Book 7, AI Boss: Managing People

This chapter gathers what AI Boss (Book 7) establishes about Hansa Mutual: an organisation whose AI boss is mostly built already and directs people today, the five management functions as they apply to claims handlers, recruiters and salespeople, the decisions and records that govern them, and the negotiated relationship with the Konzernbetriebsrat. It draws on the volume's people-management profile, its list of facts established in drafted chapters, its chapters 1 to 33 and its Appendix A. Everything dated after November 2026 is illustrative fiction, as the volume itself marks it.

The volume in brief

AI Boss asks what happens when artificial intelligence takes the manager's chair: directing, pacing, scoring, rewarding, disciplining and dismissing people, and when people in turn become the managers of AI agents. Its thesis is that AI may direct work but may not, on its own, decide about people. The practice, algorithmic management, splits into five management functions: M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care. The functions that carry consequences for a person always have a named human decider who can be identified, contested and held to account. The manager's role is transformed, not abolished, from supervisor of tasks to accountable leader of people and conductor of agents. Level 4 is never used for M3 Consequence (7, Nordqvist, 2026).

The role of Hansa Mutual in the volume

Hansa Mutual is the volume's case for an AI boss that is mostly built and already manages people. Four things carry the case. The claims-triage agent (Mode 5) allocates claims to individual handlers in Germany and produces throughput figures per handler. TalentLens, a procured recruitment tool, ranks applications and is a high-risk system under Annex III 4(a). A sales-performance score is one of several inputs to variable pay. From January 2027 a pilot, HM-AB-PILOT-01, has two teams, the Hamburg claims team under Jana Petersen and an IT service desk team, managing Mode 5 agents themselves. In autumn 2026 the executive board asked the chief HR officer and the chief risk officer for a group policy for AI in management and invited the Konzernbetriebsrat to review it (AIDR-0650, 2026-11-23, first report 2027-04-30) (AI Boss, Chapters 1 and 30) (7, Nordqvist, 2026).

The volume describes Hansa Mutual's approach as correction-first (Chapter 30) and co-determination-first (Chapter 33): find what the agent already does that nobody decided, write it into profiles, negotiate classes of change with the Konzernbetriebsrat and measure practice against recorded degree. The binding constraint is co-determination across seven markets, which means one group works council, the European Works Council and six other national arrangements. The profiles moved inward: the same degrees of delegation with more conditions written in (AI Boss, Chapters 30 and 33) (7, Nordqvist, 2026).

What the volume establishes

The workforce and the HR estate

Headcount is about 9,500 and unchanged from the earlier volumes. Claims has about 3,600 people, of whom about 1,400 work at the Polish service centre; sales and distribution about 2,100; underwriting about 1,100; IT about 1,300; corporate functions about 1,400. By country: Germany about 4,600, Poland 1,700, the Netherlands 900, the UK 800, Sweden 600, Switzerland 500, Norway 400. Claims handlers work in teams of 15 to 25 under team leaders. Salespeople, including employed field sales and the contact centre, have variable pay. The group chief HR officer sits on the executive board and leads group HR in Hamburg, with an HR lead in every national entity. The HR systems are Hansa People (a cloud HRIS on a commercial suite), workforce management for contact centres and claims, a group applicant tracking system containing TalentLens (screening and a video interview module), the performance cycle in the HRIS and a group learning platform with AI literacy and oversight training localised per entity. The volume notes that there is no Austrian entity (AI Boss, Chapter 6) (7, Nordqvist, 2026).

The management systems

System Mode Function Facts the volume gives
Claims-triage agent (AIDR-0024, AIDR-0034, AIDR-0063) Mode 5 M1 Direction Allocates claims to individuals within allocation rules v2026.4, enforced by the claims system's rule engine outside the agent; classed not high-risk for its insurance function and reclassified as high-risk for its allocation function under Annex III 4(b) in AIDR-0651 (2026-12-14); DLG-HM-CTA-01 applies in Germany only; at the Polish service centre team leaders allocate from queues
TalentLens Mode 4, procured M3 Consequence input High-risk under Annex III 4(a); used in Poland for service centre roles and in Germany for graduate and apprentice recruitment; affect and voice analysis is switched off group-wide, including the UK and Switzerland, under Article 5(1)(f) and checked after every release; recruiters decide
Sales-performance score (DLG-HM-SPS-01) Mode 4, built on the group platform M2 Evaluation Conversion, retention and complaint indicators; one of several inputs to variable pay
Pilot agents (wl-hm-clm-intake, wl-hm-clm-recovery) Mode 5 Agents managed by people Document intake agent and recovery agent that writes subrogation letters, under mandate MND-HM-CLM-01; the team leader is responsible for the output
Office assistants Mode 2 Ordinary use
Wellbeing and collaboration analytics Mode 0 M5 Care In the licensed office suite; switched off group-wide pending an agreement (FTR-HM-CA-01, AIDR-0670)

