Dominic Chiappe · People, capability & transformation

Thinking about how organisations perform in an AI-enabled world

AyEye Opinion · Exploration paper ·

Who builds the harbour

AI and the future of global HR service centres

Broad palette-knife strokes of ochre, orange and magenta form dunes beside teal water beneath a red sun.
Original artwork: The Shape of What Remains. Chiappe × OpenAI.

LecXie research paper | 30 October 2026

How can AI transform global HR service centres from processing operations into higher value capabilities and what leadership and partnership model makes that work?

The argument in brief

AI could change the economics of global HR services without bringing those services back onshore. It can reduce repetitive work while increasing the importance of policy expertise, data stewardship, exception handling and system assurance. Those capabilities can remain in India or other global delivery locations. Their value should be assessed through access to context, authority and development rather than presumed from proximity to headquarters.

The central leadership question is therefore not simply what to automate. It is what the organisation must continue to understand, decide and control when more execution is performed by software or partners. This paper proposes a model in which firms retain clear decision ownership, develop distributed specialist capability and pay partners for independently verifiable improvements rather than activity alone. It also proposes separating authority over improvement and assurance from the commercial interests of running the service where that strengthens challenge. The objective is independent control and contestable delivery, not an impermeable boundary between teams. Portability is the practical test: can another qualified team sustain the service using assets and rights the client can actually transfer?

Evidence supports selected task improvements. It does not yet support a general claim that AI produces better reward decisions or that global HR centres can safely remove a predictable percentage of staff. Outcomes differ across tasks and workers. Commercial models based on outputs and shared gains also predate generative AI. The opportunity is to adapt them to a different work mix rather than announce their invention.

This is desk research based on public primary studies, employer disclosures, regulatory material and supplier accounts. It contains no original interviews or confidential client evidence. Findings, supplier claims and proposed designs are distinguished throughout. The accompanying evidence ledger records methods and limitations. It is not a commissioned assessment of any named employer.

A beach is not an operating model

Imagine a beach before the developers arrive. The promise is access, employment and prosperity. The risk is that those who already know the shoreline lose influence over what it becomes. A harbour changes more than the movement of boats. Someone decides which vessels can enter, who pays for protection and who bears the consequences when the defences fail.

Global service centres are not untouched beaches. Many already contain experienced specialists, sophisticated controls and significant business knowledge. Describing them all as low value processing factories would erase the capability on which any successful transformation depends.

The harbour question is nevertheless useful. An AI front door is visible. The infrastructure beneath it is less so: permissions, policy versions, source data, integrations, escalation routes, legal interpretation and the people entitled to stop an unsafe action. If those foundations are neglected, a fluent interface can conceal a poorly governed service. If they are well designed, employees may get faster help while specialists spend more time on problems that genuinely require their judgement.

The issue is who builds that infrastructure and who gains from it. A lower supplier bill can coexist with more unpaid effort by managers. Faster processing can coexist with less employee recourse. Automation can enrich a technology provider while weakening the client's ability to change providers. None is an inevitable outcome. All belong in the business case.

What the evidence establishes

The strongest evidence concerns bounded tasks rather than wholesale organisational redesign. In a published study of 5,172 customer support agents, AI assistance increased issues resolved per hour by 15% on average. Benefits differed across workers. The most experienced and highest skilled saw small speed gains with small quality declines. This is credible evidence for assisted service work, not an HR savings forecast. [S01]

Direct HR evidence is more commercially interested. IBM reports 94% containment of common questions and 11.5 million AskHR interactions in 2024. Its case study describes transaction automation as well as answers. IBM is both the user and seller of the technology. The figures show reported deployment at scale but do not demonstrate that 94% of HR work has disappeared or that every contained interaction was correctly resolved. Its transformation spans years and multiple technologies. [S02]

The ILO's 2025 occupational exposure index is useful for identifying tasks to investigate. It is not a forecast of redundancies. Exposure measures what technology might affect under specified assumptions rather than what employers will adopt or what workers will lose. [S06]

Labour-market evidence is similarly uneven. A Danish study linking adoption surveys to administrative records found task changes without detectable average earnings or recorded-hours effects through December 2024. Its confidence intervals rule out average effects larger than 2% in magnitude over the study window. The occupations included HR professionals but the study is not an evaluation of offshore HR contracts. [E01]

A different US payroll study reports an early-career employment shortfall in AI-exposed occupations. The August 2026 revision presents the pattern as descriptive rather than a causal estimate. Its importance is the risk to entry routes, not proof that AI caused a specified number of redundancies. These findings concern different countries and populations. Neither should be promoted into a universal forecast. [E02]

A disciplined reading keeps four questions separate. Can the system perform the task? Does it improve the real workflow? Does the improvement create financial or service value after all costs? Is that value achieved without unacceptable harm? Demonstrating the first does not settle the other three.

The technology label also matters. A deterministic benefits eligibility rule is not a language model. Statistical pay equity analysis is not automatically generative AI. A chatbot retrieving approved documents is not equivalent to an agent authorised to change a record. The risk and investment case should follow the action performed rather than the supplier's vocabulary.

Offshore work is not the same as outsourced work

Location, ownership and authority are separate design choices. A company can employ its own specialists in an overseas capability centre. An external provider can deliver work beside headquarters. A distributed team can be closely embedded in business decisions even when its members are not in the same building.

That distinction changes the AI argument. If the source of value in a contract is mainly the number of repetitive labour hours supplied, automation can put pressure on revenue and staffing. But offshore teams may also hold the process knowledge needed to train, evaluate and improve the new system. Removing them before that knowledge is secured can destroy the capability being purchased.

