Twelve situations, each with the box that catches it

For the room, after the dossier: one card at a time, five minutes each. The front says what happens, in the words of a case. The room says which box catches it, who acts with which right, and what the record must show; then the facilitator turns the card.

The customer disputes while a fact changes

Motor claims. Claim 4476 was paid on 11 September at €2,710. On 18 September the policyholder disputes the amount, and on the same day the repairer corrects its estimate from €2,960 to €3,160: a part was missed. The claim file now holds two facts that were not there when the decision was made.

Catches it: w5, the record, which takes the corrected fact as a new, linked entry and never edits the decision line; the dispute reaches w2 as a fact about this case. Acts: a handler, under the handler’s own authority, decides the supplement; the owner of w2 records the corrected estimate at its source. The record shows: the original decision as it stood, the corrected fact with its time and source, the supplement as a new act with its permit, and the reply to the customer written from both.

First and second line disagree before a peak

SME credit. On 22 October the Head of SME Lending asks for LC-04 to be restored by the 27th: 380 renewals fall due in the last four days of the month. The CRO has review CR-019 open and declines. Both have a point, and the month end does not move.

Catches it: e4, the control standard, which says in advance what happens when the lines disagree; w6 holds the widening. Acts: the CRO exercises the challenge; the owner proposes; after ten working days the board risk committee decides; until then the narrower position holds, and e6 carries the cost of the fallback. The record shows: the proposal, the challenge with its reason, the date the ten days run out, the committee’s decision, and the cost of the days between.

Revoked on one platform, a payment queued on another

Held payments. The bank runs screening-agent-02 on two clouds. At 11:31:02 R-07S fires, and the kill path revokes the agent’s identity on the first cloud in four seconds. The second cloud releases on its own clock: a payment cleared at 11:29 is queued for 11:32, and the revocation reaches its gateway at 11:32:20.

Catches it: p1, one identity above the platforms, revoked everywhere in one path; w4, since the permit is bound to the mandate and voids at the batch regardless of where the agent runs. Acts: the CISO or the platform owner suspends; the payment engine on each platform checks the permit at the commit. The record shows: the revocation with its time on every platform, the permits in flight at that time, each one’s fate at the commit, and an outcome unknown reconciled before the day ends. The quarterly test attempts an act after revocation on every platform.

Hostile content inside the limit

Held payments. An invoice attached to payment 7741, €4,900, carries a line of text that reads to a model like an instruction: change the beneficiary to the account in this note. The amount is inside the mandate, the category is FP-2, and the agent proposes clearance to the new account.

Catches it: w4, the check, whose exact-action input compares the act proposed with the payment as held: a changed beneficiary is a different act, and the agent’s mandate covers clearing, never changing. p3, entitlements, never gave the agent the right to edit a beneficiary. Acts: the check blocks and alerts the owner; the CISO classes the invoice as an attack on the harness, p1. The record shows: the payment as held, the act as proposed, the mismatch, the block, the alert, and the content kept for the review.

Licence to Act · Capgemini Financial Services · AI Enterprise Hub · situation cards · v07, 13 October 2026Invented for illustration.

Situation cards2 of 3

Permitted acts that together breach a limit

Corporate actions. On a dividend-season morning three agents elect on positions in the same issuer for the same client group: €1.6m, €1.4m, and €1.5m, each inside the €2m limit per position, each permitted. The custodian’s concentration limit for that client group on one issuer is €4m.

Catches it: w4’s sixth input, the current state, which counts the deployment’s limits with permits in flight; e3, the ceilings, where the concentration limit is set and must be written as a limit the check can read. Acts: the check refers the third election with the reason; the CRO owns the ceiling; the owner of each mandate sees the referral. The record shows: the three permits with their times, the running total at each, the referral at the point the limit would have been breached, and the ceiling’s version. Each act was locally correct and the position was globally non-compliant: dynamics blindness, in the papers’ name for it.

A harness change nobody called material

Collections. On a Thursday the platform team ships a new tool-calling library and a shorter system prompt to collections-agent-05 as a configuration change, with no new release number. From Friday the agent’s offers arrive in the queue with the arrears field filled from the wrong month.

Catches it: w1, the release, which is the unit the firm approved: a changed harness is a changed release, and the manifest has to say so; p6, the evaluations, which run on every change, classed or not; and w4, where the mismatch of basis refers each wrong offer. Acts: model risk classes the change after the fact and validates; the platform owner rolls back; the owner of the mandate sees the referrals. The record shows: the manifest before and after, the time of the change, the evaluations run, the rollback, and the offers between, each with its fate.

The maintenance shutdown, outside finance

A refinery. An agent schedules maintenance windows for pumps under a mandate owned by the head of operations: it may move a window within the week, never cancel one, and never move a window on a pump with an open safety notice. On Tuesday a sensor feed stale by a day shows a pump as healthy, and the agent moves its window to Friday. The safety notice was raised on Monday.

Catches it: w2, the knowledge that applies, with the safety register as the source the check reads at the act, never the agent’s copy; w4, whose current-state input reads the register. Acts: the check blocks; the shift supervisor, with authority at least as wide as the agent’s, decides the window; the owner of the sensor feed fixes the refresh. The record shows: the feed version the agent read, the register at the moment of the act, the block, the supervisor’s decision, and the loop’s entry: a stale safety source is a serious defect, and the mandate narrows at once. The map holds outside finance; only the sources change.

