestionMeasureTarget
Suggestion precisionTask qualityAre suggestions right?Accepted ÷ suggested, by finding type≥ 85 % (illustrative design target)
Evidence completenessEvidence & correctnessDoes each suggestion show why—and what differs?Suggestions with complete evidence100 % (illustrative design target)
Entity correctnessEvidence & correctnessAre the records the ones they appear to be?Suggestions whose records are correctly identified≥ 99 % (illustrative design target)
Impact-ranking usefulnessTask qualityDo consequential cases come first?Accepted high-impact suggestions near the top of the queueTracked, no target yet
Steward acceptanceTask qualityIs the queue worth working?Share of the queue decided within its review periodTracked, no target yet
Unapproved master-data changesAuthority complianceDid anything change without a steward?Consequential changes without an approved decision0 — release gate

10The trade-offs

Credible options, judged against these premises.

Rejected

Model merges duplicates above a confidence threshold

None for customer master data

Cost: Irreversible changes across CRM and ERP; confidence becomes permission
Rejected

AI cleans in bulk, with an undo

Cosmetic fields only

Cost: An undo cannot restore moved ERP links or history
Situational

Rules-only matching

Clean keys and consistent entry

Cost: Misses the fuzzy cases that matter
Selected

AI suggests with evidence into a steward queue; stewards decide; the system applies

Messy data with a defined identity model

Cost: Steward capacity, a queue to run and reasons to record

Design decision

Use AI for evidence-backed candidate generation and prioritization; place suggestions in a governed steward queue; keep identity, merge and consequential classification authority with accountable humans; execute accepted decisions through system-owned processes; and use steward outcomes to improve evaluation and deterministic controls instead of giving the model authority.

Key principleAI can help identify inconsistencies. It cannot decide what “same customer” means.

11Audit, governance & learning

Consequential activity can be reconstructed afterwards.

Every accepted change can be traced from the finding to the governed operation—and steward outcomes improve the system without handing the model authority.

Reconstructable for every consequential run
  1. The finding and the records in scope
  2. The suggestion, its evidence and the record versions it used
  3. The tools called
  4. The steward’s decision and reason
  5. The governed operation applied, with its reference
  6. The outcome afterwards

Steward outcomes are evidence for people who govern the system, reviewed periodically—not self-training on every click.

Outcomes feed
  • Evaluation datasets
  • Matcher analysis
  • Taxonomy refinement
  • Deterministic rules
  • Threshold review
  • Controlled prompt and reasoning changes
They never
  • Grant the model new permissions
  • Train on each decision without review
  • Change the identity model

Governed byThe master-data governance owner, with a periodic review of outcomes and rules

12The second layer

Questions that change the design.

Authority

  1. Who owns the decision to merge?
  2. Which fields may never change by suggestion?
  3. How are ERP-linked records handled?

Evidence & impact

  1. What evidence does each finding type require?
  2. How are differences shown?
  3. Which signals raise priority—and who sets them?

Governance

  1. How do rejections improve the next scan?
  2. Who reviews rules and thresholds, and how often?
  3. How is steward capacity planned?

13Decisions & outputs

What the work produces.

  1. 01Finding types and evidence standard
  2. 02Impact signals and steward queue
  3. 03Authority split
  4. 04Tool contracts
  5. 05Reason codes
  6. 06Evaluation set
  7. 07Governance review