An IMF working paper identifies Philippine BPO as particularly exposed when occupational data are mapped to AI susceptibility. It measures potential exposure rather than observed displacement. It cannot establish the future of India's captive centres or a particular HR portfolio. [E03]

The alternatives are not confined to replacing people or preserving every existing task. A centre could develop roles in knowledge operations, policy testing, data quality, model evaluation and complex case resolution. These are plausible design choices rather than guaranteed employment offsets. New specialist roles may require fewer people, different qualifications or a transition period that displaced staff cannot easily bridge.

Retained expertise should mean an organisation preserves the ability to set requirements, judge performance and intervene. It need not mean every expert sits onshore. Local expertise remains important where language, employment institutions, consultation or employee relationships demand it. Those needs should be established case by case rather than used as a blanket argument for geographical relocation.

The commercial problem is older than AI

Outsourcing has never consisted exclusively of buying headcount. Genpact discussed hybrid effort, output and outcome models in 2017, including transaction fees combined with outcome-linked risk and reward. That is evidence of an established commercial idea, not proof of universal adoption or success. [S04]

Its 2025 filing describes contracts under which productivity benefits can be retained by the supplier while other provisions require some gains to pass to the client. It also describes an industry still dependent on personnel and skills. AI therefore enters an existing negotiation about investment, risk and the division of gains. [S03]

Nor is the transition uniform. Infosys BPM classified 73.8% of consolidated 2024–25 revenue as time and material and 26.2% as fixed price. Neither category is a direct measure of outcome pricing. The figures caution against claiming that the entire market has already left labour-related commercial models behind. [S05]

The genuinely useful question is whether a contract rewards the provider for making a service unnecessary. Payment per contact may discourage prevention. Payment per resolved case may encourage premature closure. A fixed fee may reward efficiency but also underinvestment. Pure gainsharing can provoke arguments over baselines and savings that arose from the client's own changes.

AI intensifies these tensions because it can alter both the volume and difficulty of work. If simple queries disappear, the remaining human queue becomes more complex. Cost per human-handled ticket can rise while total service value improves. Average handling time can deteriorate for the same reason. This paper calls that the shrinking denominator problem. Case-mix adjustment is an established idea. The contribution here is to connect it with changing capability and commercial incentives across HR portfolios, not to claim a new statistical discovery.

A contract needs to distinguish reduced demand from reduced access. An employee who gives up is not a successfully avoided contact. A manager who completes the work after the service centre rejects a case has not created a saving unless that effort is counted.

Different portfolios require different models

The following assessments are proposed operating designs informed by the evidence. They are not claims that every listed application has been proven at scale. A sensible portfolio starts with the employee problem and the consequences of error.

People Services

AI can help employees find the applicable policy, route a request, prepare a case summary or complete a tightly authorised transaction. The strongest starting points are high volume requests with reliable source material, limited ambiguity and a clear way to reverse mistakes. Better search and simpler forms may sometimes be sufficient.

A global centre can own the quality of the knowledge base and evaluate answers against country-specific cases. Specialists need authority to correct the service rather than merely process escalations from it. Sensitive cases such as bereavement, disability or a serious payroll error need an accessible human route that does not require repeated disclosure of distressing information.

Partners can supply platforms, integration engineering, multilingual support or managed operations. The client should own the definition of successful resolution and the escalation promise. Evidence should include repeat contacts, correctness, abandonment, employee effort and time to a competent human. Containment is a useful operational measure only when read beside these outcomes.

Reward and benefits

There is a material difference between explaining a pension document and deciding what someone should be paid. AI can assist with document interpretation, scenario preparation, data reconciliation and draft communications. Benefits navigation may reduce confusion but it should not drift unnoticed into personalised financial or medical advice.

Pay structures, job evaluation assumptions and the reasons for exceptions need accountable owners. Human judgement is not automatically unbiased. It must be supported by explicit criteria, reproducible analysis and scrutiny of outcomes. An ILO critical review argues that AI in HR can encode poorly chosen objectives and biased data. That is a conceptual challenge rather than a measured estimate of failure rates. [S07]

The EU pay transparency framework reinforces the importance of objective, gender-neutral criteria. It does not certify any algorithm as fair. Its national implementation and precise obligations require jurisdiction-specific checking. [S09]

Reward specialists can be distributed globally. Business leaders should remain involved in decisions requiring organisational context while local expertise interprets country arrangements. Partners may provide market data, administration and specialist modelling. They should not be paid simply for reducing reward expenditure. Correcting inequity can increase cost. Better evidence includes calculation reproducibility, fewer corrections, clearer explanations and the quality of reviewed decisions. A narrower analytical tool may outperform a conversational agent.

Talent acquisition

The potential spans scheduling, candidate communication, evidence extraction and structured interview support. Those are not interchangeable with autonomous screening or ranking. Reducing administrative delay can improve access. A faster rejection process can also scale an invalid selection rule.

A September 2026 working-paper revision offers more specific evidence. In a Philippine recruitment-process-outsourcing experiment, 67,056 eligible applications were randomised. AI-led interviews were followed by human hiring decisions. Offer rates were 9.73% in the AI arm against 8.70% in the human-interview arm. This is a 1.03 percentage point difference, not evidence that the AI itself made better employment judgements. The study covers one partner and a short employment horizon with commercial research relationships disclosed. [P01]

Retained capability should include job analysis, assessment validity, accessibility and the interpretation of hiring evidence. Recruiters and hiring managers need time and authority to challenge recommendations. Global teams can maintain assessment operations and evaluate performance while business teams define the work to be done.

