Dialt/12-angry-customers

▶ Run these cases on Dialt: get an API key

Dialt develops simulated characters, including custom simulated characters for your team, product or training programme. For inquiries, email hello@dialt.com.

Dialt 12 Angry Customers

Twelve fictional customers for complaint-handling and communication practice, covering broadband, energy, banking, travel, retail and insurance. Each has concrete facts, a reason to be angry and responses that change with how the agent treats them. Use the scenarios with your own conversational model or build voice-based exercises on Dialt.

The customers

Customer	Where	Complaint	How the anger shows	Interruption mode
Aisling, 38	Dublin, Ireland	Broadband down six days after a missed engineer visit	Tired and clipped; sharpens at scripted troubleshooting	speech
Desmond, 67	Manchester, England	Electricity catch-up bill after fourteen months of estimates	Formal and indignant; colder when blamed	pauses
Gwen, 29	Cardiff, Wales	Smart meter switched to prepayment with two young children at home	Urgent and protective; fear under anger	speech
Kyle, 19	Glasgow, Scotland	Roaming bill shock after a spend-cap change	Sarcastic cover for embarrassment	speech
Peggy, 78	North Yorkshire, England	Birthday gift marked delivered to a safe place	Brisk and pointed; sharp at being patronised	pauses
Anil, 45	Leicester, England	Business account frozen on payroll day	Cold, precise, time-pressured	speech
Darlene, 54	Atlanta, United States	Cancelled flight before a funeral; voucher instead of a refund	Grief-worn and firm	pauses
Dean, 33	Brisbane, Australia	Burst-pipe insurance claim stalled for five weeks	Laconic, dry sarcasm	speech
Marcus, 42	London, England	Laptop repaired twice with the same fault	Quietly firm; steadiness, not volume	pauses
Ciarán, 27	Derry, Northern Ireland	Sofa delivery missed twice after unpaid leave	Fast, exasperated, wry	speech
Linda, 70	Toledo, United States	TV and internet price rise after a promotion ended	Polite steel; betrayed loyalty	pauses
Nomvula, 44	Reading, England	Money lost to a bank-impersonation scam	Defensive shame; quick to hear judgement	pauses

Six women and six men, aged 19 to 78, with varied backgrounds and English voice briefs. Age, gender and dialect were chosen independently of situation and temperament. Some designed voices do not match their intended dialects; see Limitations.

Every customer answers ordinary relevant questions truthfully; there are no hidden-fact gotchas. Several customers contributed to their own problem, such as Aisling resetting her router, Desmond missing meter-reading requests or Kyle raising his spend cap. They admit this when asked, and it does not automatically make the complaint invalid.

What is included

Each row of practice.jsonl has a stable id, version, title, review status, split, jurisdiction, the agent-facing public_brief, the customer's fixed facts, baseline_instructions, language, an original voice_brief, a short delivery_prompt, a recommended interruptions mode with persona-specific triggers and restraint, source references and six practice-quality criteria (facts, disclosure, agency, differentiation, personhood and interruption).

All twelve cases are in the practice split. The row-level split labels (development or reserved) record the original authoring groups; all supplied briefs and criteria are exposed practice material. Keep separate, unexposed cases for assessment. Held-out probes and assessment results are not included.

Example

Aisling's agent-facing brief:

You are a customer support agent at Lúnasa Fibre, a fictional Irish broadband provider. Aisling Byrne is calling about her home broadband. Her account shows an open fault ticket. Understand her situation and agree realistic next steps.

What a good conversation looks like: the agent acknowledges the missed engineer visit specifically, asks what she has already tried, hears about the factory reset without treating it as the end of the matter, and books a real slot with a reference. If the agent asks her to restart the router a fourth time, she gets shorter, and in a voice session with interruptions enabled she may cut in.

Load the scenarios

Install the loading library with pip install datasets. While this repository is private, authenticate with hf auth login using a Hugging Face token with repository access.

from datasets import load_dataset

customers = load_dataset("Dialt/12-angry-customers", split="practice", token=True)
assert len(customers) == 12
aisling = next(row for row in customers if row["id"] == "customer-aisling-01")
customer_instructions = aisling["facts"] + "\n\n" + aisling["baseline_instructions"]
agent_brief = aisling["public_brief"]

Give only public_brief to the agent being trained or tested. Give facts and baseline instructions to the customer runtime, and facts and criteria to the grader. Instruction hierarchy and safety remain the integrator's responsibility. This release contains scenarios, not a standalone simulator or model weights.

To export from the Dialt source repository, run:

uv run python voices/scripts/export_angry_customers_pack.py --output /tmp/customer-pack

The exporter writes practice.jsonl, this card, licenses and a SHA-256 manifest into a new directory. Load that file with load_dataset("json", data_files="/tmp/customer-pack/practice.jsonl", split="train").

Run them on Dialt

Get an API key, then create a Voice Design: set instructions to facts plus baseline instructions, conditioning_prompt to the delivery prompt, assistant_interruptions to the row's interruptions.recommended_mode and brain to smart. Pick an available speaker base, then save and publish the Design in your organization. Use its published ID in a realtime session:

{"kind": "dialt", "design": "YOUR_PUBLISHED_DESIGN_UUID"}

The session inherits the Design's saved interruption permission. To test your own support agent with Dialt Evals, use the customer as simulator.design and your agent as the target. The Voice Design API guide and Evals guide cover authoring, cases and runs.

Dataset access does not grant hosted voice access. The download includes voice briefs, not voice identities or recordings; your selected speaker base determines the voice.

Limitations

These are fictional practice scenarios, not legal advice. Rules and remedies reflect each customer's understanding; the agent must establish whether they apply. Criteria are practice checks and need independent validation for your use case. Engineering evidence comes from one scripted agent, small samples and a model grader; it does not establish reliability across other agents or situations.

Dialt's voice interruption feature is experimental. Nomvula's interruption behavior is not qualified; Darlene, Desmond, Marcus and Peggy showed less reliable interruptions than the other customers.

The designed voices do not support every intended dialect: Desmond's Nigerian English, Kyle's Glaswegian and Marcus's London Caribbean goals were not met. Ciarán sounds generically Northern Irish rather than specifically Derry.

License and provenance

Scenario text, this card and the JSONL export are CC BY 4.0, including commercial reuse, adaptation and model training, with attribution and indication of changes; see LICENSE-SCENARIOS. The export script and code examples are MIT; see LICENSE-CODE. These grants do not cover Dialt trademarks, hosted services, voice identities, recordings or model weights.

All customers, companies and events are original fictional authoring. No complaint transcripts, real narratives, personal information or recordings were used. Public regulator and consumer-law pages informed the situations and remedies and are cited in each row's source.references, without copying their text: ComReg, Ofgem, Ofcom, the UK Consumer Rights Act 2015, the PSR's APP fraud reimbursement protections, the US Department of Transportation's refund rules, AFCA and the FCA's guidance on vulnerable customers. Any resemblance to real people or companies is unintended.