AI agent requirements: a template for your business

“Automate our administration” is too broad for a first test. “Prepare a draft follow-up every morning for quotes that are due” identifies an input, a trigger and an output. A document template or a CRM rule may sometimes be enough: keep that option open during scoping.

Anthropic recommends starting with a simple solution and adding complexity only when it provides a verifiable benefit. Our template turns that principle into business questions without prescribing a provider or subscription.

The file is free, requires no form and can be edited in a text editor. It is a discussion framework, not a universal specification or a compliance guarantee. Leave unknowns explicit instead of filling gaps with assumptions.

1. The task and its owner

Describe who does the work, how often and through which steps. Name the person who can judge the result. Record current preparation, processing and correction time; this baseline is not a promise of savings.

2. Inputs and access

List the required documents and applications, their owners and update frequency. Distinguish read, draft creation and editing permissions. Access to an entire workspace should not be the default. Test documents must be authorised for this use.

3. Output and human approval

Define the format, output location and references needed for review. Specify who approves sending a message, editing a record or making a commercial commitment. Also state which decisions must remain human.

4. Errors to test

Prepare a normal record, an incomplete one, a contradiction and an unavailable source. Define the expected behaviour in advance: produce a draft, flag an unknown or stop. Evaluate proposals on the same cases, not just on a successful demonstration.

5. Costs, maintenance and recovery

Separate scoping, development, subscriptions, tool usage, training and maintenance. Define who receives alerts, how to disable the workflow and how to resume manually. Ask what will be handed over: instructions, documentation, access and test cases.

A small business wants to prepare follow-ups from its CRM. The pilot reads only open quotes and authorised notes. It places drafts in a review queue without sending them or changing amounts. If a status is contradictory or information is missing, the record is flagged for manual handling. This is an educational example, not a client result.

The criterion is not “the text sounds natural”. The reviewer checks the recipient, reference, facts and absence of new commitments. They also compare total working time with the current method. Acceptable thresholds are agreed before testing; this guide does not invent them for you.

Ask each provider to answer the same template. Compare included functions, dependencies, human interventions and exclusions. One offer may exclude data migration or training; another may add unnecessary complexity. Prices are comparable only with their assumptions and limits.

The Champions Academy CRM illustrates software built around real processes, without claiming to be an autonomous agent. For your project, explore internal administration, the business AI agent scope and workflow training.

To specify reading, drafts and permitted actions, use our AI agent and CRM permissions matrix, with six tests before release.

Reference: Anthropic: designing agents, accessed on 30 September 2026. The template and example are Partna’s editorial proposal. They do not replace an examination of your systems or validation of obligations specific to your business.