Workforce Arbitrage in Customer Care

At T1 and bounded T2 task complexity, the Stewardship Model — one operator overseeing an agentic stack — produces a modelled 95.4% cost reduction in our home market of Ireland against a domestic human baseline at reference volumes. This is the deployable arbitrage. The same framework applied to Italy, a lower-cost EU market, produces 95.6% — confirming the structural finding generalises across EU cost structures, with the absolute saving scaling against the human baseline being displaced. At T3, where legal accountability, regulatory liaison, and relationship-critical judgment remain binding constraints, AI assistance compresses cycle time by approximately 30%; the Steward governs the system but does not replace specialist judgment. This publication quantifies all three regimes — pure human workforce, full Stewardship, and AI-assisted human-led — across twelve task types spanning T1, T2, and T3 task tiers, with cost benchmarks for Ireland and Italy. The modelled arbitrage is tier-specific. The size of the captured arbitrage depends entirely on which tier the work occupies, and on the architectural discipline of the operator that captures it.

Status of the analysis. This publication presents a reference model, not a deployment report. Cost structures are sourced from Citizens Information, PwC Worldwide Tax Summaries, Morgan McKinley, and Hays salary guides. Productivity benchmarks are sourced from SQM Group and ICMI. Agentic FCR figures are Arco model assumptions for bounded workflows with audited guardrails, measured under Resolution Integrity rather than raw containment. The arbitrage is calculated against published cost structures and modelled containment; the deployment that captures it is structurally feasible under the conditions disclosed in the methodology.

Key Findings

Function-level finding (all tiers including T3 reframed as AI-assisted human-led, Ireland baseline): €772K annual saving / 66.5% cost reduction. Italy comparative: €723K / 67.4%. Sensitivity analysis confirms arbitrage holds above 55% at T2 and 67% at T1 even under stress assumptions in either market.

* Structurally deployable today under disclosed conditions.

What We Set Out To Measure

This analysis quantifies the cost and throughput delta between three operating models for a customer care function: a traditional human workforce benchmarked against Irish and Italian employer cost structures, a pure agentic stack, and the Stewardship Model — one operator overseeing an agentic core. Customer care concentrates the full task complexity spectrum within a single operating unit, from fully scriptable T1 interactions to regulated, judgment-critical T3 escalations. The task tier framework maps directly onto the T1 / T2 / T3 classification used across our portfolio.

Ireland is the primary baseline because it is our domicile and the market we know best. Italy is the substantive comparative case because it is a lower-cost EU market with a different employer-contribution structure — useful for testing whether the framework's findings generalise across EU labour cost profiles or are specific to one cost regime. The third option a competitor commonly raises — offshore BPO labour at $7–16 per hour — is treated explicitly in Section 7.

The analysis covers twelve ticket types across three tiers. T1 tasks are routine, scripted, and high-volume: password resets, FAQ responses, order tracking, and basic billing queries. T2 tasks require moderate judgment and policy interpretation: technical troubleshooting, returns processing, complaint handling, and billing disputes. T3 tasks carry regulatory accountability, specialist knowledge, or relationship-critical judgment: technical escalations, legal and compliance complaints, enterprise account issues, and data security incidents.

The analysis covers twelve ticket types across three tiers. T1 tasks are routine, scripted, and high-volume: password resets, FAQ responses, order tracking, and basic billing queries. T2 tasks require moderate judgment and policy interpretation: technical troubleshooting, returns processing, complaint handling, and billing disputes. T3 tasks carry regulatory accountability, specialist knowledge, or relationship-critical judgment: technical escalations, legal and compliance complaints, enterprise account issues, and data security incidents.

Section 1 — Throughput: The Scale of the Gap

The throughput differential between human agents and the agentic stack is not incremental. Throughput figures depend on task type, not market — productivity is identical whether the agent is in Dublin or Milan. Cost per ticket varies by market and is treated in Section 2.

