Acme Copilot (Sample)

Accuracy

Declared accuracy metrics for Acme Copilot and the methodology used to measure them. Macro-F1 of 0.91, measured on a 4,000-ticket held-out set refreshed each quarter.

Robustness

Resilience to errors, faults, and inconsistencies, including feedback loops in systems that continue to learn. Model stack under test: OpenAI gpt-4.1 (drafting the reply); Anthropic claude-sonnet-5 (ticket classification)

Cybersecurity

Measures against attempts to alter use, outputs, or performance by exploiting vulnerabilities.

Training, validation, and testing data

Data sets used by Acme Copilot are subject to governance practices appropriate to the intended purpose.

Data flows

Ticket text and the customer's help-centre content enter the system over TLS. Inference runs in eu-west-1 with no data retained by the model providers under zero-retention terms. Drafts and the agent's final edit are stored for 30 days for quality review, then deleted. No customer content is used for training.

Quality criteria

Relevance, representativeness, and, to the extent possible, freedom from errors and completeness.

Bias examination

Examination in view of possible biases likely to affect health and safety or lead to discrimination.

Oversight measures

Acme Copilot is designed so it can be effectively overseen by natural persons during the period in which it is in use.

Every draft is reviewed by a named support agent before it reaches a customer. Agents see the retrieved sources alongside the draft, can edit or discard it, and can disable drafting per-queue at any time. Agents complete a 30-minute briefing on automation bias and on the ticket types where the system is known to underperform, notably billing disputes and multi-issue threads.

Capabilities of the overseer

Ability to understand capacities and limitations, remain aware of automation bias, correctly interpret output, and intervene or halt operation.

Automatic recording

Events in Acme Copilot are logged over the lifetime of the system to a degree appropriate to its intended purpose.

Scope

Ticket text and the customer's help-centre content enter the system over TLS. Inference runs in eu-west-1 with no data retained by the model providers under zero-retention terms. Drafts and the agent's final edit are stored for 30 days for quality review, then deleted. No customer content is used for training.

Retention

Logs under the provider's control are retained for the period appropriate to the intended purpose. Logs are retained for 18 months, then deleted automatically.

Risk management system

A continuous iterative process run across the entire lifecycle of Acme Copilot, requiring regular systematic review and updating.

Identification and analysis

Known and reasonably foreseeable risks to health, safety, and fundamental rights arising from the intended purpose: Support-ticket drafting for B2B software teams of 10 to 200 agents, in English, German and Portuguese. It is not designed or validated for medical, legal, financial-advice, employment or credit contexts, and is not intended for consumer-facing use without an agent in the loop.

Mitigation

Adoption of targeted measures, and residual risk judged acceptable. Oversight measures in place: Every draft is reviewed by a named support agent before it reaches a customer. Agents see the retrieved sources alongside the draft, can edit or discard it, and can disable drafting per-queue at any time. Agents complete a 30-minute briefing on automation bias and on the ticket types where the system is known to underperform, notably billing disputes and multi-issue threads.

General description

Acme Copilot: Acme Copilot drafts replies to inbound customer support tickets. It reads the ticket, retrieves the three most relevant help-centre articles, and proposes a reply for a human agent to send, edit, or discard. It never sends a reply on its own.

Intended purpose: Support-ticket drafting for B2B software teams of 10 to 200 agents, in English, German and Portuguese. It is not designed or validated for medical, legal, financial-advice, employment or credit contexts, and is not intended for consumer-facing use without an agent in the loop.

Development process

Model stack and computational resources: OpenAI gpt-4.1 (drafting the reply); Anthropic claude-sonnet-5 (ticket classification)

Data flows: Ticket text and the customer's help-centre content enter the system over TLS. Inference runs in eu-west-1 with no data retained by the model providers under zero-retention terms. Drafts and the agent's final edit are stored for 30 days for quality review, then deleted. No customer content is used for training.

Retention

Documentation is kept at the disposal of national competent authorities in EU, UK, Switzerland for the retention period required.

Purpose

This notice describes the intended purpose of Acme Copilot, its provider, and the information supplied to deployers.

Intended purpose

Support-ticket drafting for B2B software teams of 10 to 200 agents, in English, German and Portuguese. It is not designed or validated for medical, legal, financial-advice, employment or credit contexts, and is not intended for consumer-facing use without an agent in the loop.

System description

Acme Copilot drafts replies to inbound customer support tickets. It reads the ticket, retrieves the three most relevant help-centre articles, and proposes a reply for a human agent to send, edit, or discard. It never sends a reply on its own.

Model stack: OpenAI gpt-4.1 (drafting the reply); Anthropic claude-sonnet-5 (ticket classification)

Information provided to deployers

Characteristics, capabilities, and known limitations of performance, including the groups of persons on whom the system is intended to be used. Placed on the market in: EU, UK, Switzerland.