Numbers and Evidence Age

The fourteen volumes give Hansa Mutual a great many numbers: headcounts, budgets, rates, thresholds, drill times, sample results. This chapter gathers them, says which of them the dossier treats as established and which as illustrative, and then does what the series itself insists on for every claim: it gives each number's evidence age, the clock by which the number stops being a fact about the organisation and becomes a fact about a past quarter. The chapter draws on all volumes. It ends with the figures on which the volumes disagree, which are recorded in Chapter 27 and are not resolved here.

How the numbers are labelled

The chapter applies two rules, and both are the dossier's own. A figure is established when it belongs to the profile of the organisation and is repeated unchanged by at least two volumes: the headcount of about 9,500, the seven countries, the split by country and function, the size of the SOC, the red team and the SRE function, the roughly 20 handlers of a claims team. A figure is illustrative when a single volume gives it, when the volume labels it illustrative, or when it is a result, a budget or a measurement in a year after 16 November 2026. Every money figure in the series is illustrative: AI Operations says so of its budget table, and no other volume gives an established one. The label travels with the number in every table below.

The second rule is the evidence age. The series treats a claim about the organisation, such as "the stop path works", "the restore is verified", "coverage is complete", as a claim with a date on it. Each volume gives its own clock, and the clocks differ by what they protect. Figure 23.1 sets them side by side.

Evidence age clocks the series gives Hansa Mutual

Figure 23.1: The clocks of the series. A claim is current while its evidence is younger than the clock for its kind; the clocks are the volumes' own, not the dossier's.

The organisation in numbers

Item Figure Status Source
Employees; people served About 9,500; about 2 million Established (7, Nordqvist, 2026) (15, Nordqvist, 2026)
Countries Seven: four EU (DE, NL, SE, PL), one EEA (NO), two outside (UK, CH) Established (2, Nordqvist, 2026) (15, Nordqvist, 2026)
By country DE 4,600; PL 1,700; NL 900; UK 800; SE 600; CH 500; NO 400 Established (7, Nordqvist, 2026)
By function Claims 3,600 (of which about 1,400 in Poland); sales and distribution 2,100; underwriting 1,100; IT 1,300; corporate 1,400 Established (7, Nordqvist, 2026)
Managers and layers 715 managers (1 in 13); about 28 function and country heads, 160 department heads, 520 team leaders; claims teams of 15 to 25 Established (11, Nordqvist, 2026) (7, Nordqvist, 2026)
Boards Executive board 7; supervisory board 20, of whom 10 employee representatives Established (executive); illustrative (supervisory, legal basis to be verified) (11, Nordqvist, 2026)
Security teams SOC about 38 (22 analysts, 6 seniors); red team 5; threat intelligence 3; detection engineering 4; architects 6 Established (4, Nordqvist, 2026) (8, Nordqvist, 2026) (9, Nordqvist, 2026) (6, Nordqvist, 2026)
Operations Group IT about 1,300; infrastructure and cloud about 420; SRE about 95; service desk about 140; product-line operations about 310 Established (IT, SRE); illustrative (others) (10, Nordqvist, 2026) (12, Nordqvist, 2026)
Application development About 480, of whom 290 developers; 62 delivery teams; 190 external developers; platform team of 24; 12 COBOL and PL/I maintainers; about 6.5 million lines Illustrative (one volume) (12, Nordqvist, 2026)
Incident commanders 12 on a follow-the-sun rota Illustrative (10, Nordqvist, 2026)
Pilot handlers About 320 for the claims assistant; 1,900 with access to the correspondence assistant Established (320); illustrative (1,900) (2, Nordqvist, 2026) (3, Nordqvist, 2026)
AI estate 17 AI uses, of which six Mode 5, in the landscape model for 2027 Illustrative (15, Nordqvist, 2026)
Licences 8,100 embedded-assistant licences; 340 coding-assistant licences; about 120 engineers on agentic tools Illustrative (11, Nordqvist, 2026) (12, Nordqvist, 2026)

Energy, carbon and benefit (Book 14)

All figures of AI Environment are illustrative, and each carries the date and the boundary at which it was measured (14, Nordqvist, 2026). The location-based factor is 350 g CO2e a kWh unless stated and the PUE of Frankfurt is 1.38.