At T1, an average human agent resolves between 37 and 65 tickets daily. The equivalent agentic workflow handles between 1,600 and 3,100 of the same tasks in the same 8-hour window — a 43× to 48× multiplier.

Agentic FCR rates are Arco model assumptions for bounded workflows with source-grounded retrieval and audited guardrails, measured under Resolution Integrity (closures that hold against reopen and downstream-exception checks) rather than raw containment. Industry benchmarks for human FCR cluster around 70% (SQM Group); published agentic FCR ranges from 55% to 80%, often without integrity verification. Sensitivity analysis in Section 4 tests outcomes under FCR assumptions 5 to 15 percentage points lower.

At T2, the multiplier compresses but remains substantial. Human average output ranges from 11 to 17 tickets per day. Agentic throughput sits between 250 and 500 daily, producing a multiplier of 23× to 29×. The compression relative to T1 reflects increased orchestration overhead and the elevated rate of exception handling.

At T3, the throughput multiplier becomes misleading. Raw agentic volume ranges from 8 to 40 cases per day, yielding multipliers of 2× to 8×. But agentic First Contact Resolution rates fall to 30–55%. Volume without resolution is not throughput. The economic value at T3 comes from cycle-time compression on human-led work, not raw substitution. Section 3 develops this distinction. The Escalation Rate — the proportion of agentic transactions requiring human intervention — is the operational metric that governs where each tier sits on the substitution curve. At T1 the target rate is approximately 1:100; at T2, 1:5 to 1:10; at T3, mandatory human accountability on every case. The Intervention Threshold is the design parameter that sets the target; the realised Escalation Rate is the measurement that confirms the design held.

Section 2 — Human Cost: Ireland Baseline

Ireland total employer cost is calculated as gross salary multiplied by 1.1625, comprising 11.25% Class A employer PRSI (effective October 2025; rising to 11.40% from October 2026) plus 5% benefits and allowances. Salary benchmarks reflect the Dublin and Cork markets from Morgan McKinley and Hays Ireland 2026 data.

At reference volumes (T1: 500 tickets/day, T2: 100/day, T3: 20/day, 220 working days), the Irish all-human baseline requires 24.78 FTE — 10.61 at T1, 7.75 at T2, 6.42 at T3 — with a total annual loaded cost of €1,159,860. Per-ticket human costs range from €2.44 (FAQ) to €4.28 (basic billing) at T1, from €11.50 to €22.10 at T2, and from €61.30 to €153.24 at T3. These are the unit economics the autonomous alternative competes against in our home market.

Section 3 — Italy as Comparative Market

Italy provides the substantive cross-EU comparison. It is a lower-cost market with a different contribution structure — higher employer social contributions (~33% INPS) but lower gross salaries — making it a useful test of whether the framework's findings generalise across EU labour cost profiles.

Cross-market comparison

Total employer cost in Ireland is 5.3% to 11.2% higher than Italy at every role level. Gross salaries are 25% to 32% higher — materially wider, because Italy's higher employer social contributions (~33% INPS) absorb a substantial share of the gross-salary gap. The two figures describe the same underlying asymmetry through different lenses; conflating them produces overstatement, separating them preserves the directional claim with claim discipline intact.

At reference volumes, the Italian all-human baseline costs €1,073,197 — €86,663 less than Ireland. The structural arbitrage available to an autonomous competitor scales with the absolute gross labour cost rather than the social-charge gap. A market with high gross salaries and modest social contributions presents a wider absolute arbitrage than a market with the inverse profile — and Ireland sits in the former category despite the lower headline contribution rate. This generalises beyond the two markets modelled here: the labour cost a Stewardship Model competes against is the gross compensation of the workers it replaces, not the net wage they receive.

