Suppliers, Platforms and Sovereignty

This chapter gathers what the fourteen volumes establish about who supplies Hansa Mutual with models, cloud, hosting, observability, managed operations and engineering capacity, how the group contracts for them, how it plans to leave them, and where it places its models and data. It draws chiefly on AI Governance (Chapters 6, 7 and 16A), AI Platform and Security (Chapters 4, 12 to 15), AI Threat (Chapters 15, 21 and 26), AI Operations (Chapters 8, 24 and 29), AI Leader (Chapter 26) and AI Developer (Chapters 24 and 26), and on Landscape for the supply view of the group.

The chapter turns on one contrast that the series draws between its two organisations. Nordvik Region mostly procures, and its lever is the contract. Hansa Mutual builds, and its lever is the platform (AI Governance, Chapter 1) (2, Nordqvist, 2026) (AI Platform and Security, Chapter 24) (3, Nordqvist, 2026). The chapter shows what that means in practice: where the platform gives the group power over a supplier, and where it does not. As elsewhere in this dossier, the profile and the identifiers are established; the dates, drills and outcomes of 2027 and later are illustrative.

The group's supply map

Hansa Mutual's AI estate sits on a shared group platform operated from Hamburg and Frankfurt. Group IT runs it, funds it centrally and debits the national entities, with showback quarterly and chargeback annually (AI Governance, Chapter 16A) (2, Nordqvist, 2026). Under it lie five groups of supply, set out in Figure 6.1 and Table 6.1.

Supply map of Hansa Mutual: platform, hosting, operations suppliers and engineering suppliers

Figure 6.1: The supply map of Hansa Mutual, with the DORA register of information at the base (illustrative from 2027; sources: (10, Nordqvist, 2026), Chapter 2; (12, Nordqvist, 2026), Chapter 2).

Supply Supplier or place What it carries Contract or record First volume
Own data centre hansa-dc-frankfurt self-hosted and fallback models; the immutable backup vault and the clean-room recovery segment; the mainframe for policy administration own Books 2, 4, 12
Primary model route A general-purpose model in a hyperscaler's EU region, pinned (registry://gpai-large-eu/2026-04, eu-central) claims assistant, claims-triage agent VND-2025-118; exception EXC-0004 Books 2 and 3
Fallback model route open-weight-medium/3.1, hosted in Frankfurt fallback for the claims assistant route route claims-assistant Book 3
Cloud A EU regions (fictional placeholder) claims, customer portal, burst capacity for the group AI platform register of information Book 10
Cloud B EU regions (fictional placeholder) analytics, disaster recovery for claims core register of information Book 10
Managed operations Alster Managed Services night operations, monitoring and batch for legacy policy administration SVC-HM-003 in Germany and the Netherlands HM-PROC-MSP-2025-03 Book 10
Observability Lumenscope (software as a service) metrics, logs, traces and an AI assistant; 15 days hot, 24 months archive in Frankfurt HM-PROC-OBS-2026-01 Book 10
IT service management ServiceHub (software as a service) ticketing, with a priority suggestion in Mode 0 register of information Book 10
Recruitment screening TalentLens, Hansa People (HRIS) screening and interview support supplier, outside the EU Book 7 and Landscape
Nearshore development Wiślana Software about 110 developers framework agreement HM-PROC-DEV-2026-02 Book 12
Consulting developers two consultancies about 80 developers not stated Book 12
AI development tooling CodeMeridian (IDE assistant) 340 licences from January 2030 HM-PROC-AIDEV-2029-01, LIC-HM-DEV-01, exit EXIT-005 Book 12

Table 6.1: The suppliers and places of supply of Hansa Mutual (sources: (3, Nordqvist, 2026), Chapters 4 and 12; (10, Nordqvist, 2026), Chapter 2; (12, Nordqvist, 2026), Chapter 2; (7, Nordqvist, 2026), Chapter 6).