Partners contribute sourcing reach, workflow technology and specialist assessment capability. The test is not just time to hire. It is whether suitable applicants receive a fair opportunity and whether the selected people succeed in the role. Measure stage-level outcomes, adjustments, candidate experience and later performance with care about confounding factors. The ICO's recruitment audits found weaknesses in accuracy testing, data collection and responsibility allocation. Those audits did not cover generative AI. [S10]

Learning and skills

AI can help staff practise unfamiliar tasks, find relevant material and receive feedback. It can assist skills mapping but inferred capability should not be mistaken for demonstrated competence. A fluent work product produced with assistance is different from knowledge a person can apply independently.

A workplace experiment published in October 2026 provides a warning against equating assisted performance with learning. Among 487 consultants, AI access improved data-science task performance but did not establish improved technical knowledge without the tool. Its commercial affiliations are disclosed in the evidence ledger. The publicly accessible abstract supports this bounded finding, not a claim about long-term capability. [P05]

The customer-support study also found evidence of learning that persisted during AI outages. That contrast reinforces the importance of the task and learning design rather than a universal verdict. [S01]

Education studies show that design matters. A Harvard physics experiment found benefits from a deliberately engineered tutor. A separate Turkish school experiment found that unguarded AI assistance could improve practice performance while worsening later unaided performance. Neither result can be imported numerically into corporate learning. Together they justify testing the learning process rather than counting access to an assistant as upskilling. [P03][P04]

The centre's role could expand from course administration into learning design, skills evidence and evaluation. That requires educators and domain experts as well as platform administrators. Business teams must identify where improved capability changes performance. External partners can contribute content, practice environments and specialist tools but cannot define organisational competence by product usage alone.

Completion rates and learning hours are weak substitutes for skill. Use validated assessments, delayed retention checks and evidence of transfer to work. Compare AI support against coaching, better documentation and supervised practice. Where errors have serious consequences, test what the employee can do without the tool and how effectively they recognise its mistakes.

Mobility and restructuring

AI could assemble case information, identify missing documents, compare policy options and prepare communications for review. Yet immigration, tax, employee consultation and redundancy decisions have consequences that cannot be reduced to faster document generation.

Distributed specialists can maintain case operations and jurisdictional knowledge. Business leaders retain responsibility for the purpose and consequences of restructuring. Qualified advisers provide legal or tax interpretation where needed. A general model should not be treated as a current source of country-specific legal advice.

Partners can provide destination services, relocation logistics and specialist advisory work. Commercial measures should favour accurate, timely cases and dependable employee support rather than the largest number of moves or the cheapest exits. Compare automation with standardised case templates and better coordination. Public independent evidence of AI improving end-to-end mobility or restructuring outcomes remains limited in this review. The appropriate response is a bounded pilot, not a manufactured ROI estimate.

Data and integration

AI can assist mapping, documentation, test generation and anomaly investigation. However, data governance is not an automation afterthought. A system that confidently joins the wrong employee records can make every downstream service less trustworthy.

The organisation should retain ownership of definitions, permissions, authoritative sources and acceptance criteria. Global centres can lead these capabilities. Partners can provide engineering capacity and platform expertise but the client needs enough technical competence to test their work and change supplier if necessary.

Success means reliable flows and controlled access, not simply more integrations built. Measure reconciliation failures, release defects, recovery time, access-control test results and the ability to reconstruct a consequential action. Interpret incident counts alongside detection and reporting quality rather than assuming fewer reports always mean a safer system. Some work is best handled by conventional validation rules or deterministic integration logic. The language model should not become the untested system of record.

A portfolio comparison

This matrix summarises proposed choices rather than proven implementation results. Specialist location follows language, jurisdiction, capability and access to context. It is not automatically headquarters.

Portfolio Retained expertise Partner contribution Decisive value evidence
People Services Policy knowledge and complex cases Platform and managed delivery Correct resolution and employee effort
Reward and benefits Pay criteria and exception judgement Market data and administration Reproducibility and justified decisions
Talent acquisition Job analysis and assessment validity Sourcing and assessment tools Fair access and subsequent performance
Learning and skills Proficiency standards and learning design Content and practice environments Retained skill and work transfer
Mobility and restructuring Business decisions and local interpretation Specialist advice and logistics Accurate cases and employee support
Data and integration Definitions and authorised actions Engineering and platform operations Reconciliation and recoverability

The leadership model should follow decision rights

A senior leader does not need to approve every automated transaction. The leader does need to make the boundary between delegated execution and retained judgement explicit. Four kinds of ownership are necessary.

First, an accountable business or policy owner decides what the service is intended to achieve and which trade-offs are acceptable. Second, a product owner manages the whole employee journey rather than a single application. Third, specialist and technical teams maintain evidence, controls and operational performance. Fourth, independent assurance and employee representation test whether the arrangement works for people as well as for the dashboard.

Independent assurance is not the same as an independent improvement supplier. A designer can have an interest in demonstrating that its own solution succeeds. NIST’s AI risk framework supports regular assessment involving experts outside frontline development and/or independent assessors. That supports independent challenge rather than a requirement to appoint a second supplier. [S17]

These responsibilities can sit across several locations. The risk is a gap between them. A supplier may say it only provides a model. Technology may say it only runs the platform. HR may say a manager approved the result. The employee experiences one decision. An operating model should identify the person who can explain and remedy that decision before deployment.

