Part two · Licence to Act · the interactive map
1 / 1
Licence to Act · Capgemini Financial Services · AI Enterprise Hub · October 2026 · witold.reichhart@capgemini.com
03
The control plane map · part three

Inside the essentials

Each essential taken apart into the pieces an architect builds, the second line sets limits on, and an auditor examines. A slide for each essential, two for knowledge, two for the record, and two for the loop, and each ends with what the market sells for it.

1 · The agent itself2 · Its knowledge3 · Its mandate4 · The check5 · The record6 · The loop
Part three · what this part shows

For specialists: the parts of each essential, and who owns them.

Part one lists the parts of each essential and their owners, and part two follows one decision through the essentials. This part opens them: each slide takes one essential apart into the pieces an architect builds, the second line sets limits on, and an auditor examines, and the last slide says which pieces a firm can buy and which three joins it has to build. It assumes part one; part two is useful and optional.

  1. 3
    Essential 1: it is never just the agentThe model, its harness, guardrails, identity, and gateway are approved together, as one release.
  2. 4
    Essential 2: the kinds of knowledgeThe law as read, policy, contract terms, and the case, each with an owner, versions, and a date from which it applies.
  3. 5
    Essential 2: claims, not documentsThe check reads one dated claim, traced to the clause it comes from. The signed wording stays the source.
  4. 6
    Essential 3: authority as an objectToday the mandate is scattered across four places. Here it is one object, with a version and one owner.
  5. 7
    Essential 4: one gate, in the path of the actThe check between the agent’s proposal and the system that executes it, with no other route to that system.
  6. 8
    Essential 5: a trace and a recordA trace says what was called. A record says on what authority, keyed to the mandate version and the build.
  7. 9
    Essential 5: the record line, field by fieldFourteen fields, from the second of the decision to retention, filled for one act of the motor case.
  8. 10
    Essential 6: signals, rules, and the return pathsWhat the loop reads, the rules it applies, and the four places a fix goes back to.
  9. 11
    Essential 6: narrow on a rule, widen on a signatureThe first rule, autonomy may not rise above evidence, and the two directions follow different rules.
  10. 12
    Most of it can be bought; three joins have to be builtWhere the products stop, and what the firm builds itself.
At the end you can answer
  • Why does a firm approve a release, and never just a model?
  • What does the check need from the mandate to say no before the act?
  • What does a record hold that a trace does not?
Reading on your own

About 20 minutes. Every slide builds in steps: → moves through them before it moves on. Click a line on the left to go straight to that slide.

M opens the contents, F finds a word, and T switches between light and dark.

Short of time? Choose by role: risk and compliance, slides 4 and 5; architects, 7, 8, 9, and 12; audit and the examiner, 9; model risk, 3 and 11.

Then part four turns the map to your firm.

Inside the essentials · essential 1

It is never just the agent.

What acts is a model inside a harness, behind guardrails, with an identity, reaching the firm’s systems through a gateway. The firm approves all of it together, as one release.

THE RELEASE · VALIDATED AS ONE UNIT GUARDRAILS what goes inwhat comes out THE HARNESS looptoolscontext memorysandboxverification the modeland its routing Its own identityacting for a business unit, with short-lived credentials Every call through a gatewayto the firm’s systems and tools, logged and limited In the agent portfolioregistered, approved or not, with a named owner One switch stops itand the work goes back to people

In a controlled study on software-engineering tasks, a 100-task subset of SWE-bench Verified with three near-tied models, the harness explained 7.8 times more of the variance in results than the model, a ratio its authors call not universal (Zhang et al., 2026). So the firm approves a release, not a model: a new model, prompt, or tool is a new release, validated again.

What people call itagent harness · guardrails · agent identity · AI gateway · agent registry
In the market, 09.2026Microsoft Agent 365 and Entra Agent ID; Google’s Agent Gateway, Agent Identity, and Agent Registry; AWS Bedrock AgentCore and Guardrails; Okta, SailPoint, and CyberArk for agent identity; NVIDIA NeMo, Guardrails AI, and Lakera for content checks.
Inside the essentials · essential 2

The kinds of knowledge, each with an owner.

An agent relies on more than documents. Each kind has its own owner, its own versions, and a date from which it applies.

The law, as the firm reads itconduct rules, sanctions regimes, data protection, the AI Act
20252026
compliance and legal · e2
Policy and proceduredelegated authority, credit or underwriting policy, the manual
v7v8
the business owner and risk · e2
Product and contract termsthe policy wording, the facility agreement, the client’s standing instruction
v13v14
product, and the team that holds the record · e2, w2
The case itselfthe claim file, the held payment, the latest accounts, the event notice
filednow
operations, in the system of record · w2 · the act’s input, never the estate’s content
What applies · w2to this case, on the date that matters

The version in force on the date of the loss, the payment, or the decision, not the latest one, and not the most similar.