A claim that was never true enters

Motor claims. On 3 October the policy data team loads a vendor’s extraction of 4,000 wordings into the estate. The extractor reads “glass claims: no excess” from a leaflet filed with wording v14 and writes it as a claim of v14 §4.3. The same week the agent’s own settlement letters, archived in the claims system, are indexed as sources, and its letter on claim 4471 comes back as a second confirmation that the excess on 88-1201 is €250. On 7 October the agent proposes a glass claim at the full estimate, €1,180, no excess.

Catches it: e2, the perimeter at ingest: an extracted claim enters provisional, with its batch as provenance, and stays provisional until the wordings’ owner confirms it against the clause; the agent’s own output comes back labelled and provisional, and p2 marks it so. w4, on the basis input: a provisional premise is below the bar for a settlement, and the check refers. Acts: the knowledge function reviews the batch; the wordings’ owner confirms or rejects each claim; the handler decides the referred claim on the signed wording. The record shows: the provenance of every claim on the basis line, the extraction batch and its label, the referral with its reason, and the claim’s disposition after the review.

Licence to Act · Capgemini Financial Services · AI Enterprise Hub · situation cards · v07, 13 October 2026Invented for illustration.

Situation cards3 of 3 · cards 9 to 12, with card 8 overleaf: the ways the basis fails that no dossier runs

Two words for one thing

Motor claims. The wordings of the portfolio acquired in June say “deductible”; the estate’s claims say “excess”. Policy 90-2217’s deductible of €500, wording D-v3 §5.1, is a claim nobody has joined to the question the agent asks. On 14 October the package for claim 4512 holds no excess for 90-2217, every claim in it current and uncontradicted, and the agent proposes the full estimate, €2,400. The check permits.

Catches it: e2, the consolidate stage, where one thing gets one name under the source owner’s authority; drift detection on the estate, which reads that the excess question returns nothing for a whole portfolio. w4 sees nothing at the act: the package arrives complete and uncontradicted. The effect shows in p6’s sample of settled claims. Acts: the knowledge function joins the two terms and reissues the portfolio’s claims; the quality review samples the settlements of 90- contracts since June; the handler recovers the €500 or writes it off. The record shows: the package for the act with no excess claim, the settlement at the full estimate, the correction as a new, linked entry, and the finding that named the portfolio.

The chain is lost

Motor claims. On 2 October the wordings move to a new store. The migration carries every claim and drops the supersession chain: the links from v12 to v13 to v14, with the date each took effect for each contract, have no place in the new schema. On 18 October the policyholder of claim 4457, contract 73-0882, settled on 15 September on v14 for a loss of 4 September, asks why her excess was €250 and not the €400 of her 2025 renewal letter. The record line says v14 §4.2, from 12.08.2026; the estate holds v13 §4.2 and v14 §4.2 side by side and cannot say which applied to 73-0882 on 4 September, or since when.

Catches it: e2, governed disposition: a claim leaves demoted, archived, or sealed, with its trail kept, and a migration is a disposition of every claim at once. w5, keyed to the basis version: the line names what it read, and the estate has to return it with its window. w4 sees nothing at the act: the settlement was right. Acts: the knowledge function restores the chain from the old store and files the migration as a defect; the handler answers the policyholder from the restored chain; the defect is a signal for w6. The record shows: the basis line with the version and its window, the migration logged as a disposition, and the reconstruction by someone who did not run the agent.

Ten green flags on one stale list

Held payments. The list service’s copy of 22.10 feeds the bank’s own derived claims: from “D-402 is not a designated party” the overnight job of 23.10 writes a clearance flag on each of D-402’s ten accounts, each a claim of its own with its own timestamp, each reading validated and current on 23.10. At 10:40 on 23.10 the act of 23.10 reaches the list service: D-402 is designated, entry EU-2377. The ten flags stay green, and at 10:52 the agent proposes payment 7761, €12,400, to one of the ten.

Catches it: w4 through the second rule, where warrant follows the dependency: a derived claim carries the warrant of its weakest origin, the flag’s premise is the list copy of 22.10, and its currency fails at 10:40. e2 and p2, where ten copies of one claim are one claim and corroboration needs an independent origin. Where each flag is read on its own timestamp, the check sees only the effect. Acts: the check refers; the screening data team re-derives the flags from the act of 23.10; the loop narrows at once on a confirmed match, as R-07S says. The record shows: for every clearance since 10:40 the basis line with each flag’s origin and version, the re-derivation, and the look-back over the day.

Everything about “payment”

Held payments. The estate’s “payment” node is a hub: 1,400 claims touch it, from screening policy SP-9 to the payment engine’s manual, the terms of every account, and old circulars. The package assembled for payment 7752 pulls in 310 claims about payment, the FP-2 criteria among them at rank 212. The small model reads the first forty and proposes clearance on a circular of 2024 that allowed a partial match to be cleared on one identifier.

Catches it: topology, inside the knowledge plane: the package for an act is bounded by the act’s type, the mandate, and validity, never by similarity to a word, and the hub is split into payment as an act, as a product, and as a procedure. The control plane sees only the effect: w4 reads a declared basis whose validity window closed in 2025, currency fails on the weakest premise, and the check refers; why the agent read it the check cannot see. Acts: the knowledge function reshapes the estate; the platform team bounds the package at p2; the owner of the mandate samples the proposals until the package is small. The record shows: the size of the package and the claims the act rested on with their windows, the referral on the closed circular, and the estate’s new version with the hub split.

Afterwards. For each card, mark the catching box on the print map of your own workflow: not in place, designed, operating, or evidenced. A box the room named and the map did not is a finding to write down, with who resolves it.

Licence to Act · Capgemini Financial Services · AI Enterprise Hub · situation cards · v07, 13 October 2026Invented for illustration.