Section 4 — Workforce Arbitrage: The Composite View

The Stewardship Model cost per ticket allocates the Steward's time across tiers by the proportion of judgment burden each generates. T1 tasks consume approximately 5% of Steward time at reference volumes; T2 approximately 35%; T3 approximately 60%. At Irish reference volumes, a single Steward overseeing the agentic stack absorbs 18.4 FTE worth of T1+T2 work — a 46× substitution ratio at the point where the Stewardship Model fully applies. The substitution captured here is Labor-to-Compute Substitution in operation: variable human labour costs replaced with fixed or near-zero marginal compute costs for the same unit of operational output.

Per-ticket economics — Ireland

Annual savings — Ireland baseline and Italy comparative

The framework produces consistent percentage capture across both markets — 95.4% Ireland, 95.6% Italy at T1+T2. The absolute saving is approximately €40,000 larger annually in Ireland because the human baseline being displaced is bigger. This is what generalises: structural arbitrage scaling with absolute cost, not regulatory geography.

Section 5 — The Honest Assessment

The T1 and T2 verdicts are unambiguous within bounded conditions. Password resets, FAQ responses, order tracking, and basic billing queries are deployable today under workflows that include identity verification gates, hard escalation triggers on vulnerable-customer signals, and continuous QA sampling. Returns processing, technical troubleshooting on known failure patterns, and complaint handling on standard cases are deployable under the Stewardship Model with the Steward acting as exception authority on 12% to 20% of transactions. Below these bounds the substitution holds; above them, the cost-of-failure exceeds the cost-of-handling.

The T3 assessment requires precision on a point that matters for credibility. Legal and compliance complaints carry a constraint that is not a technology limitation. Many jurisdictions require a named human party to be accountable for the handling and resolution of formal complaints. An agentic workflow can draft responses, retrieve relevant precedent, collate evidence, and structure resolution proposals — but it cannot be the accountable party. This will not change as models improve. Enterprise account management and data security incidents share a different constraint: the stakes of a single failure are high enough that premature agentification carries unacceptable commercial and reputational risk.

At these tiers, the Steward is not the exception handler. The Steward is the product. The economic value of AI at T3 is approximately 30% cycle-time compression on human-led work — measured as faster case onboarding through retrieval, faster drafting of legally precise responses, and faster evidence collation for incident reports. At reference volumes, this reduces specialist FTE requirement from 6.4 to 4.5 and produces approximately 18% T3 cost reduction in both Ireland and Italy. That is a real and defensible number; it is not a labour-substitution number. The shift the AI captures at T3 is from execution time to judgment time — the specialist does less of the mechanical work and more of the work only a specialist can do.

Where the model breaks

The Stewardship Model is not invulnerable. Eight failure modes can compress or eliminate the arbitrage in either market: Contextual Friction when inputs require subjective judgment that resists deterministic classification, the Coordination Trap when AI tools accelerate tasks but human dependencies in the workflow architecture remain, Logic Decay when the data environment shifts under unchanged code, Handoff Friction at integration boundaries, knowledge-base quality erosion at scale, identity-verification gaps even at T1, Resolution Integrity breakdown — where agentic ticket closures fail to hold under reopen and downstream-exception measurement, exposing a gap between reported FCR and actual resolution, and vendor price/performance drift. Each is recoverable with the right architectural decision. None are recoverable after deployment if they were not named before deployment.

Section 6 — Sensitivity: How the Model Holds

Single-point estimates are vulnerable to focused critique even when directionally correct. The reference model is therefore stress-tested across three scenarios. The variables that move are agentic FCR rates, Steward loaded cost, automation operating cost, and QA/rework load. The Steward cost in the stress scenario reflects Dublin/Cork market premium for an experienced operator (€122,000 fully loaded). The arbitrage thesis is only credible if it survives conservative assumptions in our home market.

The Base case captures 90.4% of Irish T1+T2 cost. The Conservative case — Steward at €95,000, automation at 2× base, 15% QA load, FCR rates 5 to 10 percentage points lower than Base — captures 79.0%. The Stress case — Steward at €122,000, automation at 5× base, 30% QA load, FCR rates compressed further — captures 61.4%. The arbitrage holds at more than 55% of T2 cost and more than 67% of T1 cost even under the stress scenario. That is the floor in our home market.