Employees should be involved in identifying failure modes before management settles the operating model. Consultation should influence design and evaluation with protected routes to raise concerns, not merely collect reactions after a decision. OECD surveys associate consultation and training with better reported outcomes but do not establish causation. A later employer survey records concerns about accountability, explainability and worker health. Employer perceptions are not a substitute for direct employee evidence. [E04][E05]

A transition plan should assign funding between client and provider for paid learning, supervised practice, redeployment support and entry routes into specialist work. It should not assume every affected worker can move into an engineering or assurance role. Otherwise an organisation may automate the junior tasks through which people previously acquired expertise and then discover that it has no pipeline of experienced reviewers. This is a strategic risk to test rather than a claim that every routine role must be preserved.

Meaningful human judgement requires competence, context, time and authority. An overloaded reviewer who routinely approves a recommendation is not a reliable control. Review queues need staffing models and quality assessment. Override rates should be interpreted rather than treated automatically as defects: a rise may reveal a broken model or a healthier willingness to challenge it.

NIST's generative AI profile identifies confidently erroneous output and related human reliance as risks requiring management. A polished explanation is not evidence that the underlying answer is true. [S08] For HR, the practical response is traceability to approved evidence, constrained actions and explicit escalation when the answer is uncertain.

The EU AI Act uses a risk-based approach to employment systems. Not every HR assistant is high risk merely because it is used in HR. Current Commission material gives 2 December 2027 as the application date for Annex III high-risk rules following the 2026 amendments. Existing privacy and employment obligations are not suspended by that timetable. This paper is not a legal compliance opinion. [S12]

A partnership model that can survive success

The proposed model combines a dependable base service with a separately governed improvement programme. It should not force every portfolio into the same charging mechanism.

The base service pays for availability, competent capacity and continuity. It protects the human escalation route and the specialist work that remains when automation succeeds. The improvement programme funds specific changes with an agreed baseline, evaluation design and division of benefits. Shared gains become payable only when quality and access conditions are met.

This resembles established hybrid contracting rather than a wholly new commercial category. Its AI-specific adaptation is to account for changing case mix, model dependencies, evaluation work and permissions for autonomous actions. The UK Sourcing Playbook is useful here because it links pricing to risk allocation and warns against fixing prices when scope is not fixed. It is guidance rather than proof of commercial results. [S11]

Contracts should state who supplies and updates knowledge, who approves model changes and who investigates incidents. They should distinguish savings from changes in volume, wages or policy. Audit rights must cover the evidence needed to verify outcomes without granting unnecessary access to personal data. Retention and deletion rules should apply to prompts and logs as well as to the original HR record.

Critical services also need a rehearsed manual or deterministic fallback. Owning the data is insufficient if nobody retains the skill and capacity to resolve a serious payroll or benefit error when an AI service is unavailable.

Independence without isolation

The organisation that improves a service need not be the organisation that runs it. Separation can create room to challenge processes that an incumbent is paid to maintain. But appointing two suppliers does not establish independence if both depend on the same proprietary platform or if the client cannot judge either provider’s work.

An information barrier restricts access to confidential material. Commercial separation changes incentives and reporting relationships. Technical portability makes components replaceable. These are different controls. The proposed model uses targeted confidentiality safeguards but preserves operational feedback across the improvement and delivery boundary.

There is a useful precedent in the UK DDaT Playbook. Its discussion of pilots recognises advantages from specialist suppliers and independent evaluation while warning that multiple procurements can increase cost and duplication without effective knowledge transfer or a consistent integrator. This is procurement guidance, not evidence that separated suppliers produce better HR outcomes. [S13]

A strict build/run divide can also discard valuable knowledge. Google’s SRE account describes operators using their production experience to help development teams reduce repetitive work. That practitioner experience is not an HR trial but provides a counterargument to isolating designers from operations. [S16] Improvement partners should remain answerable for supportability while operators should contribute to redesign. One named service owner must resolve cross-provider dependencies and incidents rather than allow each supplier to meet its own target while the employee journey fails.

Separation is most plausible where interfaces are stable, improvement incentives conflict with the run contract and the client has sufficient integration capability. Integrated delivery may be preferable where changes are frequent and tightly coupled or continuity demands close collaboration. Either arrangement needs independent challenge. These are design hypotheses to test against cost, quality and change speed rather than universal sourcing rules.

Open services and a credible change of provider

Open services, as used here, means services designed to connect, be inspected and change providers without losing organisational knowledge or control. This is a design principle rather than a claimed new standard. It combines documented interfaces, portable assets and clear contractual rights with protected employee data. It does not require every component to be open source or confidential information to be public.

A practical expression could be a capability service interface for a mixed human and AI workforce. For a request such as assessing a reward exception, the service contract would specify inputs, required evidence, permissions, quality criteria, escalation and the accountable decision owner. Some exchanges could use a technical API. The wider contract also governs human dialogue and challenge. The delivery mix behind it could change without rewriting the business’s expectations. This is a proposed operating design, not an established solution. It standardises access and governance rather than assuming every judgement can be standardised or that people are interchangeable components. Specialists and employees still need room to supply context, contest assumptions and influence the work.