Item Figure Boundary and date Evidence clock
Frankfurt installed IT power; accelerators About 2.8 MW; 96 accelerators in 12 servers (installed 2029-03, planned life five years); utilisation 0.52 against a target of 0.70 and a floor alarm of 0.45 Profile of 2032 Hardware record 90 days
September 2032, platform racks 62 MWh IT, 85.6 MWh at facility boundary, 30.0 t CO2e location-based; about 3.4 million requests, about 25 Wh a request with idle capacity Facility boundary Month
September 2032, GPU racks 49,800 kWh Rack Month (OT-HM-76)
Claims assistant, FPR-HM-0001 0.996 Wh a question (1,096 kWh a month) on an evaluation set, 2032-10-18; 3.91 Wh before and 2.42 Wh after levers (about 4,303 and 2,664 kWh a month) on a sentinel pool, 2032-10-28; never added Node boundary; rack-IT boundary Footprint 90 days
Claims-triage agent, FPR-HM-0002 Bottom-up 2.3 Wh (1.1 to 4.8) and 0.81 g a claim; top-down 21 Wh and 10.3 g; about 13 times at one boundary Facility and rack, 2032-10-06 Footprint 90 days
Classifier model, FPR-HM-0007 880 kWh IT, 1,215 kWh facility, 425 kg location-based, 85 kg market-based; 0.087 g a use over 24 months As of 2032-09-30 Review 2033-01-04
Cascade, MCR-HM-0001 0.996 Wh a question at 22 per cent escalation; quality 96.8 per cent (95.9 to 97.7) against a requirement of 96.0; break-even escalation 0.67 1,500 questions from four claims specialists, 2032-10-18 Valid to 2033-01-16
Triage benefit, BEN-HM-0002 Correct routing 93.5 per cent for the agent against 91.8 per cent for rules over 6,000 claims in eight weeks to 2032-10-31; about 2,380 extra correct routes a month; 48 g per extra route Compare fails for the built use Expired 2033-05-14
Hardware options, HWL-HM-0001 Keep 826 MWh and 289 t a year; consolidate to nine servers 710 MWh and 249 t; replace with four 365 MWh and 128 t; stepwise 489 MWh and 171 t; energy cap 830 MWh a year Valid to 2033-02-06 90 days
Cloud A re-scoring batch About 9,500 accelerator hours: supplier tool 0.41 t market-based, own estimate 2.2 t location-based 2032 Disclosure 12 months
AI inference and platform spend About EUR 17 million in the 12 months to August 2032 (EUR 9 million in the profile of 2028) Illustrative Year
Claim CLM-HM-0001 revision 2 Claims assistant, test pool, IT energy a question in the rack, 3.9 to 2.4 Wh (minus 38 per cent) Valid to 2033-01-26, withdrew itself At most the reporting year

Illustrative (14, Nordqvist, 2026). The figures that differ by the direction or boundary of counting are the book's finding, and it never adds them. Two sit in the table of conflicts below and in Chapter 27.