The starting delegation table is: M1 claims allocation at Level 3 Bounded; M1 pilot at Level 2 for the agents' own actions; M2 throughput and sales score at Level 1; M3 TalentLens at Level 1 and the sales score at Level 0 as a pay input; M4 learning platform recommendations at Level 0; M5 none (AI Boss, Chapters 6 and 30) (7, Nordqvist, 2026).

People and representation

Two named people appear. Jana Petersen is the claims team leader in Hamburg, with about 20 handlers (motor and property) and twelve years in claims settlement. From January 2027 she also manages the two pilot agents, she overrides the allocation agent in MDR-HM-0001, and from 1 July 2027 her role charter ROLE-HM-MGR-01 (AIDR-0683) protects 35 per cent of her time for M4 Development and M5 Care (0.18 plus 0.17), with about two hours a week for agent verification (0.08 conducting agents). Thomas Brandt has chaired the Konzernbetriebsrat since 2022, is a former underwriter, negotiated the group IT agreement KBV-IT-2024 and sits in the European Works Council. On 2026-10-12 he requested the TalentLens selection criteria as a BetrVG §95 guideline, a quarterly contest report (AIDR-0681), team-level figures only and an empty list of M3 decisions that throughput may feed (AI Boss, Chapters 1, 20 and 30) (7, Nordqvist, 2026).

Unnamed roles include the group head of claims operations (owner of the claims-triage agent, chain owner and owner of DLG-HM-CTA-01), the group head of sales HR (owner of DLG-HM-SPS-01), an HR business partner for sales who reviews outside the line, the recruitment lead, the HR lead Germany, the agent operator on call, a deputy team leader with two years in the job, nine recruiters at the Polish service centre, a Dutch field salesperson who contested their score (CNT-HM-0001) with a works council member as support, the chief risk officer, the group AI governance board and the executive board.

The representation structures in the profile are the Konzernbetriebsrat and local works councils (BetrVG §87(1) no. 6 and §§94, 95, 99, 102), KBV-IT-2024, the European Works Council (Directive 2009/38/EC), the Dutch ondernemingsraad (consent under WOR Article 27(1)(k) and (l)), the Swedish unions (MBL), Norway's working environment act rules on control measures, the Polish company union and employee council (monitoring rules in the collective agreement or work regulations; Labour Code Articles 22² and 22³; staff informed two weeks ahead), the UK staff forum and the Swiss employee committee. The annex to KBV-IT-2024 on the claims agent's figures makes individual throughput visible only to the handler and the team leader, forbids ranking lists, forbids use as the only basis for grade, pay or discipline, sets retention at 13 weeks, requires an access log with quarterly review, makes any new metric or recipient subject to consent, and limits Switzerland to team figures and Poland to the work rules (AI Boss, Chapters 6 and 13) (7, Nordqvist, 2026).

The group AI Act baseline for all seven markets is AIDR-0657: law in the EU entities, policy in Norway, the UK and Switzerland. Hansa Mutual is the provider of the allocation function (AI Boss, Chapters 21 and 32) (7, Nordqvist, 2026).

Stop rights

Stop rights are distributed by role and measured in minutes: the team leader pauses their own team at once; the agent operator stops allocation to people within 15 minutes, with a fallback to a team leader queue; Jana Petersen can revoke the pilot agents' identities within 5 minutes; the chief HR officer can stop any management system group-wide within 30 minutes; and the chain stop path SPT-HM-CHN-01 takes 10 minutes (AI Boss, Chapters 16, 17 and 20, Appendix A) (7, Nordqvist, 2026).