Search · p2finds what is similar

The most similar passage, from any version. Usually right, wrong exactly when a version has just changed.

The owner · e2decides what applies

Each source owner keeps the record of what is in force and from when. The check, w4, asks that record at the moment of the act.

What people call itenterprise context · RAG · the knowledge layer · data governance · golden source
In the market, 09.2026Catalogs and lineage (Collibra, Informatica) govern the data; permission-aware search (Glean, Elastic) ends at retrieval; vector stores (Pinecone, Weaviate) take their governance from the systems around them; Palantir’s ontology and Databricks Unity Catalog come nearest to governed information.
Inside the essentials · essential 2

Govern the claim, not the document.

A document is a container. What can be checked, dated, and owned is the single assertion inside it: a claim, in the knowledge-layer sense. Each claim is traced to the clause it comes from, and what the firm writes for the case becomes a view on the table.

The documentMotor policy wording · v14well written, versioned, approved, and still one paragraph can be out of date
decompose
ClaimSourceOwnerStandingValidConflicts
Excess on policy 88-1201 is €250wording v14 §4.2policy data teamvalidatedfrom 12.08.2026none
Policy 88-1201 covers accidental damagewording v14 §2.1policy data teamvalidatedfrom 12.08.2026none
The repair estimate for claim 4471 is €3,650the claim fileclaims operationsvalidatedat 11.09.2026none
The customer receives the reasons for the decisionconduct rules, as readcompliancefoundationalin forcenone
Excess on policy 88-1201 is €400wording v13 §4.2policy data teamvalidatedto 11.08.2026superseded
a view: what the agent receives for claim 4471 · €3,650 − €250 = €3,400
What every claim carriesprovenance · source authority · evidential warrant · a validity window · contradiction status · a named authority · a governance status: what it may be used for today
The spec is a view; the instrument bindsFor what the firm writes, a letter, a product sheet, a spec, claims are the base table and the document is a view on it. A document that binds, the signed wording, the contract, the regulation, stays the source: where it and a claim disagree, the document wins and the claim is corrected. The agent receives a bounded package of claims, read by the check against a bar.
The set’s second rule · standing scales with consequenceThe higher the consequence of the act, the higher the standing its basis must have, on both axes: how far the evidence supports the claim, and who asserted it under what mandate. The first rule widens the agent over quarters and narrows it at once; this one binds the basis at every act. Chapter 05 takes it apart.
Where this comes fromthe Governed Intelligence work: the claim as the governed unit, the seven-requirement contract, and the spec as a view (Reichhart and Gelas, 2026)
Inside the essentials · essential 3

Authority as an object, with a version.

Today an agent’s authority is scattered across four places. Each answers part of the question; none answers all of it.

A role assignmentwho it is
A policy filewhat is allowed in general
An access reviewwhat it can reach
A system promptwhat it was told
into one object
Mandate MC-05version 05.2 · since 15.06.2026
Owner
Head of Motor Claims
Agent
claims-agent-01, for Motor Claims
Action
decide routine settlements; the payment service executes
Level
decide, within the limit
Limit
€5,000 a settlement · €1.5m a day
Conditions
wording confirmed · estimate verified · not paid before
Review trigger
three referrals for one reason in seven days
Revocation
the owner, or a rule
Challenger
the CRO, as the second line
Restore
after 20 clean proposals from a test chosen for the cause and run by model risk, on the owner’s signature
Expiry
31.12.2026, or the next release
Deputy
the Head of Claims Operations
The chain of delegation
Ceilingsenterprise: how far any mandate may go
The mandateworkflow: what this agent may decide
Entitlementsplatform: which systems it may reach

What was this agent permitted to do at 10:02 on 17 September, who owns that answer, and which version was in force? Only an object with an owner and a version can answer, as at any minute.

What people call itpermissions · entitlements · delegated authority · autonomy level · approval matrix
In the market, 09.2026Policy engines (OPA and Styra, Cedar, PlainID) keep policy as versioned code, the nearest thing to a versioned mandate if the policy is written that way; Palantir’s action types carry permissions; Okta, SailPoint, and CyberArk set identity and access, not authority for this decision.
Inside the essentials · essential 4

One gate, in the path of the act.