The landscape model gives the same supply in outline. Thirteen of the 17 units sit inside the group's own boundary, and four rest on suppliers outside the EU: conversational assistants, recruitment screening, embedded workplace assistants and ambient AI. A "model chip" to a model provider outside the EU is attached to the claims assistant and to the agentic work tools, both of which depend on the exception EXC-0004 and the origin assessment GEO-2026-017. The United Kingdom and Swiss subsidiaries and the Polish centre are defined as supplier-like entities in the model with the role of operations, but no unit refers to them, so the supply view draws no containers for them, which is itself an open thread (Landscape, Chapters 4 and 16) (15, Nordqvist, 2026).

Model origin and the exception

Hansa's models come in three kinds, and the volumes treat them differently.

A hosted model from a provider, used through the group gateway. The claims assistant (AIDR-0021) runs on a model in a hyperscaler's EU region. The group accepted that a hosted model could not meet its principle P3 (reproducibility and lineage), and recorded the gap in a time-limited exception, EXC-0004, approved on 2026-06-02 and expiring on 2027-06-01, with compensating controls instead of ignoring the principle (AI Governance, Chapter 4) (2, Nordqvist, 2026). The route is pinned to a model version, uses no semantic cache, stores prompts masked, forbids non-EEA regions, and keeps logs for 3,650 days in the group and in market-specific places for the United Kingdom (uk-south) and Switzerland (ch-north) (AI Platform and Security, Chapter 12) (3, Nordqvist, 2026).

A self-hosted open-weight model. The assessment GEO-2026-017 of 2026-10-20 concerns a model of Chinese origin (named Example-Coder-32B in the volume), developed outside the EU and run self-hosted. The group approved it with controls (egress denied, no tools) although it would have refused a hosted service from the same developer. The lesson the volume draws is that origin matters less than how the model is run; the review is due again on 2027-04-20 (AI Governance, Chapter 6) (2, Nordqvist, 2026). The agentic work tools in Book 3 use the same kind of model under the name code-32b, with a distinct gateway hostname for the protected workspace (AI Platform and Security, Chapter 15) (3, Nordqvist, 2026).

A model the group builds on or trains. The underwriting-support model for life and health is a high-risk system under Annex III point 5(c), and Hansa is its provider (AI Red Team, Chapter 21) (4, Nordqvist, 2026). It is not a general-purpose model, and the group delivers none to others (AI Threat, Chapter 19) (5, Nordqvist, 2026).

Where models and data are placed

The placement decisions of Book 3 are the clearest case of the platform as lever. For the summarisation of claims correspondence (about 40,000 claims a month, four languages, about one claim in five with health data) the group scored a cloud model at 40 and a self-hosted model at 51 on a weighted score; quality was 0.91 against 0.88 (cloud against self-hosted) with a threshold of 0.85, and utilisation 45 per cent on two nodes. The decision was a self-hosted baseline for all claims, with cloud burst only for claims without health data, under rules P1 and P3, with a review after twelve months (AI Platform and Security, Chapter 14) (3, Nordqvist, 2026). The agentic work tools classify the workspace as standard, confidential or protected; the cloud route is for standard only, the Polish centre's claims data is confidential, life and health data under GDPR Article 9 is protected, and the United Kingdom and Switzerland need a transfer assessment (AI Platform and Security, Chapter 15) (3, Nordqvist, 2026).

Capacity follows the placement. dc-frankfurt-l holds eight reserved accelerators for interactive and fallback inference; eu-cloud-m holds four reserved and twelve burst; agents run at most 150 concurrent tasks. Peak utilisation was 0.58 for the data centre and 0.71 for the cloud partition; the platform switches to a small model above 0.80 (AI Platform and Security, Chapter 13) (3, Nordqvist, 2026).

Contracts, tenders and the register

The group is a financial entity under DORA in its EU entities. Its register of information (Article 28(3)) covers its AI providers, its exit strategies follow Article 28(8) and its contract clauses follow Article 30 (AI Governance, Chapters 6 and 7) (2, Nordqvist, 2026). Three examples show what that looks like.