Stop rights over Hansa Mutual's management systems, by role and time limit, with the rules engine placed outside the claims-triage agent.

Figure 13.1: Stop rights over the claims-triage agent and the pilot agents, from DRF-HM-01 and SPT-HM-CHN-01 (illustrative, effective from 2027-04-15).

Decisions and records

Decision records

Record Date Subject
AIDR-0650 2026-11-23 The executive board's request; first report 2027-04-30
AIDR-0651 2026-12-14 Reclassification of the allocation; throughput-based capacity factor removed in release 7.0
AIDR-0652 December 2026 Terms of the pilot
AIDR-0653 2026-12-08 Disparate-impact test on every release
AIDR-0654, AIDR-0655 Group monitoring principle with entity overlays; DRF-HM-01
AIDR-0656 Managerial judgement in the team leader framework: 10 per cent of allocation unassisted (numbered AIDR-0670 in Chapter 14 before renumbering)
AIDR-0657 2027-04-30 AI Act baseline
AIDR-0670 2026-12-07 Analytics switched off, FTR-HM-CA-01
AIDR-0671 Chain record CHN-HM-CLM-01
AIDR-0674 Evaluation plan EVP-HM-AB-2027-01
AIDR-0675 The fewest-open-claims rule disclosed; daily cap per handler
AIDR-0676 2027-02-22 Article 22 analysis of TalentLens and the allocation
AIDR-0677 2026-12-14 Change classes for KBV-IT-2024
AIDR-0678 2027-03-31 Equality block in DLG-HM-REC-01
AIDR-0679 Management decision record agreement with the Konzernbetriebsrat; register v1.1.0
AIDR-0680, AIDR-0681 April 2027; 2027-01-25 TalentLens reading rule; contest report and DLG-HM-SPS-01 version 4
AIDR-0682 2026-12-21 Notices and transparency register
AIDR-0683, AIDR-0685, AIDR-0686 2027-06-28; 2027-04-30; 2027-11 Team leader charter and horizon register; claims allocation v1.2.0; MND-HM-CLM-01 to 2028-06-30
AIDR-0690 2027-10-01 24-month plan

Earlier decisions the volume builds on are AIDR-0021 (claims assistant), AIDR-0024, AIDR-0034 and AIDR-0063. Delegation profiles go through versions: DLG-HM-CTA-01 1.0.0 (2026-12-14), 1.1.0 (February 2027, with PREF-01) and 1.2.0 (2027-04-30); DLG-HM-REC-01 1.3, 1.4 (2027-03-31) and 1.5 (April 2027); DLG-HM-SPS-01 2, 3 (after CNT-HM-0001) and 4 (2027-01-25); and DLG-HM-AB-01 (AI Boss, Chapters 14, 30 and 32, Appendix A) (7, Nordqvist, 2026).

The decision rights file and change classes

DRF-HM-01 (AIDR-0655) takes effect on 2027-04-15 and is reviewed by 2028-04-14. It caps M3 at "any: L2_Supervised, adverse: L1_Assisted, never L4". Its classes are HM-M1-CLAIM-ALLOCATE, HM-M2-THROUGHPUT (may_feed is an empty list), HM-M3-HIRE-SCREEN (at most 12 reviews per hour) and HM-M3-VARIABLE-PAY (Level 0). Stop rights in the file are team leader reassign at once, chief HR officer 30 minutes and agent operator 15 minutes.

The negotiation of AIDR-0677 is the volume's most reusable design. Hansa Mutual negotiated how change is governed and not every version. Class 1 changes need the Konzernbetriebsrat's consent: a higher degree of delegation, behavioural inputs, a new individual metric or recipient, or any M3 use. Class 2 changes are decided by the joint IT committee within 10 working days: the daily cap of 8 to 20 per handler, and model releases that pass the allocation and disparate-impact tests. Class 3 changes are information only. The owner cannot lower a class and the works council can raise one (AI Boss, Chapters 20 and 30) (7, Nordqvist, 2026).