Money

Item Figure Status Source
Group IT About EUR 310 million a year; operations about 140 million; cloud about 48 million; observability 6.5 million; managed service provider 11 million; AI inference and platform 9 million Illustrative (10, Nordqvist, 2026)
Showback, 2026 Q3 EUR 818,000, with a fixed control plane of EUR 180,000 a quarter and unallocated cost capped at 5 per cent Illustrative (2, Nordqvist, 2026)
Correspondence assistant Platform fee EUR 4,200 a month; use-case budget 310,000; 2027 frame 540,000; cost per answer p50 0.21, p90 0.47, ceiling 0.60 Illustrative (2, Nordqvist, 2026)
Claims assistant EUR 0.031 per answer at pilot against a budget of 0.04; gateway route EUR 18,000 a month with alert at 0.8 and hard stop at 1.1 Illustrative (2, Nordqvist, 2026) (3, Nordqvist, 2026)
Triage agent Envelope EUR 6,500 a week and 84,500 a quarter, daily cap 1,300; cost per completed action p50 0.18, p90 0.41, ceiling 0.60 (2027 Q1); controlled-rollout ceiling EUR 4,000 a month Illustrative (2, Nordqvist, 2026) (3, Nordqvist, 2026)
Triage service Per triaged case p50 EUR 0.04, p95 0.09; per day 1,900; per month 38,000 Illustrative (10, Nordqvist, 2026)
Agent budgets HM-SEC-001 EUR 9,000 a month; HM-OPS-001 6,000 (indicative); MND-HM-01 6,000; emulation runner 4,500 (54,000 a year) plus about 1,200 for a second model family Illustrative (8, Nordqvist, 2026) (10, Nordqvist, 2026) (6, Nordqvist, 2026) (9, Nordqvist, 2026)
Smaller envelopes Platform engineering Mode 3: EUR 5,600 a month for 14 seats; coding run EUR 6; replica uplift CAP-HM-2028-014 EUR 1,900 a month Illustrative (2, Nordqvist, 2026) (12, Nordqvist, 2026) (10, Nordqvist, 2026)

Limits and thresholds

A limit is a number that a mandate or a policy fixes. These are the figures most often cited by later volumes.

Subject Limit Source
Triage agent At most 400 actions an hour; may not approve payments; 20 steps, EUR 0.50 and 240 seconds per task; new versions replay 260 recorded reports; stop after three policy denials (2, Nordqvist, 2026) (3, Nordqvist, 2026)
Approval monitoring Flag an approval faster than 4 seconds; review if the approval rate exceeds 98 per cent over four weeks (3, Nordqvist, 2026)
IT-operations agent At most 3 restarts an hour per namespace; 2 to 12 replicas; 30 steps; 4 write actions; 20 minutes an incident; rollback needs on-call approval within 10 minutes; recommend-only again after more than two agent-caused incidents a quarter (3, Nordqvist, 2026)
HM-SEC-001 Stop target 120 seconds; 5 workstations an hour; catalogue limit 10 actions an incident and 20 an hour; missed-threat rate at most 0.02, false-containment rate at most 0.01 (4, Nordqvist, 2026)
Chaos experiments Production at most 1 per cent of targets, 3 concurrent, 30 minutes; test share 0.20 with 20 safe runs before production (4, Nordqvist, 2026)
Payments Call-back from EUR 10,000; 24-hour cooling-off after a payee change (5, Nordqvist, 2026)
Review protocol At most 12 open proposals a reviewer (8 at night); 5 per cent seeded errors; catch threshold 0.80 (8, Nordqvist, 2026)
Agent-boss pilot Seeded catch 0.85; rejection rate 0.02; median approval 30 seconds; 40 letters a day (25 temporarily) (7, Nordqvist, 2026)
Operations Error allowance 1.5 per cent; budget review at 70 per cent; at most 20 pages a shift, at least half actionable; page acknowledged in 5 minutes (10, Nordqvist, 2026)
Leadership delegation Level 3 only for travel under EUR 2,500 a trip and board scheduling; 10 per cent weekly sample; 10 seeded errors a quarter; stop if catch falls below 90 per cent (11, Nordqvist, 2026)
Coding run 45 minutes, 200 tool calls, 25 files, EUR 6 (12, Nordqvist, 2026)

How long it took to stop

The stop path is the figure the series measures most often, and the measurements run through the whole chronology. All rows from 2027 are illustrative.

