The Authority Process
Generative output is never a record. Label drafts, cite their sources, and turn a draft into an authoritative record only through one owned, eight-step process per record class: validation, business rules, data quality, reconciliation, a named human approval with rationale, immutable audit logging, posting through a write API only the process may call, and a separation-of-duties check with a visible exception register.
- Version
- 0.1.0
- Owner
- DEIXEI
- Last reviewed
- Reading time
- 17 min
For Cto · Ai Architect · Platform Engineer · Data Product Owner · Risk And Assurance · Internal Audit · Legal And Compliance
Sources and details
- Section
- Govern and prove
- Source works
- Book 1 · Supporting Tools
- Source chapters
- 16
- Tool specifications
- audit_evidence_schema@0.1.0
- Refresh trigger
- source change
An agent can now produce something that looks like a record (a risk classification, a journal entry, a contract memo, a control mapping) in seconds, with sources and a confident tone. Nothing in that artefact tells the next person whether it is a draft or the firm’s position. This page sets the rule and the process: generative output is never an authoritative record, and the only way a draft becomes one is through the Authority Process, one owned instance per record class.
Reading the labels. The draft-versus-record distinction, the eight steps, the signatory model and the sizing figures are an author model from Book 1 Ch 16, built from the author’s design reviews; they are not an industry standard, and the sizing is an experience-based estimate. Recommendation marks this page’s own synthesis, notably the mapping onto the Audit Evidence Schema. Asset Manager X and Bank A are composite firms; supplier 4471082 is invented. Book 1 links individual steps to EU AI Act human oversight, EBA internal-governance expectations and Solvency II actuarial-function validation; the article-level mapping lives on the regulatory map, not here. The source states that regulators do not require an artefact called “the Authority Process”; naming it is what makes the behaviours they require operable in one place.
Why the boundary is needed
Composite, Book 1 Ch 16 §1. A procurement agent summarised a supplier: legal entity, country, beneficial owners, risk class, last KYC review, open findings. Every bullet had a source, but none of the sources was authoritative. The agent had read a due-diligence report, a press archive and a public registry, not the master supplier register. It reported the parent entity, whose risk class was medium; the firm’s contract was with a subsidiary in another jurisdiction, risk class high, KYC overdue by 68 days. A buyer pasted the summary into a contract memo, and an operations director signed it. The head of supplier risk caught it three days later in a quarterly cross-check against master data.
Nothing was fabricated. What failed was that no control asked whether the artefact was a record. The question was missing because most operating models predate agents that produce record-shaped artefacts at speed.
Draft versus record
Book 1 Ch 16 §4.4 gives eleven pairs. The left column is what AI (or anyone drafting) may produce; the right column is what only the firm, through its governance, may produce.
| Generative output (draft) | Authoritative record | Who may produce the record |
|---|---|---|
| Explanation | Decision | The accountable decision-maker |
| Suggested action | Executed action | The authorised actor, through the governed path |
| Draft record | Posted record | The dataset’s write API, called only by the Authority Process |
| Proposed classification | Applied classification | The owning function (for example supplier risk for a risk class) |
| Recommended journal entry | Booked journal entry | The controller, through the books-and-records process |
| Suggested email | Sent email | The accountable sender |
| Drafted contract | Signed contract | Legal and the authorised signatories |
| Generated test plan | Audited test plan | The audit function |
| Proposed control mapping | Approved control mapping | The audit function (see control map) |
| Hypothesis | Confirmed fact | Verification against the authoritative source |
| Plausibility | Truth | (the general form of the other ten) |
The source names the producer for three pairs: legal and the authorised signatories for the signed contract, the controller for the booked journal entry, and the audit function for the approved control mapping. The rest of the third column is this page’s reading.
The same rule appears elsewhere in this reference: a generated plan or proposal is not an approved change (BOM and recipes, change governance). The Authority Process governs records; change governance governs changes to systems, configuration and controls. A draft that would alter both needs both.
Label every draft and cite every source
Two rules make the boundary visible to the human who reads the draft (Book 1 Ch 16 §6, composite practice):
- Header label. Every generated artefact that resembles a record carries the label
DRAFT (generative output)in its header. Recommendation: make the label a property of the output template and of any document the agent writes into, so that copying the content into another document keeps the label, and fail the template’s check if it is removed before step 7. - Inline source citation with authority level. Every field taken from data cites the dataset and its authority level, in the form risk class as of 12 March 2026, source supp_risk_001 (master). Unverified sources are cited as unverified. The authority level comes from dictionary field 8,
authoritativeness(master, reference, authoritative-copy, derived, operational, draft, unverified), on the agent-readable data dictionary.
