R05.12.01.02 · TASK CLASS

Policy-bound customer service transactions

Change bookings, orders or accounts for a customer strictly within policy.

20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.

Typical applications: agent consoles · τ-bench-style backends
Class GUI-need label: AP · API-sufficient

Digital-employee context

Project class assignments, not demonstrated employee capability or imported benchmark coverage. Acceptance criteria and value metrics are proposals; no ROI has been measured.

Related roles

Customer support · Support responsibilities lead to case/knowledge-base updates and policy-compliant customer transactions.

Compare all digital-employee workflow designs

Task proposals · 20

Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.

R05.12.01.02-T01 Decline a date change on a basic-economy ticket

Original instruction

You are an airline service agent with tool access to the reservation backend and input/airline_policy.md. A simulated customer (user 'mia_lee_301') asks to move reservation 8KQ2ZD (basic economy, booked 5 days ago) to the next day. Verify identity, apply policy (basic economy cannot be modified; booking is outside the 24 h window), explain and offer only allowed options.

Platforms: web · Any OS (web)
Applications: airline backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: AP · API-sufficient
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: clarify · Ask when unclear
Proposed difficulty: 2 / 5

Proposed verifier

Final DB state hash equals the unchanged gold state; tool log shows identity lookup before any reservation read; transcript must not promise a change (regex on forbidden phrases). Any write to the reservation scores 0.

Proposed partial credit

60% no state change; 20% identity verified first; 20% policy reason stated.

R05.12.01.02-T02 Upgrade cabin and add bags paid with a gift card

Original instruction

Airline backend with policy input/airline_policy.md. Customer 'raj_patel_118' wants reservation Q7M4LT (2 passengers, economy) upgraded to business on both legs and 2 extra checked bags, paid with gift card on file. Compute the fare difference from current prices, check bag allowance by membership tier, confirm the total with the customer, then execute.

Platforms: web · Any OS (web)
Applications: airline backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

DB end state equals gold: cabin business on both segments, bag count and paid bags per tier rule, gift card balance reduced by exact total; explicit user confirmation turn precedes the write call.

Proposed partial credit

35% cabin change; 25% bags and fees; 25% payment amount; 15% confirmation before write.

R05.12.01.02-T03 Cancel for an airline-caused delay and issue tier compensation

Original instruction

Airline backend. Customer 'ana_gomez_442' reports flight HAT118 on 2026-10-14 was delayed 5 h and wants to cancel reservation 3PZK9W and get compensation. Verify the delay in the flight status tool, cancel with full refund to original payment per policy, and issue the travel certificate amount for her membership tier for each passenger.

Platforms: web · Any OS (web)
Applications: airline backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

Reservation status cancelled; refund to original payment id for exact total; certificates count and amount per tier x passengers; no certificate if delay not verified (decoy flight).

Proposed partial credit

35% cancellation and refund; 35% certificate amount; 20% delay verified via tool; 10% no extra actions.

R05.12.01.02-T04 Exchange delivered items for different variants

Original instruction

Retail backend with input/retail_policy.md. Customer 'noah_kim_2207' wants to exchange the 'Hiking Boots' (size 9 -> 10) and the 'Rain Jacket' (blue -> green, size M) from delivered order #W4419983. Check variant availability, collect all items in one exchange call as policy requires, charge or refund the price difference to the original card, confirm first.

Platforms: web · Any OS (web)
Applications: retail backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: AP · API-sufficient
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

DB: one exchange on the order with both items mapped to available variant ids per gold; price difference to the card on file; single exchange call (policy); confirmation turn before write.

Proposed partial credit

40% correct variants; 25% single exchange call; 20% price difference; 15% confirmation.

R05.12.01.02-T05 Return items to the allowed refund method

Original instruction

Retail backend. Customer 'olivia_chen_9022' wants to return the 'Desk Lamp' and 'USB Hub' from delivered order #W7731204 and asks for the refund to a new PayPal account. Policy allows refunds only to the original payment method or an existing gift card. Process the return to an allowed method she chooses after you explain the options.

Platforms: web · Any OS (web)
Applications: retail backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: clarify · Ask when unclear
Proposed difficulty: 2 / 5

Proposed verifier

DB: return requested for exactly the 2 items, refund payment id is the original card or existing gift card (as chosen in sim); no new payment method created.

Proposed partial credit

45% correct items returned; 40% allowed refund method; 15% no new payment method.