Date Subject Measured Target Result Source
2026 Q4 SPT-HM-ITOPS-2026-Q4 Manual switch 9 s; honeytoken 31 s; gateway anomaly 74 s with one process left; production drill 11 s None given; maximum token lifetime 5 minutes Failed; autonomy change blocked until a retest passes (5, Nordqvist, 2026)
2027 Q1 STOP-HM-ITOPS-2027-Q1-01 17 minutes (notice 9, decide 2, act 5, verify 1) 5 minutes Failed; re-drill in June 4 minutes (6, Nordqvist, 2026)
2027-04-12 Real stop of HM-SEC-001 57 seconds 120 seconds Held (8, Nordqvist, 2026)
2027-04-27 Drill 3 min 40 s 2 minutes Missed; 2027-06-03 re-test 1 min 46 s (8, Nordqvist, 2026)
2027-07-13 Tabletop Acknowledged 18 minutes; decided 51 10; 30 Missed (8, Nordqvist, 2026)
2028-02-15 Emulation runner 38 seconds 60 seconds Held (9, Nordqvist, 2026)
2028-10-24 SPT-HM-ITOPS-2028-Q4-01 4 min 10 s 5 minutes Held; 2029-01-17 3 min 55 s (10, Nordqvist, 2026)
2032-11-15 Triage agent, EBG-HM-0001 (Frankfurt; burst route in Cloud A) 40 seconds; 11 minutes on the burst route 5 minutes Failed; the burst key was held at the provider (14, Nordqvist, 2026)
2032-11-29 Same, after the key moved into the gateway 2 min 10 s 5 minutes Held; repeated 2033-01-12; valid to 2033-02-27 (14, Nordqvist, 2026)
2029-10-23 Triage chain ACC-HM-2029-01 6 minutes 10 minutes Held (11, Nordqvist, 2026)

Evidence age

The table follows the clocks of Figure 23.1 and gives the state of each claim at the end of the volume's own year.

Claim Evidence and date Clock State
Technique is verified 25 of 34 in BTP-HM-2027-Q3; 23 on 2028-02-22; 27 on 2028-03-31 90 days for priority techniques, 180 for others Oldest passing result 81 days against 90; four more due within 30 days (9, Nordqvist, 2026)
The loop is independent IND-HM-2028-Q1, 2028-02-22 90 days, to 2028-05-22 Open at the report date; renewal needed (9, Nordqvist, 2026)
Coverage of one technique COV-HM-0112 Expires 2028-05-09 13 days left on 2028-04-26 (9, Nordqvist, 2026)
Dependency updates are reversible Revert rehearsal for MND-HM-DEV-01 90 days Lapsed 2030-12-31; class at Level 2 until the rehearsal of 2031-01-08 (12, Nordqvist, 2026)
Exit of the coding assistant Exit test 0.91 against 0.90 on 150 tasks, 2031-01-19 180 days Expires 2031-07-18 (derived) (12, Nordqvist, 2026)
Payments failover RST-HM-2027-007, 2027-10-05 180 days Expired 2028-04-02; still expired at the end of the volume's year (August 2029) (10, Nordqvist, 2026)
Claims core restore RST-HM-2029-001, 2029-06-12 (3 h 41 min against 4 h; RPO 10 against 15 min) 180 days Verified to 2029-12-09 (10, Nordqvist, 2026)
Agent verification is independent Observer declaration renewed 2029-02-11; regression suite 24 of 24 on 2029-02-13 90 days; 30 days for a regression result Current at the report date (10, Nordqvist, 2026)
Stop path works Drills of 2028-10-24 and 2029-01-17 Currency 90 days Held (10, Nordqvist, 2026)
Triage chain is accountable ACC-HM-2029-01, 2029-10-23: 17 of 20 letters traceable in four hours 90 days, to 2030-01-21 No re-run recorded (11, Nordqvist, 2026)
Board-pack summaries are faithful 28 of 30 on 2029-09-26 To 2030-03-26 No retest recorded (11, Nordqvist, 2026)
Adoption ADP-HM-2029-01 Expired 2029-11-30 First board pack showed the sampled 38 per cent, marked expired (11, Nordqvist, 2026)
Rollback and restore of the builder tool Rehearsal 12 min against 30 on 2031-11-12; restore 18 against 30 on 2031-11-14 90 days Valid to 2032-02-10 and 2032-02-12 (13, Nordqvist, 2026)
Footprint, benefit, drill, disclosure, claim and hardware records (Book 14) Footprint 90 days; benefit 180; drill 90; supplier disclosure 12 months; claim at most the reporting year; hardware record 90; review 14 days before expiry, one extension of at most 30 days As stated EBG-HM-0001 drill valid to 2033-02-27; MCR-HM-0001 to 2033-01-16 and renewed 2033-01-13; HWL-HM-0001 to 2033-02-06; BEN-HM-0002 expired 2033-05-14, one extension to 2033-06-13 (14, Nordqvist, 2026)
Takeover of the builder tool Test of 2031-12-09, 5 h 50 min item time One year Valid to 2032-12-09; remark R2 due 2032-01-13 (13, Nordqvist, 2026)
Control tests CTRL-AGT-002 tested quarterly Result at most 100 days old Rule only (6, Nordqvist, 2026)
Review of the MCP server SR-2026-077, 2026-08-19 Next 2027-02-19 Planned (3, Nordqvist, 2026)
Exceptions EXC-0004; EXC-2026-012; WAIVER-2026-007 Expire 2027-06-01; 2026-12-11; 2026-11-20 The waiver runs 92 days against a limit of 90 (2, Nordqvist, 2026)

