How should a business choose an AI platform?

Choose an AI platform by testing it against a defined business task, the information it needs and the way your organisation will operate it. ChatGPT, Claude, Gemini, Grok, Microsoft tools and custom agents belong on a shortlist only where they meet those requirements. A strong answer in a demonstration is one part of the decision.

The purchase also establishes data access, working practices, support obligations and ongoing costs. Those decisions deserve attention before a tool becomes part of daily work.

Write a selection brief the business can agree on

Start with the outcome, users and scope. Describe the information the tool needs, where that information lives and whether the tool will advise a person or take an action.

For an illustrative internal guidance assistant, useful requirements might include finding the approved procedure, showing the source, respecting staff access permissions and referring an unclear question to the responsible team. Those requirements are more useful than asking which model gives the most impressive answer.

Separate essential requirements from preferences. If a platform cannot meet an access requirement or connect to an essential system, a better writing style may not compensate. Include business owners, intended users and the people responsible for technology and information in that judgement.

Assess the product and edition you will actually use

Product names can cover different consumer, business and developer offerings. Assess the purchased edition, settings, contract and connected services. Do not assume that an organisation’s business account and an employee’s personal account provide the same arrangements.

The following primary sources illustrate why that distinction matters. Product documentation was reviewed on 25 September 2026; confirm the current terms and configuration during procurement.

• ChatGPT: OpenAI’s business data documentation says business data is not used to train its models by default and describes controls for different offerings. Check which controls are included in the proposed plan.
• Claude: Anthropic’s commercial data retention guidance distinguishes API handling from products that retain conversations for continued use. Specify which service will receive your information.
• Gemini: Google’s Workspace privacy guidance applies to qualifying Workspace editions and distinguishes experiences within its services. Confirm the account and edition covered by the arrangement.
• Grok: The Grok Business user guide distinguishes personal and team workspaces. Establish which workspace staff will use and which business terms apply to it.
• Microsoft Copilot in Microsoft 365: Microsoft’s data and privacy documentation describes access through existing permissions, making the quality of those permissions part of implementation readiness. Confirm the specific offering and licence, including whether the proposal covers Microsoft 365 Copilot or Copilot Chat.

These are procurement checks, not a ranking or a declaration that a product is suitable for every business. Ask about retention, deletion, access administration, data processing, connected tools and the evidence available when something goes wrong.

Test the complete workflow

Give shortlisted options the same representative tasks and agreed source material. Use information you are authorised to provide. Include incomplete requests, conflicting documents and cases that should be referred to a person.

Assess whether the result is correct, usable and traceable to the information supplied. Record the work needed to prepare inputs, review outputs and fix mistakes. A polished response that takes longer to verify may not improve the process.

Test integration separately from response quality. Confirm how information enters and leaves the system, how permissions are applied and how a failed action is detected. A tool that drafts a good response has not thereby demonstrated that it can safely send the response or update a customer record.

Decide whether a custom agent is justified

A custom agent may be worth considering when the work needs a particular combination of tools, permissions and steps that an existing product cannot meet adequately. Define those requirements before selecting a framework.

Custom development creates responsibilities for testing, monitoring, access management and maintenance. Agree which actions the agent can take, which require approval, and how the workflow recovers from partial completion. There should be a named owner and an operating plan after the initial build.

Keep predictable calculations and business rules in components designed to enforce them. An agent can coordinate parts of a process without becoming the authority for every decision within it.

Compare ownership costs and make a reviewable decision

Include licences and usage alongside integration, administration, training, human review and support. Assess the effect of adding users or increasing workload, and how the business could move its information and processes later.

The recommendation should record the chosen approach, the alternatives considered, the test evidence, outstanding conditions and the person accountable for operation. It may support one platform, a limited combination or further preparation before purchase.

My AI advisory and implementation work connects these technology choices with commercial priorities, governance and the changes people will need to make.