The perfect apology and the impossible refund

A customer disputes a duplicate charge. The service responds in seconds, names the frustration, apologises warmly and promises that the money will be returned by Friday. The exchange reads beautifully. The refund system has no request. The service principal lacks authority to issue one. On Friday the customer starts again by telephone, where an employee cannot see the promise.

Generation makes this failure cheap to reproduce. Text, voice, images and personalised replies can be supplied at almost no marginal effort. Yet the organisation's scarce resources remain stubbornly physical and institutional: current account state, lawful permission, specialist attention, decision authority, capacity to act, and a repair path when an apparent promise diverges from reality. Fluent language can disguise the scarcity rather than remove it.

The design unit is the customer promise transition. This is any claim, offer or action that changes what a customer can reasonably expect from the organisation. A balance explanation may carry little new commitment if it reflects an authenticated ledger. A refund deadline, eligibility statement or assurance that an action is complete carries much more. The latter must be bound to an authoritative source, an actor permitted to commit, an executable operation and a receipt that follows the journey.

This is not a chatbot-design problem. A journey can begin in search, continue through messaging, pass to a contact centre and complete in a back-office case. The Generated Interaction Operating System spans that route. It identifies the customer and unresolved objective; constrains data and claims; separates truth from inference; protects consent and channel choice; gates promises; composes language; transfers evidence; chooses the resolution capability; records commitments; monitors outcomes; and repairs common defects.

Put the contract before the composer

The Interaction Contract declares what one class of interaction may do before a model sees the request. It names the customer goal and organisational purpose, eligible journeys, authentication required by claim, permitted data, authoritative sources, maximum freshness, allowed recommendations, available remedies, disclosure, accessibility routes, customer correction and transfer conditions. Tone belongs in the contract, but tone is not authority to simulate understanding or a relationship the service cannot support.

Paper 071 used an interaction contract as one of six conditions for product-market fit. Paper 161 develops a service-specific contract for promise-bearing touchpoints. It must survive channel transfer and connect claims to fulfilment. Papers on entitlements and data exfiltration govern what information may be retrieved or released. Here, the additional question is what a customer may reasonably rely upon after receiving that information.

Break fluent answers into claims before they reach the customer

A generated paragraph can hide one dangerous sentence among nine correct ones. The Interaction Claim Graph decomposes the response into customer-supplied facts, observed journey events, authoritative account facts, retrieved information, calculations, policy interpretations, model inferences, options, reserved judgements, offers, promises, actions and fulfilment. Each node receives a source, freshness, verifier, permitted wording, authority, consequence and failure route.

The composer works only from approved nodes. It may join a verified balance, a deterministic interest calculation and a current policy condition into plain language. It must label a model inference as uncertain and cannot convert a possible cause into account state. An explanation is not evidence. A customer correction changes the state resolver or opens a dispute; it is not merely added to the transcript and forgotten.

The Promise and Authority Gate recognises when proposed language changes expectation. It locates current state, confirms identity and purpose, separates information from recommendation and commitment, verifies the principal's authority, validates amount and deadline, reserves judgement, forbids premature completion, records the promise and later compares it with reality. Identity, consent, price, eligibility, remedy, irreversible action and protected rights are hard gates.

Transfer the journey, not the burden

Generation should continue only while the channel can carry the evidence, authority and care needed for resolution. A customer request for a person is itself a transfer trigger. So are repeated loops, conflicting sources, vulnerability, accessibility need, negotiation, institutional accountability and a consequence window shorter than the receiving route's delay. Transfer is successful when it reaches the smallest capability that can resolve the objective without making the customer reconstruct the case.

The Continuity Packet contains verified identity level, the customer's goal in their own terms, current state, attempts and results, verified facts and conflicts, options and choices, open promises, permissions, accessibility and language needs, urgency, transfer reason, evidence links, prohibited actions and receiving acceptance. A generated narrative may help orientation, but it cannot replace the records that support material claims.

Thought experiment 1: the shorter conversation that creates more work

A chatbot's containment rises because it ends sessions after presenting three help articles. The dashboard reads success. Now include every customer who returns by phone, abandons silently, repeats the same question or misses a deadline. The apparent efficiency dissolves. The automation reduced visible interaction inside one channel while exporting investigation and reconstruction work to customers and staff.