Dashboards against samples

The series repeatedly sets a figure from a dashboard or a before-and-after measurement against a figure from a sample, and the gap is the teaching. These pairs are the dossier's clearest cases of a number that was true of an instrument and not of the organisation.

Claim Instrument figure Sample figure Source
Weekly use of embedded assistants 74 per cent (logins; June 2029) 38 per cent use of any function (1,200 of 8,100 licences; August 2029) (11, Nordqvist, 2026)
Code that is AI-assisted 42 per cent (dashboard, February 2030) 27 per cent (40 pull requests, August 2030; persistence not measured); second sample 2031-08-26 within five percentage points (12, Nordqvist, 2026)
Pay proposals and the sales score 97 per cent matched the score band, median 48 seconds, in one Dutch unit Recorded grade was Level 0 Advisory (7, Nordqvist, 2026)
Review quality Median approval 28 seconds on agent pull requests; team leader's Monday review 38 seconds against a four-minute allowance Seeded errors caught 0.86 against 0.80 in the SOC; 7 of 12 (58 per cent) in board-pack summaries (6, Nordqvist, 2026) (11, Nordqvist, 2026) (8, Nordqvist, 2026)
Verified techniques 31 green cells on a dashboard 27 of 34 verified once age is shown (9, Nordqvist, 2026)
Carbon of one batch in the cloud 0.41 t market-based in the supplier's tool 2.2 t location-based on the group's own estimate; about five times, not comparable (14, Nordqvist, 2026)
Energy of the triage agent a claim 2.3 Wh bottom-up 21 Wh top-down, about 13 times at one boundary; likely idle capacity, not measured (14, Nordqvist, 2026)
Time to complete a claim file Median 6.8 days (212 claims, September 2031) to 5.1 (198 claims, December) Before and after, no control group; valid for the month, remeasure within 90 days (13, Nordqvist, 2026)

Where the figures conflict

Subject Values Working value
Token lifetime of an agent 5 minutes (AIDR-0027, containment profile); 15 minutes (permission manifest, 2027 start) Dossier reading: 15 minutes in the permission manifest at the start, 5 after the board action of 2027-03-31; the 5-minute values of earlier volumes stay open (3, Nordqvist, 2026) (4, Nordqvist, 2026) (6, Nordqvist, 2026)
Level 3 rate limit for HM-SEC-001 5 workstations an hour; 5 targets; 10 actions; 20 actions Kept apart: each belongs to a different artefact (4, Nordqvist, 2026)
Stop target 120 seconds (HM-SEC-001); 60 seconds (emulation runner) Per system (9, Nordqvist, 2026)
Number of Mode 5 agents Two (AI Governance, AI Threat); three (a table in The Transformed Security Architect); four (its Chapter 8); a team of five SOC agents (AI Blue); six (the landscape model) Six in the model; Chapter 27 (2, Nordqvist, 2026) (5, Nordqvist, 2026) (6, Nordqvist, 2026) (8, Nordqvist, 2026) (15, Nordqvist, 2026)
Payment control Call-back and second approver above EUR 10,000 (AI Red Team); call-back from 10,000 and second approval from 100,000 (AI Threat) The later volume (4, Nordqvist, 2026) (5, Nordqvist, 2026)
Incident of stale terms 37 answers over 72 hours in Sweden; 41 hours and 312 drafts; 212 clauses Three versions kept (2, Nordqvist, 2026)
Level 3 actions of HM-SEC-001 Two (AI Red Team); three (AI Blue Chapter 13); five delegations in four rows (DLG-HM-SOC-01) Later volume (4, Nordqvist, 2026) (8, Nordqvist, 2026)
Median agent-PR approval 28 seconds; under 30; "seconds" 28 seconds (6, Nordqvist, 2026)
Measurement base of September 2032 Platform racks 62 MWh; GPU racks 49,800 kWh; about 3.4 million requests against about 1.5 million in the breakdown Open (OT-HM-76) (14, Nordqvist, 2026)
Claims-assistant population About 320 handlers in the pilot unit; about 2,900 active handlers and about 140 in the UK and Switzerland Open (OT-HM-77) (14, Nordqvist, 2026)
Language of the claims assistant Five languages; four (de, en, sv, da); evaluation slices for two Open (2, Nordqvist, 2026) (3, Nordqvist, 2026)