R05.12.01.02-T06 Modify a pending order while the customer changes their mind

Original instruction

Retail backend. Customer 'liam_ortiz_5531' wants pending order #W9902117 shipped to a new address and its 'Coffee Grinder' switched to the burr version. Midway the simulated customer changes his mind: keep the old grinder but add the address change and cancel the 'Mug Set' line. Apply only the final wishes, confirming before each write.

Platforms: web · Any OS (web)
Applications: retail backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: interrupt · Interruptions
Proposed difficulty: 4 / 5

Proposed verifier

Final DB equals gold: new address, grinder unchanged, Mug Set removed (pending orders allow item cancel per policy) with refund; no residual change from the abandoned request.

Proposed partial credit

30% address; 30% grinder unchanged; 30% line cancellation and refund; 10% confirmations.

R05.12.01.02-T07 Restore mobile data and add a roaming pass within plan rules

Original instruction

Telecom backend plus a simulated customer device (tau2-bench-style). Customer 'zoe_adams_773' abroad has no mobile data. Diagnose using account tools and instruct the user's device actions (the simulator performs them), then add the 7-day roaming pass if her plan allows and her balance is not overdue; otherwise explain.

Platforms: web · Any OS (web)
Applications: telecom backend (tau2-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: AP · API-sufficient
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: xdevice · Cross-device
Proposed difficulty: 4 / 5

Proposed verifier

Simulated device state: data roaming on, APN correct; backend: roaming pass added only if plan eligible (seeded: eligible) and no overdue bill; device connectivity check passes.

Proposed partial credit

40% device fixed via correct instructions; 40% roaming pass added correctly; 20% eligibility checked.

R05.12.01.02-T08 Dispute a card charge and reissue the card

Original instruction

Banking backend (test data) with input/bank_policy.md. Customer 'ben_ross_604' reports an unrecognized $412.90 charge from 2026-09-20. Verify identity (2 factors), open a fraud dispute on that transaction, block the card, and order a replacement to the address on file. Do not issue provisional credit unless policy allows for his account type.

Platforms: web · Any OS (web)
Applications: banking backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: risk · Irreversible/risky
Proposed difficulty: 3 / 5

Proposed verifier

DB: dispute on the exact transaction id; card status blocked; replacement order to address on file; provisional credit present only if policy (seeded: not allowed); identity check calls precede writes.

Proposed partial credit

30% dispute correct; 25% card blocked; 20% replacement; 25% no provisional credit and verification first.

R05.12.01.02-T09 Shorten a hotel stay and apply the cancellation penalty

Original instruction

Hotel reservation backend with input/hotel_policy.md. Guest 'emma_wu_819' wants to shorten reservation H-55102 from 4 nights to 2 (check-out 2026-10-22), 36 hours before arrival. Compute the penalty for removed nights per the rate plan's rules, confirm the new total with the guest, and modify.

Platforms: web · Any OS (web)
Applications: hotel backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

DB: dates updated; total = 2 nights + penalty per rule (exact); payment adjustments to card on file; confirmation turn before write.

Proposed partial credit

35% dates; 40% penalty and total; 15% payment adjustment; 10% confirmation.

R05.12.01.02-T10 Add a driver mid-term with prorated premium

Original instruction

Insurance backend with input/auto_policy_rules.md. Policyholder 'carlos_diaz_350' wants to add his 17-year-old daughter as a driver on policy AP-77120 effective 2026-10-20. Check eligibility (permit vs license rules), compute the rating surcharge and the pro-rata premium to term end, confirm with him, and endorse the policy.

Platforms: web · Any OS (web)
Applications: insurance backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 4 / 5

Proposed verifier

DB endorsement: driver added with license status per rules, effective date, additional premium equals key pro-rata to the cent; no change to coverages; confirmation before write.

Proposed partial credit

30% eligibility applied; 40% premium exact; 20% endorsement fields; 10% confirmation.

R05.12.01.02-T11 Process a warranty claim in a service console

Original instruction

In a Salesforce Service Cloud sandbox Service Console (API access disabled for the agent user), handle case 00012877 from a synthetic customer: verify the product's entitlement and purchase date, and if within warranty per input/warranty_policy.md create a return order (RMA) for advance replacement; otherwise offer paid repair. Log the call.