Records of individual decisions

MDR-HM-0001 (2027-01-12) records Jana Petersen reassigning five of nine motor claims the agent had given to a handler on an agreed reduced schedule, because the agent had read only open workload. No health data was involved; a capacity field filled in by a team leader without reasons becomes the first Class 2 change; retention is 24 months (AIDR-0679) and review outside the line takes place within 15 working days. MDR-HM-0070 (2027-02-15) records a Polish recruiter who read 64 of 212 ranked applications, 45 of them below the cut, selected number 131 and rejected number 9. CNT-HM-0001 was opened on 2026-10-19, explained on 2026-10-23, decided on 2026-11-02 and closed on 2027-01-15. Other records are DIT-HM-CTA-01, DIT-HM-REC-01, RVT-HM-CTA-2027-01 (2027-03-15 to 03-19), OQM-HM-CLM-01, OQM-HM-REC-01, OVR-HM-REC-2027-Q1-PL, FMAP-HM-CLM-HH-01, TCH-HM-CLM-01 (effective 2027-01-01), WIN-HM-CTA-DE (version 1.0, 2027-01-08, and 1.1), the transparency entries TRG-HM-CTA-DE, TRG-HM-TL-DE, TRG-HM-SPS-DE and TRG-HM-OSA-DE, HR-REC-CRIT-2026-01 (selection criteria), the rule CTA-R-14, PREF-01, HZN-HM-MGR-01 and delegation register version 1.1.0 (1.2.0 in Appendix A) (AI Boss, Chapters 12, 13, 26 and 30, Appendix A) (7, Nordqvist, 2026).

Events and figures

Established in the profile: headcount, systems, representation structures and the earlier decisions. Illustrative: the 2027 sequence and all measured figures.

Date Event
2026-05-12 Dutch ondernemingsraad consents to the sales score
2026-05-20 European Works Council informed; TalentLens annex
2026-06-01 Article 26(7) notices for TalentLens issued in advance
2026-10-12 Brandt's BetrVG §95 request
2026-10-19 to 2026-11-02 CNT-HM-0001 opened and decided
2026-11-23 AIDR-0650
2026-12-07, 12-08 FMAP-HM-CLM-HH-01 and AIDR-0670; DIT-HM-CTA-01 to the Konzernbetriebsrat
2026-12-14 AIDR-0651, release 7.0, DLG-HM-CTA-01 1.0.0, change-class annex (AIDR-0677)
2027-01-01, 01-08, 01-12 TCH-HM-CLM-01; WIN-HM-CTA-DE; MDR-HM-0001
Second half of January 2027 Pilot agents start
2027-01-25, 02-15, 02-22 AIDR-0681; MDR-HM-0070; AIDR-0676
2027-03-15 to 03-19 RVT-HM-CTA-2027-01
2027-03-31 AIDR-0678, DLG-HM-REC-01 1.4
2027-04-15 DRF-HM-01 in force
2027-04-30 First report under AIDR-0650; DLG-HM-CTA-01 1.2.0 (AIDR-0685); AIDR-0657
May 2027 CHN-HM-CLM-01 (AIDR-0671) and first SPT-HM-CHN-01 drill
2027-06-28, 06-30, 07-01 AIDR-0683; reversibility conditions met; ROLE-HM-MGR-01 effective
July 2027 MND-HM-CLM-01 1.1.0: recovery agent's budget cut from 40 to 25 letters a day until September after median approval time stayed under 30 seconds for three weeks
October 2027 EWC informed of a proposed Polish phase
November 2027 AIDR-0686: mandate to 2028-06-30
2027-12-01 Annual review of the change-class annex falls due (open at year end)

Planned: Article 26 obligations apply to Annex III systems from 2027-12-02 (including 26(7)), after the Digital Omnibus, Regulation (EU) 2026/1744; renew or end the mandate by 2028-06-30; next RVT in March 2028; DRF-HM-01 review 2028-04-14; a 12-month pilot evaluation; the plan runs 2027-10-01 to 2029-09-30; the Polish phase is a plan only (AI Boss, Chapters 28, 30 and 32) (7, Nordqvist, 2026).