The Experience Denominator begins with eligible, attempted, authenticated and admitted interactions, then follows self-service, generated, deterministic, specialist and human routes to resolution, repeat, abandonment, complaint, appeal and redress. Answer completion, task completion, authoritative action and customer resolution are separate states. A channel can complete an answer while the customer's objective remains untouched.

A promise ledger turns words into owned service

The Customer Promise Ledger is not a transcript archive. It stores the minimum evidence needed to reconstruct material commitments and protect rights: journey identifier, customer objective, channel and accessibility state, claims and supporting sources, options and exclusions, promises and conditions, authority, customer choices, transfer acceptance, action receipt, fulfilment, correction, complaint, redress and retention. Sensitive conversation remains under separate purpose and access controls.

Applied example: a pending payment across chat and phone

An authenticated customer asks why a card payment is pending. The contract permits explanation, not a settlement guarantee. The claim graph contains the current pending state, the authorisation time, general clearing information and several possible causes. Only the first two are customer-specific facts. The composer says what is verified, labels possible causes, explains when the status will next be checked and offers a dispute route if the customer does not recognise the payment.

The customer then asks for a person. The Continuity Packet carries the identity level, transaction reference, exact question, evidence already inspected, possible causes already described and the open promise to check status. The telephone employee accepts the packet, views primary records and does not repeat authentication beyond what the higher-risk action requires. If a dispute begins, the promise ledger switches from explanation to case commitment and records owner, deadline and receipt. The journey is resolved when the promised investigation action and customer notice occur, not when chat ends.

Applied example: common eligibility error at scale

A prompt change begins telling customers in several channels that a hardship option is unavailable. The answer cites a real policy but mistranslates one condition. Individual language appears plausible, so ordinary transcript sampling misses the pattern. Promise monitoring compares negative eligibility claims with subsequent specialist decisions and detects an abnormal disagreement concentrated in one customer segment.

The incident plane stops the affected prompt and knowledge combination, preserves the valid deterministic routes and identifies every exposed interaction from claim receipts rather than complaints alone. Teams prioritise customers near a deadline, issue corrections through suitable channels, restore options, extend time where authorised and record remediation. The old prompt, cached summaries and downstream scripts are removed. Reauthorisation requires the corrected policy mapping, multilingual claim tests, fresh transfer capacity and independent assurance. Average satisfaction never becomes the closure criterion.

Twenty boundary cases

Generated Customer Interaction Designer

The instrument tests a proposed interaction route, not the truth of a real customer case. Load one of ten scenarios or change the conditions. Its five outcomes distinguish factual generation, a bounded promise, evidence hold, continuity-preserving transfer and a stop where the channel cannot carry the consequence.

The strongest rival: generation is often the wrong service

The objection rejects generation by default. Remove avoidable contacts. Fix the product or policy that causes repeat demand. Use deterministic transactions for stable tasks. Generate where language adaptation, explanation or synthesis adds value. Preserve friction where consent and consequence require reflection. Route to a person when judgement, care, negotiation or institutional accountability is the scarce service. Store less conversation while preserving more verifiable promise continuity.

This is also an economic challenge to the architecture. A generated conversation can make the visible contact cheaper while moving expensive work into verification, correction, complaints and later human queues. That is not productivity if the customer must repeat the story or if staff reconstruct an undocumented promise. The relevant comparison counts the whole journey: demand avoided by a clearer product, successful deterministic completion, generated assistance, capable human work, follow-up, recovery and the cost of unresolved obligations. A slower route can be better when it reaches a durable outcome with less total effort.

Nor should continuity mean collecting everything. A complete transcript can contain intimate context, obsolete inferences and language that was never needed to resolve the case. The system should retain the smallest evidence set that lets the next actor verify identity, understand the goal, distinguish fact from inference, honour active commitments and explain the decision. Purpose limits, access controls, retention periods and correction routes belong inside the continuity design. Memory without those constraints is surveillance presented as convenience.

Thought experiment 2: the disclosure that changes nothing

Suppose a service clearly labels every reply as AI-generated. Customers now know the speaker is a machine. Does that make an invented refund, blocked complaint route or inaccessible control acceptable? It does not. Disclosure may help calibrate expectations, but research on disclosure and anthropomorphic cues is context-dependent. Trust can fall or rise without the service becoming more capable. The operating decision remains tied to claims, authority, action and rights.

