Latest
Deal Heidi cuts headcount by 18% and co‑founders leave after $475 million raiseDeal What buyers look for when they acquire an AI companyGlobal Build, buy or bring in an implementer: three ways to get an AI agentGlobal Airbnb to launch AI travel agent in first half of 2027Global AI agents versus traditional automation: what is the differenceUnited States Playful launches AI‑driven “vibe coding” platform for non‑technical creatorsGlobal AWS launches Kiro Crew bots to add persistent agents to its dev environment
United States edition
Friday, 6 November 2026
Infercline.
Advertise
The Agent Directory and Deals

How to choose an AI agent platform: questions to ask before you buy

There are hundreds of agent products and the labels blur. These are the questions that separate a platform you can rely on from one you will regret.

The Infercline Desk · Friday, 6 November 2026 Sign in to save
text
Photo by L N on Unsplash

Choosing an AI agent platform is harder than choosing most software. The market is crowded, products change every few months and the word agent is applied to almost anything. The good news is that the questions that matter are the same whatever you buy. They are about fit, control, cost, data and what happens if you need to leave.

Start by knowing the main kinds of product. Some agents come built into software you already use: Salesforce offers Agentforce inside its customer relationship tools, Microsoft offers Copilot Studio for building agents that work across Microsoft 365, and help desk vendors such as Intercom and Zendesk include service agents. Integration platforms such as Zapier offer agents that work across thousands of connected apps. Cloud providers offer platforms for technical teams to build and run their own agents, such as Amazon Bedrock AgentCore, which Amazon describes as a platform for running agents in production with tools for securing access and evaluating behaviour. Each kind suits a different business and a different level of in-house skill.

The first question is whether it does your job. Ask the vendor to run the product on one of your real tasks, using your own documents or records, during a trial. A demonstration on the vendor's example data proves little. Gartner warned in June 2025 about agent washing, where existing chatbots and automation tools are relabelled as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI were real.

The second question is control. What can the agent do on its own, and what needs a person's approval? Can you limit which systems and records it can reach? Can you see a log of every action it took and why? Can you switch it off instantly? The OWASP Gen AI Security Project lists excessive agency, giving an agent too much power, among the main risks of AI applications. A good platform makes it easy to give each agent only what its job needs.

The third question is cost, in detail. Agent platforms are priced by token, outcome, conversation, credit or seat, often in combination. At the time of writing (November 2026), Salesforce's Agentforce page listed a price per conversation and Flex Credits per action, and Microsoft's Copilot Studio page listed monthly capacity packs of credits. Ask for a worked estimate based on your own monthly volume, ask what happens if volume doubles and ask whether you can set a spending cap.

The fourth question is data. Where is your data stored and processed? Is it used to train the vendor's models? Who at the vendor can see it? These answers matter legally in every Infercline edition. South Africa's POPIA, in section 72, limits transfers of personal information to other countries unless conditions such as adequate protection or consent are met. Saudi Arabia's regulation on transferring personal data outside the Kingdom requires a risk assessment before transfer. The UK, Nigeria, Singapore, Australia, the UAE and US states have their own rules. Ask for the vendor's data processing agreement and list of sub-processors.

The fifth question is assurance. Ask what independent certifications the vendor holds. ISO/IEC 42001, published in December 2023, is the international standard for an AI management system, and some vendors now certify against it alongside long-standing security standards. A certificate is not a guarantee, but its absence is worth asking about.

The sixth question is what happens if you leave, or if the vendor changes course. Products in this market change quickly. OpenAI, for example, added a note to its AgentKit launch page on 3 June 2026 saying it was winding down its Agent Builder and Evals products, which would no longer be available from 30 November 2026. Ask how you can export your agent's instructions, data and logs, how much notice the vendor gives before retiring a product, and whether the agent depends on a single AI model you cannot change.

Why it matters to your business

An agent platform sits inside your operations, often with access to customers, money and records. Switching later is costly, and a poor choice can expose you to data protection breaches, runaway bills or a product that disappears. A short, consistent list of questions lets you compare very different products on the same terms and spot the weak answers before you sign.

What to do on Monday

  1. Write down the one job you want an agent to do first, and use it as the test for every vendor on your shortlist.
  2. Send each vendor the same six questions: fit, control, cost, data, assurance and exit. Compare the written answers side by side.
  3. Insist on a trial using your own data, with a spending cap set before it starts.
  4. Ask for the data processing agreement and sub-processor list, and check them against the data protection rules of each country you operate in.
  5. Prefer products that already sit inside software you use, unless they clearly fail the test job.

More on this topic: AI agent tools

Sources