The contract review of 2026-09-10 (draft version 3) of the exit strategy for "Model Provider Alpha" found, against the Article 30 clause list, four clauses missing and eight with gaps. The same strategy had an exit rehearsal on 2026-06-12 against Provider Alpha: an exit time of four weeks, a moderation switch time of two weeks, a maximum tolerable outage of 72 hours, an evaluation score of 0.93 against 0.92 on the original, and eleven prompt templates changed (AI Governance, Chapters 6 and 7) (2, Nordqvist, 2026).

The document extraction renewal (DEC-2026-044, 2026-11-11, use case UC-2024-012) is the worked case of AI Governance. Document extraction was 6 per cent of AI spend, its hosted service lacked a tested exit plan, and the contract renews before 2027-03-31. The board compared four options on the same quality constraint. Renewing the hosted service cost EUR 1.9 to 2.4 million over three years, with price rises capped at 8 per cent a year and a unit cost of EUR 0.11. An open-weight model on the group platform cost EUR 2.1 to 2.9 million, at a unit cost of 0.09 at full utilisation and 0.19 at pilot volume. A rules-based route for structured forms with a hosted model for the rest cost EUR 1.5 to 1.8 million, at 0.04 and 0.14. Manual extraction cost EUR 3.1 to 3.4 million, at 0.52. The decision was hybrid: scale the rules for structured forms (70 per cent of volume), run a bounded open-weight pilot of at most 15 per cent of unstructured volume, renew the hosted service for twelve months with the price cap, and reject doing nothing. The conditions were the exit plan EXIT-007 completed and rehearsed by 2027-02-05 (owner the head of procurement), the pilot boundary by the same date, and signature of the renewal on 2027-02-09 after the first condition; review on 2027-02-17, merged through PR-1263. The lesson is that missing evidence yields "approve with conditions", not refusal (AI Governance, Chapter 16A) (2, Nordqvist, 2026).

The managed services contracts are the group's weakest seams. Alster Managed Services carries night operations for the legacy policy administration SVC-HM-003 and embeds AI in its own monitoring (Mode 0); the group cannot place a policy point inside it. Lumenscope carries observability for critical or important functions and is in the register. In May 2027 its ingestion failed for 3 hours 10 minutes, with 14 services unobserved (INC-HM-2027-0572), and in April 2029 an outage of 1 hour 50 minutes (INC-HM-2029-0214) left the payment gateway and customer portal without an independent probe, so that the service desk in Poland was the only observer of the portal for 40 minutes (AI Operations, Chapters 2, 8 and 29) (10, Nordqvist, 2026). AI Operations concludes that the group places both suppliers under annex conditions before any widening of their use (AI Operations, Chapter 30) (10, Nordqvist, 2026).

AI Developer adds the engineering supply. The IDE assistant contract carries three clauses: no training on the group's code, processing in the EU and 30 days' notice of a model change. Its exit test, RST-HM-2031-0002 on 2031-01-19, scored 0.91 against 0.90 on 150 tasks; the evidence expires after 180 days, which falls on 2031-07-18. The classification under DORA of the nearshore partner Wiślana is open, and the group applies the strictest reading until it is settled (AI Developer, Chapters 24 and 26) (12, Nordqvist, 2026). A last supply fact is a small one: in the summer of 2031 a builder used a hosted application builder for a prototype with about 40 rows of real claims export data, with a link open to anyone and no contract; the builder closed it herself on 2031-09-12 and informed the data protection officer (Vibe to Dev, Chapter 2) (13, Nordqvist, 2026).

Exit plans

Hansa treats an exit plan as a claim that expires. Table 6.2 collects the plans and drills the volumes record.