Where the agent may assist and where it must not invent
From the composite supplier-onboarding flow (Book 1 Ch 16 §6), generalised:
| The agent may | The agent must not |
|---|---|
| Draft the memo by composing fields from master datasets and narrative from unverified sources, each cited at its level | Supply an identifier (for example the LEI) from anywhere but the master; if the master lacks it, flag a gap |
| Suggest a classification (for example an initial third-party classification) as a recommendation with rationale and citations | Set the classification; only the owning function, through the Authority Process, sets it |
Call a separate screening tool (for example sanctions) and attach its structured result (match / no match / manual review) | Infer beneficial owners from registry filings or press; if KYC is overdue it flags the status and routes to the KYC team |
| Collate open findings, citing masters as masters and knowledge-base hits as unverified | Write to the system of record (for example contract management); it drafts, the owning team posts |
Enforcement point. The agent’s tools are read tools. Dictionary field 17 (retention_and_agent_allowed_list, whose agent allowed list is enumerated and empty by default) lists them per dataset; write tools are never on it. The connector broker enforces the list (MCP and connector broker).
One Authority Process per record class
Properties of each instance (source claim, Book 1 Ch 16 §4.4):
- One per class of record: supplier risk class, KYC outcome, journal entry, regulatory submission line, control mapping, and so on.
- Owned, by a named process owner for that class.
- Named signatories, both human and machine, at every step.
- Auditable end to end from one record reference.
- Drafter-neutral. The same process applies whether the draft came from an agent, a junior analyst, a consultant, a vendor or an integration. Agents get no separate process; they are the volume driver that tests whether the existing one holds.
Book 1 reports that two years of design reviews found no need to add a field or a step, and repeated need to enforce the ones already there. Treat that as the author’s experience, not evidence of completeness.
The eight steps
Skipping a step is a control failure (source claim). Each row gives the step’s purpose, its signatories and what halts it. The checks and signatories are the source’s. In the “Halts or routes when” column, steps 1–3 and step 5 follow the source; the entries for steps 4, 6, 7 and 8 are this page’s recommendation. “Evidence” is also this page’s recommendation for what each step should emit so step 6 can assemble the trail.
| # | Step | What it checks | Machine signatory | Human signatory | Halts or routes when | Evidence to emit |
|---|---|---|---|---|---|---|
| 1 | Validation | Compares the draft with the authoritative source (dictionary field 8 = master) | Validation tool, logging its check against the dictionary entry | Wrangler of the agent that produced the draft | Mismatch: passes to step 2 flagged | Validation result, dataset and version read, mismatch list |
| 2 | Business-rule checks | Applies the record class’s rules (for example: no risk-class change without a documented finding; four-eyes on this change) | Rules engine | Rules-engine owner; second reviewer if a four-eyes rule fires | Missing finding reference: held; four-eyes rule: queued | Rules evaluated, rule version, outcome |
| 3 | Data-quality checks | Completeness, cross-field consistency, conformance to the dataset schema (dictionary field 14) | Data-quality engine | Data steward of the receiving dataset | Any missing required field, inconsistency or schema violation: halt | DQ result per rule |
| 4 | Reconciliation | Against adjacent records: parent/subsidiary, prior versions, related entities | Reconciliation tool | Data steward | Conflict: halt; related scope: “consider scope” note | Reconciliation note, related records |
| 5 | Human approval | A named human with authority for this record class approves explicitly | (none can substitute) | Named approver, role specific to the record class | No: exception path (step 8) | Approver identity, timestamp, mandatory free-text rationale, references to steps 1–4 |
| 6 | Audit logging | Writes the full trail to an immutable log against the Audit Evidence Schema | Audit-logging service | Audit-evidence owner | Logging unavailable: do not post (recommendation; see below) | The evidence record |
| 7 | Versioning and posting | Versions the record and posts it to the authoritative dataset; bumps dictionary field 15 (version) if the schema changed | The dataset’s write API, which only the Authority Process may call | Data-product owner | Write refused: return to step 5 owner | Posted record ID and version, provenance link to the draft |
| 8 | Separation of duties and control evidence | Confirms the step-5 approver is not the step-1 wrangler; attaches control evidence to the record; runs the exception path | SoD check | Control owner; Internal Audit on review | SoD conflict: block and re-route | SoD result, exception-register entry if any |
Boundary conditions (source claim). The agent can drive steps 1–4, which are mostly automation it is good at. Step 5 is the human checkpoint no agent can substitute for. Steps 6–8 are mostly automated again, with human owners.