The check sits between the agent’s proposal and the system that executes it. It reads the agent, its knowledge, and its mandate, and there is no other route to the system.

The agentproposes the act System of recordexecutes the act no other route Guardrailsjudge the content THE CHECK identityreleasebasis mandateexact actioncurrent state permit refer a person decides,with the reason block no act; the owner is alerted
Guardrails judge the content: what goes into the model and what comes out.
Access control judges the reach: which systems the agent can call at all.
The check judges the act, by rules, every time. The model does not judge itself.
Human in the loop is two jobs: deciding what the check refers, and reviewing samples of what it permitted.
What people call itpolicy enforcement · action gate · approval step · human in the loop · kill switch
In the market, 09.2026AWS Verified Permissions with Cedar decides per request, the nearest thing to a checkpoint; gateways (Portkey, LiteLLM, Kong) log every call but do not make the action decision; workflow engines (Pega, Camunda, Appian) and LangGraph interrupts hand referrals to a person.
Inside the essentials · essential 5

A trace says what was called. A record says on what authority.

A trace
model call · 1.2 stool: policy lookupmodel calltool: payment APIdone
what the agent called, and how long it took
A decision record
when11.09 · 08:14
who actedclaims-agent-01 · release v2.4
model routetier 1 → tier 2
basiswording v14 · estimate
authoritymandate MC-05.2
the checkpermit · six of six
the act€3,400 · receipt P-4471
cost€0.34 tokens
a personnone needed
for the customerthe reasons, in plain words
keyed to the mandate version and to the release, so the decision can be rebuilt as at that minute
Keptappend-only for the decision; a corrected fact is a new, linked entry; the record, the manifest, and the permits kept for the retention period set for decisions of that tier, personal data by purpose and retention rule; readable by people who did not run the agent
Read by sixCTO CISO CRO compliance business owner CFO
The gapWe found no record keyed to both the mandate version and the build. Traces rarely say which permission was in force when the call was allowed.
What people call itaudit trail · observability · traceability · lineage · evidence
In the market, 09.2026LangSmith, Langfuse, Arize Phoenix, W&B Weave, and Braintrust trace calls, costs, and evaluations; Datadog, Dynatrace, and Splunk bring traces into existing monitoring, not execution receipts; MLflow links traces to the app version and AWS to the runtime version; Credo AI’s Agent Governor records the policy version in force (research preview, July 2026).
Inside the essentials · essential 5

The record line, field by field.

Fourteen fields, written once for every act, in this order. The right-hand column is claim 4471 of the motor case; each dossier fills the same fields for its own case.

Before the act · what the check read
whenthe second the check decided, and the second the act committed
decided 11.09.2026 08:14:06 · committed 10:00:41, the payment run
the agent and its identitywhich agent, acting for whom, under its own identity, and how long its credential lived
claims-agent-01, acting for Motor Claims · its own identity · token issued 08:12:50, valid 15 minutes
the releasethe release id, each model by version and hash, the prompt set, the validation report
v2.4 · tier-1 model 2026.06.1 (7c41e0) · tier-2 model 2026.05.2 (e9a3b7) · prompt set claims-settle 2.4 · validated by model risk, MRV-031, 01.09.2026
the basiseach claim the act relied on, its version, and its standing at that second on both axes: evidential warrant, and authority
The excess on policy 88-1201 is €250 · wording v14 §4.2 · from 12.08.2026 · warrant validated · current · uncontradicted · authority: the signed wording · the policy data teamPolicy 88-1201 covers accidental damage · wording v14 §2.1 · from 12.08.2026 · warrant validated · current · uncontradicted · authority: the signed wording · the policy data teamThe repair estimate for claim 4471 is €3,650 · the claim file · 11.09.2026 08:13 · warrant validated against the repairer network · authority: claims operations
the mandate versionthe mandate in force at that second, with its level and its limit
MC-05.2 · since 15.06.2026 · decide · €5,000 a settlement · €1.5m a day across the deployment
the control standardthe control standard version and the rule set the check applied
CS-3 v2026.2 · rule set RS-MC-05 r14
the consequence tierthe tier the act was classified into, and the version of the taxonomy that classified it
routine · taxonomy CT-2026.1 · to €5,000, one policyholder, recoverable inside the recall window, inside the firm
At the act · what the check decided
the check’s decisionpermit, refer, or block, how many of the six inputs held, and the reason code
PERMIT · six of six · reason code OK
the exact actwhat was done, to whom, for how much, under which id, once
settle claim 4471 · €3,400 to B-082 · C4471-S1 · once
the permitits id and what it was bound to: the mandate version, the release, the basis version, the exact act; and how long it held
PR-4471-1 · bound to MC-05.2, release v2.4, wording v14 §4.2 and estimate E-4471-1, €3,400 to B-082 · valid for the next payment run
After the act · what came back
the receiptwhat the receiving system returned, and the reconciliation against it: one permit, one act, one receipt
P-4471 · €3,400 · the 10:00 run · acknowledged by the payment service 10:00:41 · reconciled against the payment ledger: one permit, one payment, one receipt
the human decisionwhere a person decided: who, under which authority, what they confirmed, and when
none needed
the costwhat the decision cost in tokens, and in a person’s time where one was needed
€0.34 in tokens · no reviewer time
retentionhow long the record, the manifest, and the permits are kept, and personal data inside them by its purpose
the record, the manifest, and the permits: ten years from the act, the period the insurer sets for settlement decisions of this tier (illustrative) · personal data by purpose and retention rule · a corrected fact is a new, linked entry