Exit plan Subject Drill or test Result Open point
EXIT-004 classification of claims correspondence: a primary US-provider model in a hyperscaler EU region, a European open-weight second source in Frankfurt April 2026 9 hours; GPU quota too low none
EXIT-004 same October 2026 4 hours; 6 replicas; about 40 minutes; second source 0.89 against 0.94, threshold raised from 0.80 to 0.90 when switching the 90-day plan HM-W4-second-source asks for a switch to the Frankfurt fallback within 48 hours, and the volumes do not link the two numbers
Model Provider Alpha strategy hosted model provider 2026-06-12 rehearsal exit 4 weeks, outage tolerance 72 hours, evaluation 0.93 against 0.92 contract: 4 clauses missing, 8 gaps; link to EXIT-004 unclear
EXIT-007 hosted document-extraction service (UC-2024-012); owner the head of procurement; draft due 2026-12-15 rehearsal due 2027-02-05 planned renewal signature 2027-02-09 follows the rehearsal
Fallback evaluation test five working days, one market, Frankfurt models date not stated claims assistant acceptable; claims-triage agent limited to simple routes the evaluation check that approved the fallback called the primary provider's model; removed (AI Threat, Chapters 26 and 30) (5, Nordqvist, 2026)
Self-hosted second model for claims-triage CAP-HM-AI-001 test 2028-10-26 pass at 0.91 none
Cloud B model fallback HM-PROC-CLB-2028-04, AIDR-0974 of 2029-02-20 the competent authority informed on 2029-02-27 before first production use informed none
EXIT-005 IDE assistant RST-HM-2031-0002, 2031-01-19 0.91 against 0.90 on 150 tasks; evidence ends 2031-07-18 cadence of the exit test is decided by the CRO in the second quarter of 2031

Table 6.2: Exit plans, drills and results (illustrative; sources: (2, Nordqvist, 2026), Chapters 6, 7 and 16A; (3, Nordqvist, 2026), Chapter 25; (5, Nordqvist, 2026); (10, Nordqvist, 2026), Chapters 15 and 29; (12, Nordqvist, 2026), Chapter 26).

The common finding is the one AI Governance states: exit plans only work if they are rehearsed. The first EXIT-004 drill in April 2026 took nine hours and exposed a GPU quota too low for the second source, and the second drill in October switched in four hours (AI Governance, Chapters 6 and 23) (2, Nordqvist, 2026). The correlated-check lesson of AI Threat adds a second: a check that approves a fallback by calling the primary supplier verifies nothing, and the group removed it (AI Threat, Chapter 26) (5, Nordqvist, 2026).

Concentration and sovereignty

Four positions of the volumes make up the group's sovereignty posture.

  1. The group holds the sovereign layer itself. Self-hosted baselines and fallback models in Frankfurt, an immutable vault and a clean-room recovery segment with its own identity provider give the group something to fall back on that no supplier can switch off (AI Red Team, Chapter 16) (4, Nordqvist, 2026).
  2. The group does not let concentration sit unexamined. THR-HM-06 records the correlated failure of a shared model supplier (expected, with an incident in the wild as the evidence rung), and the supplier-exposure record SXR-HM-2026-03 of 2026-11-02, owned by the head of the group AI platform, names a "Provider A". The register states that the second source is not equivalent for complex claims routes, which links THR-HM-06 to the erosion of skills in THR-HM-08 (AI Threat, Chapters 21, 32 and 33) (5, Nordqvist, 2026). The group's evidence request to a model provider is SER-HM-2026-01 (AI Threat, Chapter 29) (5, Nordqvist, 2026).
  3. The fallback and the restore copy must not rest on the same hyperscaler. AI Leader records the position, grounded in DORA Article 29(1): EXIT-004 is to be run again, a quality gap of 0.05 for the fallback is accepted, and the board reads the exit result in the source, not in a summary (AI Leader, Chapter 26) (11, Nordqvist, 2026).
  4. Origin is a control question, not a verdict. GEO-2026-017 is the case in point; the same developer's hosted service would have been refused (AI Governance, Chapter 6) (2, Nordqvist, 2026).

The concentration that no volume resolves is Lumenscope. A supplier of observability is, by construction, observing the thing that would reveal its own failure; AI Operations answers with an independent observer manifest (OBS-HM-0001) whose probes and heartbeat receiver sit outside it, and Chapter 7 describes it (AI Operations, Chapter 17) (10, Nordqvist, 2026).

Across the usage modes

Supply and sovereignty differ by mode in a direct way: the higher the mode, the more the group can hold itself.