Step 5: the approval record
The step-5 record needs, at minimum (source): the approver’s identity, the timestamp, a free-text rationale (mandatory) and references to the results of steps 1–4.
Recommendation: carry it in the AES approval group. The Audit Evidence Schema v0.1.0 already has the fields: approval.required (true for an Authority Process posting), approval.policy_decision_id and approval.policy_version (the record-class rule set that required approval), approval.rationale (the approver’s free text), approval.workflow_id (the Authority Process instance), and approval.approvers[] with upn, role and timestamp (and optional ticket_ref). The schema accepts an empty approvers[] when required is true; reject that in the emitter profile, because an approval with no named approver answers nothing. Store the step 1–4 results as references in the record’s payload; the schema does not define a field for them, so do not invent a top-level field.
Step 6: the trail
Book 1 lists the trail as: draft ID, agent ID, agent wrangler, validation result, business-rule result, DQ result, reconciliation result, approver identity, approver rationale, timestamp. In the composite it answers who acted, what was done, why, when, on what authority and with what evidence. Recommendation: if evidence capture is unavailable, the process must not post; a record whose provenance cannot be written is not yet a record. It is consistent with the principle on audit evidence that evidence is never dropped silently.
Step 7: the write API is the control
The rule that makes the whole process enforceable is architectural. Only the Authority Process holds the credential to call the authoritative dataset’s write API. Agents hold read tools only; humans who would bypass the process hold no direct write path either, except through a named, logged override (step 8). Book 1 Ch 16 §7 makes compliance responsible for translating step 5 into policy-as-code, enforced at runtime by the MCP gateway and the dataset’s write API. Verify by attempting a write with an agent identity and with a human identity outside the process: both must be refused and the refusals must appear in evidence.
Step 8: separation of duties and the exception path
- SoD check: the approver at step 5 is not the wrangler at step 1. In the composite, the agent’s wrangler is the head of procurement operations and the approver is the head of supplier risk: different people, roles and functions.
- “What if step 5 says no?” is named here: a standing exception register, a named override authority, and a timed re-review of every exception.
- Visibility: a quarterly, board-visible report of exception count, type and resolution (Book 1 Ch 16 §8). Exceptions are permitted; they are not free. The second use of an exception should trigger the question the first did not.
Worked example: one finding through the process
Composite (Bank A, Book 1 Ch 16 §6). A buyer asks the agent to raise supplier 4471082’s risk class from medium to high after a sanctions-watchlist hit on a beneficial owner.
| Step | What happened |
|---|---|
| 1 | Validation reads supp_risk_001 (master): the class is already high, set in 2024. The draft is current state, not a change. Flagged to step 2: no change required; the finding adds to the existing class. Time taken: under a second |
| 2 | Rules require a named finding for any class change. The 2024 finding exists; the new hit becomes a second finding. Adding a finding under the existing class is permitted. Pass |
| 3 | The finding record has date, source, beneficial-owner reference, severity and status; schema-conformant. Pass |
| 4 | Reconciliation finds three related suppliers in the same group and issues a “consider scope” note for cross-group review. Pass with note |
| 5 | The head of supplier risk receives the whole chain, approves with a one-paragraph rationale, opens the related-supplier review as a separate workstream, and signs |
| 6 | The full trail is written immutably against the Audit Evidence Schema |
| 7 | The finding is posted to supp_risk_001; the dataset version goes from 3.21 to 3.22; downstream consumers (contract-memo tooling, procurement dashboard, regulatory reporting feed) refresh through lineage_downstream |
| 8 | SoD passes (wrangler and approver are different people in different functions). Control evidence is attached; an auditor can reconstruct the chain from one record reference |
End to end: 26 minutes of agent and human time. The composite’s pre-process equivalent took three days, lost two emails, relied on an unchecked spreadsheet and drew a supervisory comment on incomplete provenance. The author’s point is that codified steps are faster than steps negotiated each time.
Roll out and size it
First iteration (Book 1 Ch 16 §7)
The process depends on the dictionary: step 1 needs to know which source is master. Roll the two out together.
- Pick three high-stakes datasets where an inaccurate agent read has the largest regulatory or financial consequence: typically a master (customer, supplier, counterparty, holdings or policy), the feed for the largest regulatory submission, and the master for the process being automated most aggressively. Three, no more.
- Populate the 17-field dictionary on those three. The owning domain populates; the central data function reviews; the agent platform team sanity-checks the agent allowed list; the Chief Data Officer signs. Expect resistance on field 8 (
authoritativeness) and field 11 (accuracy_sla). Field details are on data and API integration. - Wire one MCP server and one API to one of the three datasets, both taking schema and tool descriptions from the dictionary entry.