Scalability and provider independence are not the same achievement. A common platform can expand quickly across countries while increasing dependency on its owner. Excessive portability requirements can add cost or forfeit useful managed capabilities. Government cloud guidance explicitly recognises this trade-off and describes testing components in another environment to reassess switching costs. [S14]

The client needs usable data, policy provenance, documented interfaces, configurations, evaluation cases and operational knowledge. Runbooks must explain how the service is operated and restored. Transfers remain subject to privacy, security, regulatory and contractual constraints. The test is lawful continuity rather than unrestricted copying. Rights to continue using or licensing necessary assets to a replacement operator can matter more than owning every component. Government IP guidance warns that demanding ownership or exclusivity can discourage innovation or raise price. Supplier-owned reusable technology and client freedom to re-compete need not be incompatible if the rights are designed properly. [S15]

Distinguish changing the operator from changing the platform or model. A new operator may work on an existing platform without difficulty while replacing that platform requires substantial redesign. Data export alone establishes neither service continuity nor equivalent AI behaviour. The organisation should know which dependencies it accepts and why.

Suppliers also need a viable return. If every improvement immediately removes their revenue without compensating investment, they have little reason to build it. If they capture all the gains indefinitely, the client has little reason to collaborate. A negotiated sharing period with periodic rebasing can be more credible than a permanent promise of percentage savings. The appropriate terms depend on control, investment and risk rather than a universal formula. Periodic benchmarking or market testing should check whether the arrangement still earns its return. Improvement fees should not continue indefinitely after the improvement becomes normal service.

Proving value without hiding the work

A pilot should compare an AI-enabled design with a credible alternative. That alternative might be the existing service but it could also be a better knowledge base, a redesigned policy or conventional automation. Otherwise the organisation may pay a technology premium for an improvement that did not require AI.

Define the unit of analysis before deployment. A contact, a case, an employee journey and a pay decision are different denominators. Track the same case across channels so that escalation does not reset the clock. Separate capacity released from cash savings and record where released time actually goes.

A fully loaded economic assessment includes licences and inference, integration, data preparation, testing, assurance, specialist review, security, training, change management, transition costs and the retained client organisation. It also includes rework and employee or manager effort shifted outside the centre. Potential loss reduction should be shown separately from realised savings rather than used as a balancing figure.

Where feasible, use a phased or randomised rollout with comparable groups. Adjust for demand, country, seasonality and case complexity. Predefine acceptable error rates and stop conditions. For consequential decisions, an initial shadow period can test recommendations without allowing the system to act. A successful demonstration is not permission for every subsequent model version. A pilot with no serious incidents may still be too small to establish safety for rare harms. Scenario testing, access controls and difficult edge cases matter alongside average accuracy. Results should be examined by relevant groups and jurisdictions rather than only in aggregate.

An illustrative arithmetic example shows the problem. Suppose a service handles 10,000 cases per year at a fully loaded cost of £10 each. In the proposed design, 4,000 are correctly resolved through automation with their operating cost included in a £20,000 annual technology and assurance budget. The other 6,000 require human handling at £12 each. All 10,000 remain in the end-to-end resolution denominator. Total cost falls from £100,000 to £92,000 before transition and displaced effort. Prevented demand and unresolved demand must be tracked separately. That is an 8% gross operating reduction, not a 40% saving. The example is invented to explain the calculation and is not a forecast or benchmark.

For a proposed separated model, compare an integrated alternative on the same scope. Include retained integration staff, duplicated tooling, handoff delays and transition costs. Test a bounded service transfer: can an authorised replacement team use the agreed asset package to operate a representative workflow within an acceptable time and budget? Re-run quality and access tests after any model or platform change. A contractual exit clause is weaker evidence than a successful controlled rehearsal. These proposed tests assess freedom to change provider without assuming that switching is always desirable.

Reward, recruitment and learning require additional outcome measures. Pay accuracy is not pay fairness. Selection speed is not hiring validity. Course usage is not capability. A credible business case specifies the distinction and the evidence needed to cross it.

Three scenarios rather than one forecast

These scenarios are planning hypotheses for the next two to five years. The time windows are not predictions of adoption.

Augmented global centres over the next two years. Routine assistance expands but specialists and approvals remain central. The conditions are usable knowledge, sufficient integration and investment in learning. The upside is quicker service with stronger exception handling. The downside is work intensification as staff lose easier cases while retaining demanding ones. Watch case mix, staff wellbeing and service quality rather than assuming that smaller queues mean lighter work.

Capability networks over two to five years. Centres, business-embedded experts and specialist partners operate through shared products and evidence standards. Ownership is clearer even though delivery is distributed. The conditions are interoperable data, capable product leadership and commercially workable sharing of benefits. The upside is better coordination across hiring, skills, reward and mobility. The downside is diffuse accountability or dependence on a small platform ecosystem. Watch the ability to explain decisions and move work between providers.

Automated fragmentation within either window. Local teams buy tools independently while suppliers optimise narrow metrics. The apparent gains are fast demonstrations and reduced visible effort. The conditions are weak architecture and pressure to cut cost before evidence is available. The downside is duplicated expenditure, hidden rework and inaccessible services. Watch unresolved contacts, conflicting records, exceptions outside the formal process and rising managerial work.

None assumes that higher value work moves to headquarters. The location of new capability is an outcome of investment and authority. It should not be predetermined by a story about the end of offshore work.

Who benefits from the harbour

AI can make global HR centres more capable. It can also make their contribution harder to see. If routine work disappears, the remaining expertise may be more valuable even when conventional productivity measures look worse. If an assistant makes a worker temporarily more effective, the organisation still needs to know whether knowledge has grown or merely been rented.

The leadership task is to preserve the ability to understand and contest the system while improving the work it performs. That means developing specialists where they already exist, involving employees in service design and asking partners to prove improvements against the whole journey.