Mode Where supply lies What the group can enforce
0 Ambient AI Supplier's product (Alster's monitoring, Lumenscope default anomaly detection, ServiceHub suggestions) Registration, an owner, notice clauses in the DORA register; no policy point inside the product
1 and 2 Enterprise tiers and embedded assistants from suppliers outside the EU Acceptable-use policy, tenant policy, proxy blocking of consumer tools; interim agreement with the works council for the IDE assistant
3 Agentic work tools Gateway and sandbox owned by the group, models by classification route Routing by data class; sensitive material only on self-hosted models
4 Group-built systems on supplier or self-hosted models Gateway route, pinned model versions, evaluation gates, exit plan per system
5 Group-built agents with their own identities on group infrastructure Manifest, policy point and stop path owned by the group; model route replaceable

Table 6.3: Supply and sovereignty across the usage modes (sources: (3, Nordqvist, 2026), Chapters 12 and 15; (10, Nordqvist, 2026), Chapter 2; (12, Nordqvist, 2026), Chapter 2).

Hosting and supply for energy and carbon

AI Environment adds the supply questions that arise from where the group's AI runs. Its reading is illustrative and dated September to November 2032 (14, Nordqvist, 2026).

Supply What Book 14 gives Disclosure the group holds
Own Frankfurt data centre About 2.8 MW installed IT power, PUE about 1.38, 12 GPU servers with 96 accelerators installed in March 2029, planned life five years, utilisation 0.52 against a target of 0.70 SDR-HM-0001, 8 of 10 questions answered
Cloud A (placeholder: Ireland) Burst up to 64 accelerators on demand, plus claims and portal workloads; customer-level carbon figures are market-based, monthly, lagged and on a partly documented boundary SDR-HM-0002, 3 of 10 answered, received 2032-11-08; a re-scoring batch of about 9,500 accelerator hours reads 0.41 t market-based in the supplier's tool and 2.2 t location-based on the group's own estimate
Cloud B (placeholder: France) Analytics and disaster recovery of claims core Not separately disclosed
Regional provider with a Nordic site Compared in a 90-day pilot under AIDR-1369 to 2033-02-13 SDR-HM-0003, 6 of 10 answered; PUE 1.25 and 60 g/kWh are assumed
Hosted model providers Non-sensitive routes through the gateway; Annex XI estimates and per-request figures partial Requested, partial

Table 6.4: Supply and disclosure for Hansa Mutual's AI compute (illustrative; source: (14, Nordqvist, 2026), Chapters 19 and 23).

On the comparison sheet HCS-HM-2032-11 (Jan Berger, reviewed by Dr. Verena Lindner), with Frankfurt as 100 for 32 accelerators, a hyperscale burst scores 65, the regional provider 15, and a burst at 0.30 utilisation 95; these are assumptions that Book 14 marks as such. The decision AIDR-1369 (2032-11-15) keeps Frankfurt as the base, allows Cloud A only as burst under a monthly cap in billed accelerator hours (EBG-HM-0006) and makes no hosting choice or exclusion on environmental grounds before the answers arrive. Procurement (Stefan Wiedemann) attaches the clause family HM-PROC-ENV-01 to -10 to each renewal; the first agreement under it, in June 2033 at the renewal of the Cloud A commitment, accepts ENV-01 and ENV-04 and defers ENV-02 (energy for capacity), so that Cloud A stays on rung 3 for energy until 2034. A supplier's benchmark for a Frankfurt hardware refresh shows less gain than the 2.5 times the group had assumed and no order is placed (14, Nordqvist, 2026).

Budget as a measure of supply

The volumes give a few figures that show the weight of supply in the group's cost base, all illustrative. Group IT costs about EUR 310 million a year; infrastructure and operations about EUR 140 million; cloud consumption about EUR 48 million; Lumenscope about EUR 6.5 million; the managed service provider about EUR 11 million; AI inference and platform about EUR 9 million (AI Operations, Chapter 2) (10, Nordqvist, 2026); AI Environment gives about EUR 17 million for AI inference and platform in the 12 months to August 2032 (illustrative) (14, Nordqvist, 2026). The route claims-assistant has a budget of EUR 18,000 a month, with an alert at 0.8 and a hard stop at 1.1 (AI Platform and Security, Chapter 12) (3, Nordqvist, 2026). The triage agent's capacity record CAP-HM-AI-001 caps cost per triaged claim at the 99th percentile of EUR 0.12 and the month at EUR 38,000, with Cloud A burst at most 12 per cent (AI Operations, Chapter 15) (10, Nordqvist, 2026). Showback for the third quarter of 2026 was EUR 818,000 in total, with a fixed control-plane charge of EUR 180,000 a quarter split by entity (DE 34, NL 14, SE 12, PL 10, NO 10, UK 12, CH 8 per cent) (AI Governance, Chapter 16A) (2, Nordqvist, 2026).