Two of the fields are version keys, the mandate and the release, and they let the decision be rebuilt as at that second. The other twelve say what it rested on, what it did, and what came back. So the examiner, the auditor, and the knowledge function read one line where today they read three systems.

Where this comes fromthe Knowledge Layer paper’s manifest; every case carries both of its acts in these fields; the ids and figures are invented.
Inside the essentials · essential 6

Signals in, rules in the middle, the ways back.

The loop reads every record, and what it finds changes the parts before it. That is what makes the control plane a living system.

referrals complaints, appeals audit findings drift and quality incidents cost against value Rules,written in advance three referrals for one reasonin seven days: narrow a serious defect:narrow at once Enterprise: expand, hold, or stop 1 · reroute the model, roll back the release 2 · correct the source, at its owner 3 · narrow or restore the mandate 4 · retune the checks
The inner loopthe release, the sources, the checks · hours to days · the platform and data teams
The outer loopthe mandate · on a rule, or on a signature · the business owner, challenged by the CRO
What people call itmonitoring · evaluation · drift · model risk review · reauthorisation
In the market, 09.2026Monitoring (Arize AX, Fiddler, Galileo, Patronus) ends in an alert, and a person changes the configuration; evaluation scores (Databricks Agent Evaluation, Salesforce Testing Center) do not change permissions; Entra Conditional Access can block an agent on risk, but only the whole agent.
Inside the essentials · essential 6

Narrow on a rule, widen on a signature.

The set’s first rule: autonomy may not rise above evidence. Each level of autonomy needs a level of evidence, and the two directions follow different rules.

Narrowautomatically, on a rule written in advance, at once
AutonomyEvidence it needs
A.5full autonomy, no oversight
–no evidence condition allows it
A.4all tasks, in stated circumstances
K3governed: every act with warrant and a named authority
A.3most tasks alone; a person decides the critical ones
K2reconstructible by someone who did not run it: the check recomputed from the record and the rule set, the proposal read and never re-run
A.2partial; a person is ready to take control
K1measured: every act recorded and attributed
A.1one automated task
K1measured
A.0a tool, no delegated act
K0logs only; ordinary model risk applies
Widenonly on evidence, and only with the owner’s signature; the second line can challenge
Not independentThe agent’s own traces do not count as evidence for widening its own authority.
Irreversible actsWhere an act cannot be undone, the evidence it needs rises one level.
From the reference modelautonomy levels A.0 to A.5 and evidence levels K0 to K3, with the rule that joins them; a proposal under review, with K2’s label still open. K0, logs without attribution, is where most pilots are today.
In the market, 09.2026Microsoft’s open-source Agent Governance Toolkit drafts trust decay and quorum approval; the CSA Agentic Trust Framework demotes on incident and promotes on sign-off; access packages widen on approval, for the whole agent.
Inside the essentials · what is bought, and what is joined

Most of it can be bought. Three joins have to be built.

1 · The agent itself
agent frameworksagent identitygatewaysguardrailsregistries
2 · Its knowledge
catalogs and lineagepermission-aware searchvector storesontologies
3 · Its mandate
policy as coderoles and access reviews
4 · The check
per-request policy decisionsapproval stepsgateways
5 · The record
tracingLLM observabilityprocess logs
6 · The loop
monitoring and drift alertsevaluation scoresinventories
a versioned mandate the check can query
a record keyed to the mandate version and the build
a return path from the record back to the mandate

All three joins meet at the mandate, and they are what lets the control plane say no before the act. That is where the firm’s own work sits, on whatever stack it already runs. Next: one decision, through all six.