The organisation should be able to change a delivery partner without surrendering its knowledge or rebuilding every control. That freedom depends on competent ownership, usable rights and tested alternatives rather than supplier count. The harbour should remain usable when its operator changes. Its designers and operators still need to talk.

The most credible partnership is not the one that promises the fewest people. It is the one that can show who is better served, where value is created and who remains answerable when the system is wrong.

Life is a beach, but the sand does not stand still. It shifts to form majestic dunes. The task is not to freeze the landscape, but to build services that can adapt without losing the knowledge, trust and human judgement that make them valuable.

Source references

S01. Brynjolfsson, Li and Raymond. Generative AI at Work. Quarterly Journal of Economics. 2025. Source

S02. IBM. IBM AskHR case study. Accessed 30 October 2026. Source

S03. Genpact. Form 10-K for year ended 31 December 2025. Source

S04. Genpact. Outcome based commercial models at the point of accelerated adoption. 9 August 2017. Source

S05. Infosys BPM. Annual Report 2024–25. Page 48. Source

S06. ILO. Generative AI and jobs A 2025 update. 20 May 2025. Source

S07. Berg and Johnston. AI in human resource management The limits of empiricism. ILO Working Paper 154. 11 November 2025. Source

S08. NIST. Artificial Intelligence Risk Management Framework Generative Artificial Intelligence Profile. July 2024. Source

S09. Council of the European Union. Pay transparency in the EU. Accessed 30 October 2026. Source

S10. ICO. AI tools in recruitment Audit outcomes report. November 2024. Source

S11. Cabinet Office. The Sourcing Playbook. Source

S12. European Commission. AI Act policy and implementation overview. Accessed 30 October 2026. Source

S13. Cabinet Office. The Digital, Data and Technology Playbook. June 2023. Source

S14. Government Digital Service. Managing technical lock-in in the cloud. Accessed 30 October 2026. Source

S15. Cabinet Office. Intellectual Property Rights Guidance Note. June 2023. Source

S16. Google. The Site Reliability Workbook: Eliminating Toil. 2018. Source

S17. NIST. Artificial Intelligence Risk Management Framework 1.0: Core, Measure 1.3. January 2023. Source

P01. Jabarian and Henkel. Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews. Working paper version 11 September 2026. Source

P03. Kestin and colleagues. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports. 3 June 2025. Source

P04. Bastani and colleagues. Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS. 2025. Source

P05. Wiles and colleagues. Generative AI and the temporary upskilling of knowledge workers. Nature Human Behaviour. 26 October 2026. Source

E01. Humlum and Vestergaard. Still Waters Rapid Currents. Working paper version 13 March 2026. Source

E02. Brynjolfsson, Chandar and Chen. Canaries in the Coal Mine. Working paper August 2026 revision at a legacy URL. Source

E03. Cucio and Hennig. Artificial Intelligence and the Philippine Labor Market. IMF Working Paper 25/43. February 2025. Source

E04. Lane, Williams and Broecke. The impact of AI on the workplace. OECD. 27 March 2023. Source

E05. Milanez, Lemmens and Ruggiu. Algorithmic management in the workplace. OECD. 6 February 2025. Source

Evidence ledger for the HR service centre paper

Evidence companion | Sources checked 30 October 2026

How to read this ledger

The main paper is desk research rather than original reporting. Studies were selected for methodological clarity, direct relevance or their ability to challenge the thesis. This is not a systematic review and absence of evidence here does not establish absence everywhere. No original interviews or supplier validation were undertaken. Source interests do not invalidate findings but affect how strongly they should be used.

Observed results, self-reported deployments, exposure models and normative guidance are different evidence classes. The main paper uses its own proposed operating model rather than attributing that design to any source. Numbered identifiers match the paper.

Core evidence

S01 Service productivity

Brynjolfsson, Li and Raymond. Quarterly Journal of Economics. Published 4 February 2025. Source

Staggered introduction of an AI assistant among 5,172 customer-support agents. Outcome is issues resolved per hour, averaging 15% improvement against comparison observations without access. Effects vary by experience and skill. Peer-reviewed field research rather than an HR trial. One deployment and selection into rollout limit generalisation. The published sample and estimate differ from earlier working-paper versions. Does not establish headcount savings or reward decision quality. Its outage-based learning evidence uses chat duration with noisy estimates rather than a separate proficiency test.

S02 Direct HR deployment

IBM AskHR case study. Undated current page accessed 30 October 2026. Source

IBM reports 94% containment of common questions and 11.5 million employee interactions in 2024. It describes around 80 automated tasks. Vendor-as-customer account with no independent counterfactual or published accuracy denominator. Containment is not the share of all HR work automated. The reported ticket reduction uses a 2016 baseline and should not be attributed solely to contemporary generative AI. Useful as a scaled deployment example, not transferable ROI evidence.

S03 Current outsourcing incentives

Genpact Form 10-K for year ended 31 December 2025. Management discussion printed pages 50–51. Source

Corporate filing describes productivity-benefit retention and minimum pass-through provisions in some contracts. These are actual disclosed commercial arrangements without a representative contract sample or causal outcome estimate. Filing also describes a labour-intensive cost base. Management has a commercial interest; financial reporting does not independently validate every strategic statement. No HR-specific or AI-only attribution is available.

S04 Historical outcome models

Genpact. Published 9 August 2017. Source

Supplier discussion describes hybrid effort/output/outcome arrangements and unnamed client examples. Establishes that these models were discussed before today's generative AI wave. No audited savings, disclosed sample or counterfactual. Not evidence that outcome contracting was universal, successful or newly invented by AI.