Summary

The numbers of the dossier fall into three layers. The profile figures, headcounts, countries and the shape of the security and operations teams, are stable across volumes and are the ground on which everything else stands. The limits and budgets are a layer of fixed numbers that change by decision and carry the record that changed them. The measurements, the drills, the samples and the restore tests, are the layer that moves, and the series gives each its own clock. Two conclusions are the book's own. First, almost every number that matters for a claim about safety or competence is a time, a drill duration or an age, not a count. Second, the series teaches by pairing a flattering instrument figure with a sample, and where the pair of the table is a count (adoption, AI-assisted code, verified techniques, board-pack catch rate) the sample or the age-adjusted figure is the lower number. A reader who quotes a figure from the dossier should quote its status and its age with it.

Sources

  1. Nordqvist, G. AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. (2, Nordqvist, 2026)
  2. Nordqvist, G. AI Platform and Security: Building and Securing AI Platforms in European Enterprises and Public Authorities. (3, Nordqvist, 2026)
  3. Nordqvist, G. AI Red Team: AI as an Active Part of Cyber Defence in European Enterprises and Public Authorities. (4, Nordqvist, 2026)
  4. Nordqvist, G. AI Threat: Are the Dystopias Coming True? Evidence, Risk and Response for European Enterprises and Public Authorities. (5, Nordqvist, 2026)
  5. Nordqvist, G. The Transformed Security Architect: Role, Craft and Career in the Age of AI in European Enterprises and Public Authorities. (6, Nordqvist, 2026)
  6. Nordqvist, G. AI Boss: When Artificial Intelligence Manages People in European Enterprises and Public Authorities. (7, Nordqvist, 2026)
  7. Nordqvist, G. AI Blue: How Artificial Intelligence Transforms the Blue Team in European Enterprises and Public Authorities. (8, Nordqvist, 2026)
  8. Nordqvist, G. AI Purple: How Red and Blue Learn Together When AI Joins the Loop in European Enterprises and Public Authorities. (9, Nordqvist, 2026)
  9. Nordqvist, G. AI Operations: How Artificial Intelligence Transforms IT Operations in European Enterprises and Public Authorities. (10, Nordqvist, 2026)
  10. Nordqvist, G. AI Leader: How Leadership Roles Change When AI Is Used Broadly in European Enterprises and Public Authorities. (11, Nordqvist, 2026)
  11. Nordqvist, G. AI Developer: How the Developer Role Changes When AI Is Used Broadly in European Enterprises and Public Authorities. (12, Nordqvist, 2026)
  12. Nordqvist, G. Vibe to Dev: From Vibe Coding to Professional Software, Step by Step, for Builders Who Are Not Developers, in European Enterprises and Public Authorities. (13, Nordqvist, 2026)
  13. Nordqvist, G. AI Environment: Designing AI for the Greatest Benefit and the Least Environmental Harm in European Enterprises and Public Authorities. (14, Nordqvist, 2026)
  14. Nordqvist, G. Landscape: A Fact-Based Picture of Architecture and AI at Work, for Leaders, Architects and Citizens. (15, Nordqvist, 2026)