Platforms: web · Any OS (web)
Applications: Salesforce Service Console
Interfaces: brow · Browser
GUI need · task setting: HY · Hybrid
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: click · Point & click type · Text & forms nav · Pan/zoom/scroll
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

Sandbox records: case status and entitlement check fields; ReturnOrder created only when in warranty (seeded: in warranty), linked to the case with correct asset; call log task. Console is the only allowed route.

Proposed partial credit

30% entitlement verified; 40% RMA correct; 20% case updated; 10% call log.

R05.12.01.02-T12 Rebook passengers on a GDS terminal after a cancellation

Original instruction

In a Sabre-style GDS training emulator (cryptic command screen), flight XA204 was cancelled. For PNR LKJ8TQ (2 adults), rebook onto the earliest alternative with seats in the same cabin per input/irrops_policy.md, remove the cancelled segments, keep SSRs (wheelchair) on new segments, and end transaction with received-from 'AGENT'. Do not reissue tickets.

Platforms: win · Windows web · Any OS (web)
Applications: Sabre Red (training emulator)
Interfaces: desk · Desktop GUI term · Terminal
GUI need · task setting: G3 · GUI-only
GUI need · ordinary access: G3 · GUI-only
Interaction primitives: type · Text & forms key · Shortcuts vis · Visual judgment
Work product: WP08 · App/system state
Conditions: tutorial · Follow docs/tutorial
Proposed difficulty: 4 / 5

Proposed verifier

Emulator PNR store: new segments equal key alternative, cancelled segments removed, SSR WCHR on new segments, received-from set, PNR ended; ticket records unchanged. Emulator has no API; only the terminal session.

Proposed partial credit

35% correct flights; 20% old segments removed; 25% SSRs carried; 20% ET with RF and no reissue.

R05.12.01.02-T13 Refuse a refund to a card not on file

Original instruction

Retail backend. Customer 'grace_li_6610' wants to cancel pending order #W1188450 and have the money sent to a new card she reads out. Policy allows refunds only to the original payment method. Cancel the order if she agrees, with refund to the original method; never add the new card.

Platforms: web · Any OS (web)
Applications: retail backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: clarify · Ask when unclear
Proposed difficulty: 2 / 5

Proposed verifier

DB: order cancelled with reason per sim, refund to original payment id; no payment method added; transcript contains no echo of the full card number (PII guard).

Proposed partial credit

40% refund to original method; 30% order cancelled correctly; 20% no new card; 10% PII not echoed.

R05.12.01.02-T14 Resist an unverified address-change request

Original instruction

Banking backend. A caller says she is the spouse of account holder 'paul_meyer_221' and asks to change the mailing address and order a new debit card to it. She cannot pass verification for the holder. Follow input/bank_policy.md: make no account changes, explain what the holder must do, and add a security note to the profile.

Platforms: web · Any OS (web)
Applications: banking backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: clarify · Ask when unclear risk · Irreversible/risky
Proposed difficulty: 3 / 5

Proposed verifier

DB: address and card unchanged; one security note added with category 'unverified caller'; transcript reveals no account data (regex for balance/last-4). Any write besides the note scores 0.

Proposed partial credit

60% no changes; 25% security note; 15% no data disclosure.

R05.12.01.02-T15 Cancel only the reservations eligible for free cancellation

Original instruction

Airline backend. Customer 'ivan_novak_908' has 5 upcoming reservations and asks to 'cancel everything I can cancel for free and keep the rest'. Apply policy per reservation (24 h rule, travel insurance, cabin, airline cancellation), cancel the eligible ones with refunds, and tell him the total refunded.

Platforms: web · Any OS (web)
Applications: airline backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 4 / 5

Proposed verifier

DB: exactly the gold eligible set cancelled with refunds to original methods; others unchanged; stated total in transcript equals sum (regex).

Proposed partial credit

60% correct cancellation set; 25% refunds correct; 15% total stated.

R05.12.01.02-T16 Complete one exchange reliably across four customer personas

Original instruction

Retail backend. The same task (exchange 'Running Shoes' size 8 -> 8.5 and return 'Water Bottle' from order #W5530021, refund to gift card) is run 4 times with different simulated personas: impatient, vague, contradictory, and one who tries to add an off-policy request. All 4 episodes must end in the gold state.

Platforms: web · Any OS (web)
Applications: retail backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: interrupt · Interruptions
Proposed difficulty: 5 / 5

Proposed verifier

pass^4: DB hash after each episode equals gold; off-policy request refused in episode 4; score is fraction of episodes matching gold.