S05 Pricing mix

Infosys BPM Annual Report 2024–25. Printed page 48. Source

Consolidated revenue mix: time and material 73.8% versus 70.7% in the preceding year; fixed price 26.2% versus 29.3%. Denominator is reported service revenue classification, not contracts or employees. Corporate financial disclosure. Time and material is not identical to FTE charging; fixed price is not synonymous with outcome pricing. One company, not an industry estimate.

S06 Occupational exposure

ILO/NASK. Research brief published 20 May 2025. Source

Task-level occupational exposure model combining expert and AI input. Defines gradients of potential exposure. No before/after deployment or redundancy baseline. Institutional research, not supplier marketing. Exposure should not be translated into job loss, adoption or savings. Occupational aggregation may hide employer-specific task variation.

S07 Reward and HR objectives

Berg and Johnston. ILO Working Paper 154. Published 11 November 2025. Source

Critical review across recruitment, compensation, scheduling and performance management. Examines objectives, data and programming. Conceptual analysis rather than an experiment or prevalence study. Its argument about misaligned objectives challenges claims of automatic fairness but does not supply a measured failure rate. The authors’ analytical stance informs their interpretation.

S08 Generative AI risk

NIST AI 600-1. Published July 2024. Source

Voluntary cross-sector risk-management profile. Addresses confidently erroneous outputs and human reliance among other risks. No outcome denominator, baseline or causal estimate. Authoritative technical guidance but not proof that a particular control works or a legal certification. Application to HR decision rights is the paper's proposed design.

S09 Pay transparency

Council of the European Union policy explanation. Accessed 30 October 2026. Source

Explains the EU pay transparency framework and objective, gender-neutral pay criteria. Primary institutional explanation, not AI efficacy research. National implementation and exact employer obligations require separate legal review. The paper deliberately avoids numerical reporting thresholds or treating all countries as identically implemented.

S10 Recruitment audit

ICO. November 2024. Source

Consensual audits from August 2023 to May 2024 led to 296 recommendations and 42 advisory notes. Findings include accuracy, data minimisation and responsibility issues. No industry prevalence denominator. Generative AI was explicitly outside scope. Independent regulatory evidence of concrete weaknesses, not proof that all hiring systems discriminate.

S11 Procurement design

Cabinet Office Sourcing Playbook. June 2023 edition and current HTML. Source

Normative public procurement guidance on risk allocation, payment and delivery-model assessment. No measured AI outcome or statistical denominator. Principles can inform private contracts but are not automatically binding on private employers. Proposed hybrid HR contracts in the paper are analytical recommendations.

S12 AI Act implementation

European Commission policy overview. Accessed 30 October 2026. Source

Current institutional account reflects 2026 amendments and gives 2 December 2027 for Annex III high-risk requirements. Corroborated against Council legislative text PE-CONS 30/26, page 90 clause 40: Source. Employment use classification depends on intended purpose and exceptions. No efficacy claim. Practical legal reliance requires checking current consolidated law. This does not postpone existing privacy, equality or employment obligations.

Additional empirical evidence

P01 Voice AI in Firms

Brian Jabarian and Luca Henkel. Working paper version 11 September 2026. Source

Preregistered Philippine RPO experiment at PSG Global Solutions, March to June 2025. 67,056 eligible applications randomized from 70,884 received. Human arm 13,557; AI arm 40,103; choice arm 13,396. Offer probability 8.70% versus 9.73%: 1.03 percentage points or 12% relative. Humans made hiring decisions. One-month employment among randomized applicants was 4.98% versus 5.89%, not conditional retention. Existing fees rewarded hires remaining one month. Preprint with one labour market and short horizon. Firm supplied data; Google Cloud supported project; author later accepted unpaid partner role. No proof of long-term job quality or executive hiring validity. Applicant survey response was 14%.

P03 Designed AI tutoring

Kestin and colleagues. Scientific Reports. Published 3 June 2025. Source

Harvard introductory physics crossover RCT in autumn 2023. 233 enrolled; 194 eligible participants with consent and completed assessments. Two lessons compared AI tutoring with active classroom learning using identical content. Reported stronger learning in less time. Expert-written solutions and carefully engineered instructional prompts mattered. Platform development took months. Authors developed the intervention and declared no competing interests. University educational context. Does not establish corporate learning transfer or long-term retention. The findings do not establish that AI can replace teachers.

P04 Learning without guardrails

Bastani and colleagues. PNAS. Published online 25 June 2025. Source and Source

Field experiment involving nearly 1,000 Turkish high-school maths students. GPT Base improved assisted practice grades by 48% versus control yet reduced subsequent unassisted grades by 17%. Guardrailed GPT Tutor improved practice by 127% with the learning-harm pattern largely mitigated. Percent changes are relative grade effects, not percentage points. Peer-reviewed educational research; authors declared no competing interests. An August 2025 correction changes an affiliation, not findings. These effect sizes do not directly establish workplace training effects. Main transfer hypothesis: assess performance with assistance separately from independent capability.

P05 Temporary workplace upskilling

Wiles and colleagues. Nature Human Behaviour. Published 26 October 2026. Source

Randomized experiment with management consultants, analysis N=487. AI access and training improved data-science task scores over controls but did not match expert data scientists. No evidence of improved post-experiment technical knowledge without ChatGPT. Source interest: BCG employee time/resources; BCG and OpenAI coauthors; academic author received BCG consulting fees. Publisher abstract and disclosure verified, not full paywalled methods. These findings do not establish long-term skill erosion; precise effect sizes are not reported here. This is directly relevant to distinguishing tool-enabled performance from acquired skill, though not an HR trial.