Conversely, a human voice and name do not guarantee understanding or authority. The useful comparison is not machine versus person as a badge. It is which route can establish the customer's objective, inspect primary evidence, make the permitted decision, act and own recovery. A human reviewer approving every sentence can add cost while leaving the back-office action disconnected.

An observation would weaken this architecture: materially consequential generated interactions that reliably preserve cross-channel facts and commitments, complete authorised actions, reduce total customer effort, maintain segment outcomes and resolve complaints without claim-level gates, continuity packets or a promise ledger. Containment, satisfaction or conversational preference alone would not establish that result.

Assure the journey and give the panel stop power

The Generated Experience Assurance Docket holds the objective, population and state; Interaction Contract; claim graph and sources; disclosure, identity, consent, privacy and accessibility design; authority gate; route and capacity; Continuity Packet; promise ledger; ordinary, complex, stale, adversarial and vulnerable tests; customer effort and repeats; complaints, disputes, appeals and redress; segment and common-mode results; cost and displaced demand; residual risk; and authority for suspension, correction and retirement.

The operations panel keeps answer completion, task completion, authoritative action and customer resolution separate. It shows journeys, claims, freshness, permissions, promises, receipts, transfers, queue age, continuity, accessibility, repeat contact, broken deadlines, complaints, appeals, redress, segment differences and incident populations. Its stop control must disable the affected model, prompt, knowledge and channel combination while preserving valid alternatives.

Assurance must sample journeys, not only individual replies. A sentence can be accurate when written and still contribute to a broken experience if another channel contradicts it, the promised action never executes or the receiving team lacks authority. Tests should therefore follow representative cases from eligibility and admission through authentication, decision, action, receipt, follow-up and redress. They should include customers who abandon, switch channels, use assistive technology, communicate in another language or arrive with incomplete evidence. Those paths reveal the burden hidden by average satisfaction and containment rates.

A common-mode defect changes the response. If one prompt, retrieval rule or policy mapping can fabricate the same promise repeatedly, ordinary case-by-case correction is too narrow. The organisation must identify the exposed population, freeze the affected route, preserve evidence, issue consistent corrections and provide a practical remedy. Recovery is complete only when the erroneous expectation and any downstream action have been repaired, not when the model begins producing better sentences.

NIST's voluntary AI Risk Management Framework, which is under revision, and the Generative AI Profile support contextual testing, human roles, provenance, monitoring, feedback and incident response. WCAG 2.2 provides current accessibility criteria for web content, not proof that every customer's needs are met. The CFPB's chatbot report is sector-specific evidence about consumer-finance failure and human access. The FTC applies established substantiation principles to AI claims. These materials guide tests; none certifies an interaction route.

EU data-protection and AI rules apply according to legal scope, role, system and phased dates. Their transparency, data, rights and affected-person provisions should be interpreted with counsel for the actual service. ISO 42001, ISO 10002 and ISO 10004 inform management, complaint handling and customer-satisfaction processes, but certification or process conformance is not evidence that a disputed customer was made whole. UK government and US GAO guidance likewise remains bounded to its institutional purpose.

The Generated Interaction Reauthorisation Rule applies when purpose, population, product, policy, price, remedy, channel, model, prompt, knowledge, data use, disclosure, identity, consent, language, accessibility, tone, authority, action, transfer route, queue, right or law changes materially. It also applies after invented state, unsupported promise, false completion, blocked human access, accessibility failure, complaint spike, broken deadline, redress failure or common-mode incident.

Generate only what the organisation can stand behind

Abundant language makes a polished interaction easy. A good customer experience remains difficult because it depends on institutional memory, current truth, constrained authority, suitable channels, capable people and completed action. The organisation should therefore optimise neither conversation length nor automation containment. It should reduce the customer's effort to reach a verified and recoverable outcome.

Put the contract before the composer. Decompose claims before they become prose. Lock material promises to authority and receipts. Carry evidence and rights across channel boundaries. Preserve the customer's option to correct, refuse, complain and reach a person. When the promise and outcome diverge, find the affected population and repair the journey. Generated interaction earns its place only when the organisation can prove the facts, own the promise and complete the customer's actual objective.