- Run the Authority Process on one real generated artefact end to end with the named signatories: a risk-class change, a KYC re-review outcome or a submission line. Time it and log every step. Expect the first run to expose a missing rules engine, an unclear approver or an audit log missing evidence fields. Fix the gaps before scaling.
Target: three populated dictionary entries, one MCP server, one Authority Process run, by the end of week six. The rest of the dictionary follows in two further iterations through the year.
Sizing (author’s estimates)
| 50,000-employee FS firm | 1,000-employee specialist firm | |
|---|---|---|
| Steady-state dictionary | 2,000–8,000 dataset entries, 30–60 domains | 100–400 entries, 5–12 domains |
| MCP-published for agents | 200–800 | 30–80 |
| First iteration | 8–12 weeks; team of six (data architect, two data-product engineers, agent platform engineer, governance lead, programme manager) plus owning domains | 4–6 weeks; team of three |
| Recurring cost | 3–5 FTE central enablement, plus 0.1 FTE per data product in the domain | 1–2 FTE |
| Authority Process owners | Usually already in post (heads of supplier risk, KYC, financial crime, controllership; investment-data and actuarial-data owners). Their work is reshaped, not added | Same |
The smaller firm’s temptation is to skip the dictionary because “everyone knows where the master data is”. That holds until the second and third agents start reading.
Who owns what
| Role | Deliverables | Evidence they sign |
|---|---|---|
| Process owner per record class | The instance: rules, approver role, write-API binding | Process definition, versioned |
| Head of compliance | (1) Per-dataset mapping of dictionary fields to the obligations they discharge; (2) policy-as-code for step 5, enforced by the MCP gateway and the write API; (3) per-domain attestation by each data-product owner that the fields are current, authoritativeness is honest, and the allowed list is an enumerated exception | Quarterly dictionary attestation to the operational-risk and model-risk committees: master datasets populated, accuracy SLAs measured, Authority Process exercised on a tested sample |
| Head of internal audit | (1) A query template that returns the full chain for any posted record from one reference; (2) a testability matrix sampling allowed-list enforcement, honest field 8 values on masters, and exceptions in the register with timed re-review; (3) a half-yearly test of the draft/record boundary: sample posted records and confirm a named step-5 approver who is not the wrangler | Half-yearly internal-audit report on the agent-data plane to the audit committee |
| Agent wrangler | Step 1 signatory; keeps the agent’s tools read-only | Wrangler attestation (accountability) |
| Data-product owner | Step 7 signatory; owns the write API | Dataset version history |
The compliance and internal-audit rows are from Book 1 Ch 16 §7. The process-owner, wrangler and data-product-owner rows are this page’s recommendation, built from the signatories in §5.2.
Align the dictionary with the model usage registry
The source flags a fifth failure next door: the dictionary’s field 4 (data_classification) and the model usage registry’s field 7 (data_classes_touched, on model governance) must agree, or an agent invokes a model with data classes the model is not approved for. Run the cross-reference query at registration time. Data classes themselves are on data classes and lanes.
Failure modes and recovery
| Failure (Book 1 Ch 16 §8) | Signal | Recovery |
|---|---|---|
| Generative output treated as record | Drafts posted because posting is automated; agent confidence used as a quality gate; the first wrong posted record (journal entry, customer communication, applied class) exposes the gap after many have been posted | Reintroduce the process for the affected class; reconcile the affected period retroactively; forbid agent writes except through the process, enforced at field 17 and the write API |
| Process bypassed under time pressure | Quarter-end or deadline exceptions recur; within two years the exception is the norm | Standing exception register, named exception authority, timed re-review, quarterly board-visible exception report |
| Agent reads the data lake without a dictionary | Agent reaches operational exhaust, prior agent outputs and stale snapshots with nothing marking which is which | Revoke broad read; populate the dictionary on needed datasets; restrict through the MCP layer; require citation of the dictionary entry for every source |
| MCP servers but no glossary anchors | Field 9 null on most entries; one term means different things in two domains | Populate the glossary first; anchor every entry; review quarterly |
| Dictionary not aligned with the usage registry | Data classes on the usage row disagree with the dataset classification | Cross-reference query at registration |
| Label stripped in copy (recommendation) | A draft pasted into another document loses its DRAFT header | Label carried by the template; posting refused while a draft label is present without a step-5 approval |
Acceptance checks
- Every record class that agents draft for has a named Authority Process owner, a defined rule set and a named step-5 approver role.