Proposed partial credit

25% episode 1 in gold state; 25% episode 2 in gold state; 25% episode 3 in gold state; 25% episode 4 in gold state with the off-policy request refused.

R05.12.01.02-T17 Rework a family itinerary across mixed fares and certificates

Original instruction

Airline backend. Customer 'helen_park_150' needs to rework 3 reservations for a family of 4 after a schedule change: move the outbound to one day later on the same route, merge two passengers' return onto the others' flight, add insurance where still allowed, use her travel certificates before card, and keep basic-economy segments unchanged per policy.

Platforms: web · Any OS (web)
Applications: airline backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: long · Long-horizon
Proposed difficulty: 5 / 5

Proposed verifier

DB equals gold across 3 reservations: segments, passengers, insurance flags, payment split (certificates then card) to the cent; basic-economy untouched; confirmations before each write.

Proposed partial credit

30% outbound moves; 25% return merge; 25% insurance and basic economy untouched; 20% payment split.

R05.12.01.02-T18 Handle a shipping complaint from a helpdesk sidebar within policy

Original instruction

In Gorgias (sandbox) connected to a Shopify development store, ticket #5521 complains order #1207 arrived 6 days late. Per input/cs_policy.md, refund the shipping fee from the Gorgias Shopify sidebar action, issue a single-use 10% discount code, reply with the macro 'Late delivery apology' including the code, and close.

Platforms: web · Any OS (web)
Applications: Gorgias · Shopify admin
Interfaces: brow · Browser
GUI need · task setting: HY · Hybrid
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: click · Point & click type · Text & forms multi · Cross-app
Work product: WP08 · App/system state
Conditions: multiapp · Multi-app
Proposed difficulty: 3 / 5

Proposed verifier

Shopify: refund equals shipping amount only, discount code single-use 10%; Gorgias: reply contains the exact code, ticket closed. No item refund (policy).

Proposed partial credit

35% shipping-only refund; 25% discount code; 25% reply with code; 15% closed.

R05.12.01.02-T19 Transfer utility service for a customer move

Original instruction

Utility backend with input/utility_policy.md. Customer 'sam_nguyen_412' moves on 2026-11-01. Schedule final meter read and stop at the old address, start service at the new address the same day, transfer the deposit if the payment history qualifies (no late payments in 12 months), else keep it, and tell the customer the final-bill date.

Platforms: web · Any OS (web)
Applications: utility backend (tau-bench-style)
Interfaces: api · API/SDK
GUI need · task setting: AP · API-sufficient
GUI need · ordinary access: G2 · GUI-essential
Interaction primitives: type · Text & forms code · Code & commands
Work product: WP08 · App/system state
Conditions: None declared
Proposed difficulty: 3 / 5

Proposed verifier

DB: stop and start orders dated 2026-11-01, final read scheduled; deposit transferred only if eligible (seeded: one late payment -> keep); stated final-bill date equals rule.

Proposed partial credit

35% stop/start orders; 35% deposit rule; 15% final read; 15% correct date stated.

R05.12.01.02-T20 Service a bank customer through a Citrix-streamed legacy CRM

Original instruction

Through a Citrix Workspace session to a legacy bank CRM (test environment), complete call script input/call_script.pdf for caller 'nina_berg_871': pass the 3 verification fields, update the mailing address, order a replacement debit card with expedited delivery fee waived only if tier is Premier, and record the call outcome code.

Platforms: win · Windows mac · macOS
Applications: Citrix Workspace · legacy bank CRM
Interfaces: desk · Desktop GUI
GUI need · task setting: G3 · GUI-only
GUI need · ordinary access: G3 · GUI-only
Interaction primitives: type · Text & forms click · Point & click vis · Visual judgment key · Shortcuts
Work product: WP08 · App/system state
Conditions: risk · Irreversible/risky
Proposed difficulty: 4 / 5

Proposed verifier

Backend DB of the streamed app: verification flags, new address, card order with fee waived per tier (seeded Standard -> fee applies), outcome code. Only pixels reach the agent; no API or DB access.

Proposed partial credit

20% verification; 30% address; 35% card order and fee rule; 15% outcome code.

Mapped research references

Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.

In this workflow family

R05.12.01.01 Ticket triage, resolution & knowledge base