E01 Humlum and Vestergaard

Source: Still Waters Rapid Currents. Version: 13 March 2026. Design: Adoption surveys linked to administrative data; latest survey approximately 25,000 workers across 7,000 workplaces in 11 occupations. Difference-in-differences through December 2024; compares adopters/non-adopters and employer initiatives. Finding: Rules out average earnings/recorded-hours effects above 2% while documenting task change. Interest: Academic working paper; disclosed university centre funding. Limit: Denmark and early chatbot adoption, not causal estimates of advanced agents or offshore HR contracts. Earlier title was Large Language Models Small Labor Market Effects.

E02 Brynjolfsson Chandar and Chen

Source: Canaries in the Coal Mine. Version: PDF now dated August 2026 despite older URL. Design: Millions of US workers in ADP payroll data through June 2026. Finding: Age 22–25 employment in exposed occupations 19% below a benchmark keeping pace with less-exposed peers. Interest: Academic working paper using proprietary payroll data. Limit: Descriptive rather than causal; pre-existing trends, education controls and sample representativeness qualify inference. Not a 19% absolute economy-wide job loss. Estimates depend on the working-paper version.

E03 Cucio and Hennig

Source: Artificial Intelligence and the Philippine Labor Market. Date: February 2025, WP/25/43. Design: Exposure/complementarity indices merged with October 2022 Labour Force Survey, over 180,000 observations including approximately 80,000 employed. Finding: BPO has the highest share of high-exposure/low-complementarity jobs. Interest: IMF working paper, authors' views rather than IMF institutional position. Limit: Modelled potential, older occupational microdata and broad BPO classification. Neither observed displacement nor an India/GCC/HR-specific forecast.

E04 Lane Williams and Broecke

Source: OECD employer and worker surveys, sample description. Date: 27 March 2023; fieldwork early 2022. Design: 5,334 workers and 2,053 firms in manufacturing/finance across seven countries. Finding: Generally positive reported experience alongside job-loss concerns; consultation/training associated with better outcomes. Interest: Intergovernmental research rather than vendor promotion. Limit: Self-reported associations, sector limits and pre-ChatGPT fieldwork. The findings do not establish causal effects from worker consultation or directly generalise to current GenAI HR services.

E05 Milanez Lemmens and Ruggiu

Source: Algorithmic management in the workplace. Date: 6 February 2025. Design: Survey of over 6,000 firms in France, Germany, Italy, Japan, Spain and the US. Finding: Managers perceive benefits and concerns including unclear accountability, explainability and worker health protection. Interest: OECD policy research. Limit: Employer perceptions, not employee testimony or observed harm incidence; algorithmic management includes non-AI software. No before/after causal baseline. The findings motivate risk measurement but do not prove that AI causes specific health outcomes.

Independence and portability evidence

S13 Separation and integration trade offs

Cabinet Office. The Digital, Data and Technology Playbook. June 2023; current HTML checked 30 October 2026. Source

Commercial considerations for pilots recognise specialist competition and independent evaluation benefits from separate suppliers but also procurement cost, duplication and handoff risks without effective knowledge transfer or a consistent integrator. Public-sector normative guidance rather than an evaluated HR operating model. No causal baseline, sample or quantified separation benefit. It does not prescribe a universal split between build and run.

S14 Technical portability

Government Digital Service. Managing technical lock-in in the cloud. Current official page checked 30 October 2026. Source

Guidance balances portability against useful integration and managed capabilities. Describes testing components in another environment to reassess switching costs. No representative sample or measured benefit is supplied. The paper’s proposed HR transfer rehearsal is an application of the principle, not a reported deployment result. Moving an operator differs from replacing a platform or model.

S15 Rights rather than blanket ownership

Cabinet Office. Intellectual Property Rights Guidance Note. June 2023; checked 30 October 2026. Source

Guidance distinguishes ownership from licence rights including use by replacement suppliers. Warns about price and innovation effects of ownership or exclusivity demands. Normative public procurement source with no outcome denominator. Commercial rights must be checked against background and third-party IP. It does not establish that all AI assets can be transferred or that exclusive ownership is necessary.

S16 Counterargument from operational engineering

Google. The Site Reliability Workbook: Eliminating Toil. 2018. Source

First-party practitioner account describes operational knowledge informing engineering and collaboration to reduce repetitive work. Google has an interest in its operating practices and technology. Not independent evidence comparing separated and integrated HR suppliers. Used to challenge an impermeable build/run boundary rather than to import numerical productivity claims or Google staffing rules.

S17 Independent AI assessment

NIST. AI Risk Management Framework 1.0, issued January 2023. Core Measure 1.3, current official presentation checked 30 October 2026. Source

Voluntary framework supports regular assessment by internal experts outside frontline development and/or independent assessors. No experimental baseline or measured benefit. Supports independent challenge but does not require external suppliers or structural separation of improvement from operations. The paper’s commercial design remains its own proposal.

What remains unproven

Independent end-to-end HR returns are less developed than task-level evidence. This review does not establish AI-caused offshore HR job losses, durable reward-fairness improvements, reliable autonomous restructuring decisions or a universally superior partnership model. In particular, the sources do not establish a causal HR-specific advantage from separating improvement and run providers. It also does not establish that observed productivity improvements persist through major model or policy changes.

Legal timelines and living webpages can change. Working papers can also change titles, samples and estimates at the same URL. Implementation requires process-level baseline data, employee feedback and a credible non-AI comparator. No source in this ledger substitutes for those tests.