Figures

All are illustrative. Disparate impact in 2026 Q3 (minimum cell 20) was 0.71 (CI 0.64 to 0.78) for complex claims to part-time staff and 0.69 (0.60 to 0.79) for part-time women, caused by rule CTA-R-14, which required two consecutive working days left in the week. In one Dutch sales unit 97 per cent of pay proposals matched the score band, with a median of 48 seconds; the practice test of DLG-HM-SPS-01 reads "decisive in practice in one unit", and plan item H3 removes the score as a pay input where deviations are below 0.05. For TalentLens in Poland in Q1 2027, nine recruiters handled 3,840 ranked applications; the median time to reject below the cut was 38 seconds; 21 of 150 blind-reviewed applications would have advanced; 3 of 6 seeded cases were caught; and recruiters did 19 reviews per hour against the cap of 12. In Q4 2026 recruiters departed from the ranking in 11 per cent of shortlists across 41 posts and read below the cut for 95 per cent of the posts; display of the ranking is paused where deviation is below 0.05 for two quarters. The conditions of DLG-HM-REC-01 version 1.3 are to read at least 20 per cent below the cut, blind-review 50 rejected applications a month, keep logs for 12 months and explain within 30 days.

On reversibility, the deputy needed three times as long as the agent-assisted allocation, complex claims concentrated on four experienced handlers, and from April 2027 10 per cent of new claims are allocated by the deputy without the agent. Allocation limits are a daily cap of 8 to 20 per handler (the Class 2 bounds) and PREF-01, at most 30 per cent of a week's allocation and one complex claim a week for handlers in their first two years. The pilot thresholds in MND-HM-CLM-01 are seeded catch 0.85, rejection rate 0.02 and median approval 30 seconds, with 40 letters a day (25 temporarily). The evaluation stops if reopened claims rise 15 per cent for two quarters, uses three matched Hamburg teams over 24 months, and TCH-HM-CLM-01 covers 20 handlers, reserves claims above EUR 25,000, personal injury, disputed debt, complaints and vulnerable customers for humans, leaves 15 per cent of routine claims without an agent draft, requires four complete cases per handler per month, rotates quarterly, adds review time of 4 minutes (intake file) and 6 minutes (letter), and sets stop at 5 minutes and chain stop at 10. Retention of German employment decision records after a rejection is six months (AGG §15(4), ArbGG §61b) (AI Boss, Chapters 7, 8, 12, 22, 25, 26, 30 and 32) (7, Nordqvist, 2026).

Across the degrees of delegation

Function and system Recorded degree What practice showed
M1 claims allocation Level 3 Bounded in Germany Capacity factor built on throughput removed (AIDR-0651); reasons for routing to be shown in the handler view
M1 pilot agents Level 2 for their own actions The chain point for recovery letters ran in practice at Level 3 Bounded until the routing reason was shown (May 2027)
M2 throughput, sales score Level 1; sales score recorded Level 0 as pay input Score was decisive in practice in one Dutch unit
M3 TalentLens Level 1 Assisted Drifted under peak load: 38-second rejections, 3 of 6 seeds missed
M3 variable pay Level 0 may_feed empty
M4 learning recommendations Level 0
M5 None Analytics switched off

No system exceeds Level 1 Assisted for M3 in the plan, and Level 4 is excluded (AI Boss, Chapters 8, 30 and 32) (7, Nordqvist, 2026).

What the volume concludes for Hansa Mutual

The volume's lessons are specific. Hansa Mutual reclassified its own agent: the claims-triage agent was classed not high-risk for the insurance function and nobody had asked what the allocation to people was, so a capacity factor built on throughput was found and removed. It negotiated how change is governed (the change classes) and not every version. The pilot's conditions are that people manage agents and agents manage nobody, mandates are enforced outside the agents, the team charter forbids comparing agent and handler output, evaluation uses matched teams, and agents return to Level 1 if reopened claims rise 15 per cent for two quarters. A failure is recorded plainly: an advisory score became decisive, and nobody measured deviation until a Dutch employee, supported by the ondernemingsraad, contested it by comparison with the group median instead of the Dutch market. Practice at a different degree from the recorded one is legal exposure (Article 22, the SCHUFA judgment). Oversight in the pilot weakened when four handlers were absent and approvals fell below 30 seconds; Jana Petersen lowered the budget before the evaluation plan's stop rule required it (AI Boss, Chapters 30 and 31) (7, Nordqvist, 2026).