Open threads

  • Exit plans. AI Governance describes EXIT-004 (hyperscaler region with an open-weight second source) in Chapter 6 and a different exit strategy for "Model Provider Alpha" in Chapter 7; the link is unclear. Book 3's 90-day target of a 48-hour switch has no stated relation to the 9-hour and 4-hour drills.
  • Providers A and Alpha. AI Threat uses a fictional "Provider A" for the supplier-exposure record and "Model provider A" for the model under the triage agent; whether they are the same supplier, and whether either is "Alpha", is not stated (AI Threat, Chapters 15, 21 and 29) (5, Nordqvist, 2026).
  • DORA flag. AI Governance sets dora: law in the group charter; AI Platform and Security places the DORA flag in each market overlay and not in the baseline (AI Platform and Security, Chapter 24) (3, Nordqvist, 2026).
  • NIS 2. AI Leader holds that NIS 2 is displaced for EU financial entities; AI Operations says application depends on national transposition (AI Operations, Chapter 2) (10, Nordqvist, 2026) (AI Leader, Chapter 24) (11, Nordqvist, 2026).
  • Register entries for the builder lane. Whether the builder lane's assistant has its own entry in the register of information is open (Vibe to Dev, Chapter 19) (13, Nordqvist, 2026). The same holds for the nearshore partner.
  • Same number, two things. The placement budget EBG-HM-0004 (9,500 kWh a month at facility boundary) and the Cloud A re-scoring batch (about 9,500 accelerator hours) share a figure and no unit (OT-HM-80) (14, Nordqvist, 2026).
  • Notice clause. AI Platform and Security notes a missing notice clause for Mode 0 at the service desk, with no counterpart elsewhere (AI Platform and Security, Chapter 24) (3, Nordqvist, 2026).
  • Aged exit evidence. The exit test RST-HM-2031-0002 passed on 2031-01-19 and its evidence ended on 2031-07-18; no later volume records a retest.

Summary

Hansa Mutual contracts less than Nordvik Region and builds more, and the lever that follows is the platform. It places models by data class, keeps a self-hosted baseline and a fallback in Frankfurt, puts every supplier in a register and holds exit plans as claims that expire. The weak points are the suppliers whose products the group cannot instrument (ambient AI, the managed operator, the observability service), the exit plans that were rehearsed once and not again, and a group of identifiers and numbers that disagree between volumes. The chapter's own conclusion is that Hansa's sovereignty is a matter of evidence age. The group has the means to leave a supplier; whether it can show, today, that it could, is what each drill dates.

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. AI Boss: When Artificial Intelligence Manages People in European Enterprises and Public Authorities. (7, Nordqvist, 2026)
  6. Nordqvist, G. AI Operations: How Artificial Intelligence Transforms IT Operations in European Enterprises and Public Authorities. (10, Nordqvist, 2026)
  7. Nordqvist, G. AI Leader: How Leadership Roles Change When AI Is Used Broadly in European Enterprises and Public Authorities. (11, Nordqvist, 2026)
  8. Nordqvist, G. AI Developer: How the Developer Role Changes When AI Is Used Broadly in European Enterprises and Public Authorities. (12, Nordqvist, 2026)
  9. 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)
  10. Nordqvist, G. AI Environment: Designing AI for the Greatest Benefit and the Least Environmental Harm in European Enterprises and Public Authorities. (14, Nordqvist, 2026)
  11. Nordqvist, G. Landscape: A Fact-Based Picture of Architecture and AI at Work, for Leaders, Architects and Citizens. (15, Nordqvist, 2026)