Customer Service

As of August 2026. Phase 6. The most publicly visible agent deployment, and the most reversible.

1. The scenario

A consumer business handles contacts across chat, email and voice. Volume concentrates in a narrow set of intents: order status, returns, billing disputes, account changes, and a long tail of everything else. A knowledge base exists. Agents work from macros. Executives have seen a demo promising 70% deflection.

2. Agent team design

Agent	What it does	A x L position	Notes
Intent and triage agent	Classifies, retrieves the customer's own context, routes	A3, L2	The safest and most valuable agent in the department
Resolution agent	Handles the narrow, well-instrumented, recoverable intent set end to end	A3, L2	Scope by intent, never by channel
Account action agent	Executes bounded account changes within the customer's own scope	A3 on reversible actions, A2 on irreversible, L1	Refunds and commitments are gated. See controls
Agent-assist	Drafts and retrieves for a human handling the tail	A1, L1	Consistently the highest-return application in the department and the least demoed

Agent: Intent and triage agent
What it does: Classifies, retrieves the customer's own context, routes
A x L position: A3, L2
Notes: The safest and most valuable agent in the department

Agent: Resolution agent
What it does: Handles the narrow, well-instrumented, recoverable intent set end to end
A x L position: A3, L2
Notes: Scope by intent, never by channel

Agent: Account action agent
What it does: Executes bounded account changes within the customer's own scope
A x L position: A3 on reversible actions, A2 on irreversible, L1
Notes: Refunds and commitments are gated. See controls

Agent: Agent-assist
What it does: Drafts and retrieves for a human handling the tail
A x L position: A1, L1
Notes: Consistently the highest-return application in the department and the least demoed

3. Planes activated

Knowledge (direct: a governed knowledge base, never model memory), Human (direct: escalation to a real queue), Evidence (direct: conversations are legal evidence), Action (direct), Control (direct: a separate edge on the shared control plane), Improvement (direct: customer escalations label failures for free, making this the highest-value trace source in the enterprise), Execution (supporting).

4. Controls

Mandatory AI disclosure. Article 50 transparency duties are enforceable from 2 August 2026. No agent presents as a named human.
Customer identity and entitlement checked before any account action; the agent acts strictly within the customer's own scope.
Irreversible actions gated: refunds above a threshold, contractual commitments, account closure.
Escalation to a real human queue, wired with working-state transfer, before launch. Never manufacture containment through queue unavailability.
A vulnerable-customer path that bypasses the agent.
Full conversation retention as legal evidence.
Hardening for the public prompt-injection surface. Users here are anonymous and sometimes adversarial, unlike authenticated employees under an acceptable-use policy.

5. Economics

Per run. Roughly an order of magnitude below a human contact. Outcome-priced offerings list between $0.99 and $2.00 per automated resolution against a $6 to $12 human-handled comparator, all vendor list pricing as of mid-2026.

Per resolved outcome. Divide by resolution, and track repeat-contact rate beside it. A conversation that ends without escalation scores as contained whether the customer was helped or gave up, and re-contact multiplies true cost per issue while the dashboard improves.

6. Honest limits

Every documented reversal of an AI-first service programme set a containment or headcount target. This is the strongest single finding in the department and it is a target-setting failure, not a technology failure.
74% of surveyed senior decision-makers had already rolled back or shut down a customer AI communications agent after a governance failure, rising to 81% among organisations with mature guardrails. Maturity correlates with catching failures, not avoiding them.
Statements here bind externally. A tribunal held a company responsible for what its chatbot told a customer, rejecting the argument that the chatbot was a separate legal entity. A 2026 German appellate decision held that a chatbot is part of corporate communication rather than a third party, that liability attaches regardless of training-data provenance, and that general disclaimers do not provide sufficient protection.
No credible independent escalation-rate benchmark exists as of mid-2026. The guide states the absence rather than repeating circulating figures.
No independent production accuracy figures exist for voice agents; published latency distributions come from small vendor samples.

7. Metrics

Resolution rate. Repeat-contact rate within a defined window. Escalation by trigger. Containment reported only alongside resolution, never alone. Time to human when escalation fires. Complaint and regulator-contact rate. Cost per resolved issue.

Sources

research/R09-experience-and-channels/findings.md and sources.md. Disclosure duties: research/R11-governance-risk-sovereignty/. Trace value: ../../synthesis/learning-loops-map.md.

Source: blueprints/departments/customer-service.md in the evidence repository behind this site.

On this page