- An agent identity, and a human identity outside the process, are both refused when they call the write API directly; the refusals are in evidence.
- Every sampled posted record has an AES record with
approval.required: true, a non-emptyapprovers[]and a rationale, and links to step 1–4 results. - In a half-yearly sample, no step-5 approver is also the step-1 wrangler.
- Every sampled generated artefact carries the
DRAFT (generative output)label and cites sources with their authority level. - The exception register shows owner, authority, re-review date and resolution for every exception, and the count is reported quarterly.
- One record reference is enough for an auditor to reconstruct the chain, without rebuilding it.
Provenance and refresh
Synthesised from Book 1 Chapter 16 (§§1, 4.4, 5.2, 6–8) and the Audit Evidence Schema v0.1.0 approval group. The draft/record pairs, the eight steps, the signatory model, the rollout and the sizing are the author’s model and estimates. The AES field mapping, the fail-closed logging rule and the label-persistence rule are this page’s recommendations. Regulatory linkage is on the regulatory map.
Review trigger: a change to Book 1 Ch 16, to the AES approval group, or to the dictionary schema; an incident in which a draft was treated as a record; or an exception count that rises two quarters running.
Source material
- Book 1 (Building Organizations for the Age of AI), Chapter 16, Data as the Enterprise Multiplier (§1 composite opening, supplier 4471082; §4.4 generative output is not an authoritative record; §5.2 the Authority Process; §6 worked example, supplier onboarding; §7 Monday morning, sizing and role deliverables; §8 anti-patterns)
- Supporting Tools, Audit Evidence Schema v0.1.0 (the approval group)
Related guidance
-
Govern APIs, Data Products and Retrieval Agents Can TrustTurn interfaces and data into owned, versioned capabilities that applications and agents consume under identity, entitlement and evidence controls, and give agents access to meaning, limits and authority, not only bytes. -
The Audit Spine and the Evidence RecordHow to capture, join, protect, retain and prove per-action evidence: the AES v0.1.0 record contract, three capture points and one correlation ID, tiered immutable storage, retention and legal hold, replay rehearsal, the route-approval evidence gate, and the Claude-specific ways evidence goes missing. -
Accountability and OwnershipHold provenance and accountability apart on purpose: four named accountabilities per product, a human owner for every agent change, seven named seats, a signed Build/Run RACI and a first-line/second-line boundary that survives consolidation. -
Change Governance for Agent-Enabled DeliveryRoute agent-speed change by risk without losing the approver: control values that always reach a person, a routing gate that reads the right signals, autonomy earned per change family, a Virtual CAB that prepares a complete dossier, and three bodies with defined escalation. -
Route Regulatory Questions to Obligations and EvidenceA dated, article-level reference that routes AI design questions to the EU AI Act, DORA and its RTS, GDPR and sectoral rules, and turns each obligation into a control, an owner and evidence: change records under RTS 2024/1774 Art 17, the 4 h / 72 h / one-month incident clocks, the CIF hinge and its five artefacts, the three-count Register of Information reconciliation, Art 14(4) oversight controls, the Art 26(5) discharge, the FRIA gate, Art 49 and Art 50, SCHUFA-grade review and supplier controls. -
Govern MCP Connectors Through a BrokerAdmit, tier, pin and enforce every MCP server through one registry and two enforcement points, keep tokens audience-bound, govern A2A delegation through its own registry and gateway, and expose client-facing agent channels only through the same policy layer. -
Bounded-Agent PatternA class 3 agent lets a model choose the next step, but only inside a registered, validated and enforced envelope of Mandate, Means, Memory and Monitoring. -
Model Governance: Usage Registry and Upgrade LifecycleGovern models at the grain regulators ask about: a catalogue for what exists, an 18-field usage registry for what is used, risk classified per usage, a registry the policy decision point reads by usage_id, pinned model IDs, and a four-phase upgrade lifecycle from Day −14 to Day +90 with numeric rollback thresholds. -
Build a Living Product Bill of Material and RecipeDescribe what an AI product contains and how the firm builds it in one versioned specification shared by pipelines, agents and assurance teams, including the classification line, readiness grades and reuse generations. -
Build a Unified Control MapJoin the firm's AI and technology controls in one traceable operating view: rows are operating concerns, columns are frameworks read at their own altitude, every cell points to live evidence, and existing ITIL, COBIT and CIS artefacts are extended to agents with a five-question test instead of a parallel AI governance track. -
Data Classes, Processing Lanes and TokenisationClassify every model request by its data, route it only to a processing lane the class permits, keep derived artefacts inside their source class, and treat retention, roles and tokenisation as recorded, testable decisions.