A separate cognitive-load check examines whether one Steward can absorb the exception volume across scenarios. At Base FCR rates, the Steward handles approximately 5 T1 exceptions and 15 T2 exceptions per day — 5.4 hours of work per shift, fitting within the 6.5-hour productive day. Under Conservative FCR rates the load rises to 10.4 hours per day, requiring a second Steward. Under Stress FCR rates it rises to 19.6 hours, requiring a third Steward or tighter workflow scoping. The single-Steward claim holds at Base and breaks under Conservative. A second Steward at €76,725 still leaves the T1+T2 saving above 70% under the Conservative case. The arbitrage holds; the staffing model is not absolute.

Section 7 — Regulatory and Accountability

Regulatory accountability is a deployer-side constraint that applies at every tier — not only at T3 — and the same constraints apply identically in Ireland and Italy as EU member states. The Stewardship Model exists precisely because accountability cannot be delegated to an agent, regardless of task complexity.

In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal held the airline liable for negligent misrepresentation after its customer-service chatbot provided incorrect refund-policy information. The defence — that the chatbot was a separate legal entity — was rejected. The deployer bears full legal and financial liability for the errors of its autonomous agents, including in routine T1 and T2 interactions. As a common-law jurisdiction within the EU, Ireland is doubly exposed to this precedent.

EU AI Act Article 14 requires human oversight for high-risk AI systems. Article 50 imposes transparency obligations on AI systems that interact directly with users. Article 86 grants individuals the right to an explanation of individual decision-making. GDPR Article 22 grants the right not to be subject to a decision based solely on automated processing where that decision produces legal or similarly significant effects — a category that plausibly captures T2 cases such as billing disputes, refund denials, and account suspensions. The NIS2 Directive imposes formal incident-handling obligations across critical sectors that map directly onto T3 data-security cases.

Every one of these obligations requires a named human in the loop. Every one of them is an obligation the Stewardship Model satisfies by design. The Steward is the named human oversight party; the audit log is the transparency mechanism; Deterministic Logging is the explanation infrastructure; the Proof of Action is the immutable ledger an auditor can replay. Competitors deploying pure agentic stacks without an architectural Steward layer face structural compliance debt that compounds with every regulatory iteration. Accountability is not the cost of the model. Accountability is the model.

Section 8 — What the Arbitrage Is Made Of

Two structural forces drive the arbitrage. Both must be understood for the magnitude in this analysis to be intelligible at all.

The cost asymmetry. Inference prices for production-grade models continue to decline, and the operating cost stack — model routing, caching, retrieval optimisation, workflow design — continues to compress the cost of any compliant resolution. The relevant operational metric is not the headline token price; it is cost per compliant resolution after QA, reopen, and audit. On that metric, the trend is downward and structural. Human labour costs are not declining. They are rising — at inflation plus labour market pressure. A function running on human customer care today faces a cost structure that deteriorates annually. A function built on the Stewardship Model has a cost structure that improves annually without any architectural change. The arbitrage does not require action to widen. It widens by default.

The throughput asymmetry at scale. The agentic stack does not degrade at volume. The 100,000th T1 ticket costs the same as the first. The Coordination Tax does not compound with output because there is no human coordination required between tasks. For a human team, scaling volume means scaling headcount, supervisory layers, and the variance that comes with a larger workforce. For the Stewardship Model, scaling volume means adjusting a rate parameter and monitoring a dashboard. The cost curve is structurally different in shape, not just in level — and the difference compounds.

The third option — offshore BPO

Operating in Ireland, a customer-care function has three operating options once human delivery is the status quo. Section 2 modelled domestic delivery at Irish costs. Section 3 modelled cross-EU migration to Italy as an EU-internal cost optimisation. The third option a competitor commonly raises is offshore BPO operations at $7 to $16 per hour (Manila, Bangalore, Cebu City). The implied compression of the arbitrage margin is real but does not invalidate the structural finding for three reasons.

