How to Choose an AI Agent Platform: 2026 Checklist

The questions a vendor can only answer yes or no to

Most selection criteria published by vendors are unfalsifiable — "powerful models", "enterprise-grade", "seamless integration". A useful checklist contains only questions with a checkable answer.

1. Where is inference executed, and who are the named sub-processors? Where a vendor is headquartered tells you nothing about where your data is processed.
2. Do customer inputs train models? Ask for the clause, not the reassurance.
3. Can one agent be restricted to a subset of data? If not, every workflow inherits the blast radius of the largest one.
4. Is there a per-action audit log? Which agent, on whose behalf, what it read, what it changed, who approved it.
5. Which actions require approval by default? The default is the control. A gate that must be switched on will not be.
6. What happens at contract end? Export format, retention period, deletion guarantee.
7. Can a single workflow use more than one model? This decides whether switching models later is configuration or migration.

What it costs, and what the sticker leaves out

List prices as of August 2026: ChatGPT Business is $20 per user per month billed annually, Claude Team $20 per seat per month billed annually, and Microsoft 365 Copilot $30 per user per month on an annual commitment. The Copilot figure is an add-on — it requires a qualifying paid Microsoft 365 plan underneath it, so the all-in figure per user is materially higher. ChatGPT Enterprise has no published price and is quoted per deal.

Three costs never appear on a pricing page and routinely exceed the licence: connector and integration work to reach the systems the agent must read; the review time that approval steps consume, which is real salaried time; and the internal change effort of moving a process from how people do it now to how it will be done.

A per-seat comparison between categories is misleading for the same reason — an assistant subscription and an agentic workspace are not priced for the same job.

Red flags

Patterns that reliably indicate a platform is not ready for production use:

• No audit trail of agent actions. Without it you cannot answer, months later, what the agent did and on whose authority — which is the first question any auditor or incident review asks.
• No named sub-processor list. "We use industry-leading providers" is not a list.
• Agents that cannot be scoped. All-or-nothing data access means the permission model does no work.
• "Unlimited" with undocumented limits. There is always a limit; an undisclosed one surfaces at the worst moment.
• Demos that only ever show the happy path. Ask to see a failed run, a rejected approval, and a permission denial.

Running the evaluation itself

Evaluate on your own material. Vendor demo data is chosen to work; the question that matters is how the platform behaves on your least tidy documents, your inconsistent CRM records and your worst-audio meeting.

Give each shortlisted platform the same process and the same success measures — cycle time, share of runs needing correction, share reusable without edit. Comparing different processes across vendors produces a conclusion about the processes, not the vendors.

Include the people who will actually operate it in the trial. Adoption failure, not capability failure, is the usual cause of a shelved platform, and it is visible in a pilot if the pilot includes real users rather than the project team.

Sources

• OpenAI — ChatGPT pricing
• Anthropic — Claude pricing
• Microsoft — Microsoft 365 Copilot pricing
• GDPR, consolidated text (EUR-Lex)

All links verified on 26 August 2026. Prices are vendor list prices as of that date and do change; the vendor's own page is authoritative.