The answer to the board in Chapter 32 is that the group policy is a set of artefacts: DRF-HM-01, the AI Act baseline (AIDR-0657), the change-class annex, the transparency register (AIDR-0682) and the delegation register. The allocation carries provider obligations; the score loses its role as pay input where managers rarely deviate; the pilot expands only with a team charter and MND-HM-CLM-* mandates; and no M3 goes above Level 1 Assisted. Achievements are Brandt's empty M3 list (confirmed 2027-04-30), a team-level workload indicator with a cap per handler hour (AIDR-0685) and a mental-workload assessment of the allocation rules under Arbeitsschutzgesetz §5. The plan PLAN-HM-MGT-2027 (AIDR-0690; owner the chief HR officer; endorsed by the executive board; reviewed by the Konzernbetriebsrat; 2027-10-01 to 2029-09-30) has six items. H1 renews or ends MND-HM-CLM-01 by 2028-06-30. H2 meets provider obligations for the allocation from 2027-12-02 and shows routing reasons in the handler view, moving the chain point back to Level 2. H3 tests score deviation in each market. H4 keeps TalentLens at Level 1 until seeded catch is at least 0.85 for two quarters. H5 extends the pilot to more Hamburg teams. H6 lifts the Germany-only freeze market by market, starting with Poland, after conditions are met. Excluded: Level 4, M3 above Level 1, agents managing people, throughput as M3 input and emotion inference (AI Boss, Chapters 30 to 33) (7, Nordqvist, 2026).

Open threads

These are recorded for Chapter 27.

  • Level 4 in practice. Chapter 8 calls the sales score Level 4 in all but name in one unit and Chapter 31 says Level 4 by accident; Chapter 30 says no action of any system went at Level 4 over a person.
  • M3 ceiling. DRF-HM-01 in Appendix A says "any: L2_Supervised"; Chapters 30 and 32 and the plan's excluded list say no system raises M3 above Level 1 Assisted.
  • Plan wording. Chapter 32 describes the plan in prose and not all of H1 to H6; Appendix A and the profile list H1 to H6.
  • Renumbering. AIDR-0685 was AIDR-0670 in Chapter 14; AIDR-0653 to AIDR-0657 were AIDR-0670 and AIDR-0671; the profile states that all dates were adjusted to Chapter 30.
  • Notice date. The Article 26(7) notice for TalentLens was issued 2026-06-01, before the article applies (2027-12-02 after Regulation (EU) 2026/1744); Chapter 28 explains that the group ran the information step as soon as the system was classed as high-risk.
  • Retention. German employment records are kept six months after a rejection (Chapters 25 and 7), while MDR-HM-0001 is kept 24 months (AIDR-0679).
  • Poland. Labour Code Articles 22² and 22³ (Chapter 6) against Article 22³ (profile); allocation does not run under DLG-HM-CTA-01 in Poland while recruitment in Poland runs under TalentLens.
  • Unresolved by the volume. Whether the group's AI boss leaves Germany (an EWC question); one system with seven condition sets or seven systems; who becomes a team leader when allocation is learned through 10 per cent of the work; team leaders' oversight is reported only at department level, and Jana Petersen is the only team leader.
  • Jurisdictions. The UK and Switzerland are outside the AI Act; the group applies the baseline there as policy, and the register marks per market whether a control is law or policy.

Summary

Book 7 shows Hansa Mutual as an organisation that has already built an AI boss and then had to find out what it was. The claims-triage agent allocated work to named people, an advisory score became decisive, and a ranking tool let recruiters reject in seconds. The volume's response is not a ban but a set of artefacts: delegation profiles per system, a decision rights file with an empty list of M3 decisions throughput may feed, change classes negotiated with the Konzernbetriebsrat, stop rights measured in minutes and a pilot in which people manage agents and agents manage nobody. The measured figures, illustrative after 2026, show where oversight erodes: disparate impact of 0.69 to 0.71, 38-second rejections, approvals under 30 seconds. The binding constraint is co-determination across seven markets, and the dossier's open threads show that the volume is not consistent about whether any system acted at Level 4 in practice.

Sources

  1. Nordqvist, G. AI Boss: When Artificial Intelligence Manages People in European Enterprises and Public Authorities. (7, Nordqvist, 2026)