First, the offshore route does not exit the EU regulatory layer. A Manila BPO handling EU customer interactions remains bound by GDPR, EU AI Act transparency, and NIS2 obligations under the deployer brand. The deployer is the EU-domiciled brand; compliance overhead does not relocate with the agents.

Second, offshore operations retain the Coordination Tax. Wage cost falls but supervisory layers, training overhead, and managerial coordination compound with headcount regardless of geography. The structural problem the Stewardship Model addresses is not labour cost per se; it is human coordination of high-volume, policy-driven tasks.

Third, offshore labour is itself subject to Labor-to-Compute Substitution pressure. The same pricing-decay curve that creates the arbitrage in Ireland and Italy creates it in Manila and Bangalore. The competitive horizon for an offshore BPO is not the EU operation it currently undercuts; it is an autonomous operator that serves the same EU customers with no offshore arbitrage at all. Offshore is a transition state. Autonomous is the equilibrium.

Section 9 — Generalisation Beyond Customer Care

Customer care is the use case modelled here because it is analytically clean and publicly understood. The finding generalises. Any function where the dominant operating cost is human coordination of high-volume, policy-driven tasks — and where T1 and T2 tasks represent the majority of the revenue loop — carries the same arbitrage profile.

The customer care sector is large enough across European telecommunications, financial services, e-commerce, insurance, travel, and utilities that even narrow tier-level arbitrage produces material market-entry opportunities. The majority of headcount handles T1 and T2 tasks under supervisory layers that add coordination overhead without adding resolution capacity. This is the structural condition that the Human-to-Logic Ratio identifies in market selection: markets where human labour accounts for more than 60% of gross margin and the Coordination Surface is large, deterministic, and uniformly distributed across all incumbents. Customer care is one instance. Back-office compliance processing, logistics coordination, claims adjudication, and professional services brokerage carry the same structural signature.

The thesis is not that all of these roles disappear. It is that the operators which rebuild these functions under the Stewardship Model will run at a structural cost advantage that compounds annually against competitors who do not. This analysis quantifies the first year of that compounding in our home market. The second year, with inference costs lower and the operational logic already proven, the gap is wider — without any additional investment. That is what Operational Arbitrage looks like when it is structural rather than tactical.

Five Findings to Lock

The modelled arbitrage is structurally tier-specific. T1 captures 93.8% cost reduction in the Irish base case and 67.6% in the stress case. T2 captures 86.9% base and 55.0% stress. T3 captures approximately 18% through cycle-time compression. These are the boundaries within which any operator captures the available arbitrage; the choice to operate at the bottom or the top of those ranges is architectural, not aspirational.

The Stewardship Model is the differentiator, not pure agentic deployment. The highest-value position is reducing the human role to exception authority and architectural oversight at T1 and T2, and to specialist leverage and governance at T3. One Steward at Irish reference volumes absorbs 18.4 FTE worth of T1+T2 work — until the cognitive load on exceptions exceeds 6.5 hours per shift, at which point the architecture needs a second Steward.

T3 requires explicit honesty. Legal, regulatory, and relationship-critical cases require human judgment because accountability cannot yet be delegated to an agent. The economic value is cycle-time compression and decision support, not labour substitution. This boundary is structural; it does not move when the technology improves.

The framework generalises across EU cost structures. Ireland captures €694K (95.4%) at T1+T2; Italy captures €653K (95.6%) on the same framework. The percentage capture is consistent because the Stewardship Model is architecturally identical across markets. The absolute saving scales with the human baseline being displaced — meaning a higher-cost market produces a larger absolute return at the same percentage capture.

The cost gap is widening. Inference prices continue to fall and human labour costs continue to rise. Every month of delay compounds the arbitrage further against incumbents who continue adding headcount to manage complexity the agentic stack could absorb under bounded workflows.