How to choose an AI platform for the whole company: 24 questions

Choosing an AI platform for the whole workforce is a three-to-five-year decision, not a license purchase. It's evaluated badly when compared by features and evaluated well when compared by questions: what happens to the data, who decides on the model, what stays with the company if you switch providers. This guide brings together the 24 questions a CEO or CIO should ask any provider, including us, before signing.

Before comparing providers: three internal decisions

Loose licenses or a company layer

Giving every employee an AI license produces individual productivity. What each person discovers stays on their screen. A company layer makes what works get shared, reused and governed. These are different purchases with different prices; decide which one you want first.

A pilot department or the whole workforce

Department pilots make sense for learning, but they produce islands: each team with its own tool and nothing shared. If the goal is for AI to become company capability, the rollout is company-wide and what gets staged is usage, not access.

What you're going to measure

Distrust anyone who promises a productivity percentage before knowing your company. What you can demand is measuring the baseline —hours of prep per salesperson, days to close the books, interruptions per engineer— and comparing it at three months.

The checklist: 24 questions

They're grouped into six blocks. Every question allows a clear answer; if the provider doesn't give one, you already have your answer. The downloadable version includes a scorecard for comparing three providers side by side.

Data and security

• Where is my company's data hosted, in what region and with which cloud provider?
• Is my data used to train models, whether by the platform provider or the model providers? Is it written into the contract?
• What certifications does the platform hold today (ISO 27001, SOC 2), and which are in progress, with a date?
• Is there the provider's own data processing agreement and a published list of subprocessors, with advance notice of changes?
• When was the last penetration test done, and who did it?

Models and dependency

• Which model providers are available, and who chooses the model for each task: the platform provider, the administrator or the user?
• If the best model for a task moves to another provider, how much does switching cost: configuration or a project?
• Do the agents my company builds depend on the model they were built with?
• Can I restrict which models each team or each agent uses?

Context and permissions

• Which of my company's systems can the platform connect to, and which of those connectors are tested on a real customer, not just listed in the catalog?
• Are permissions from the source systems inherited, so each person only sees what they could already see?
• How does the context stay current: automatic sync or manual document uploads?
• For systems with no direct connector (the ERP, for example), what's the route to their data?

Agents and governance

• Can an agent built by an employee be published for the whole company, with permissions and without depending on them?
• Is there an audit log of who uses which agent, with what data and which model?
• Can the administrator lock down the catalog of tools and connections, or can any user connect external services?
• What use cases does the provider exclude in its terms under the EU AI Act (candidate screening, for example)?

Cost and spend control

• Does the price include model usage, or is it billed separately? What happens if a user consumes far more than average?
• Can I see usage by user, team, model and tool?
• What prerequisites does the price have (licenses for another suite, user minimums, an annual commitment)?
• Is there a free trial, and what does it include?

Rollout and exit

• Who does the rollout, how long does it take, and what does it need from my team (identity, connectors, a lead per department)?
• What training and what materials does my staff get, in their own language?
• If I leave the service, what can I take with me: conversations, agents, knowledge? In what format, and on what timeline is my data deleted?

How to score it

The scorecard in the downloadable version assigns each question a weight (1 to 3) based on what matters to your company, and a score (0 to 2) per provider: 0 if they don't answer or don't have it, 1 if they have it with limitations, 2 if they have it and can show it. The weighted sum doesn't decide for you, but it orders the conversation and puts in writing why you chose what you chose. Two pieces of advice: get the answers in writing, and always ask "which customer is this actually working for today?" It's the question that separates the catalog from reality.

How Azimut answers

We'd rather you ask us these questions than avoid them. Hosted on Google Cloud, Madrid region. Four model providers under enterprise contracts that prohibit training on your data, chosen by the user per task. Agents published company-wide, independent of the model. Permissions inherited from the source systems and a connection catalog closed by default. Credits per seat and visible usage by user, team, model and tool. And two uncomfortable answers we'll still give you: we don't have ISO 27001 certification yet (it's planned, and we'll tell you the expected date instead of a vague promise), and we'll only offer you the integrations we've tested with you.

Download the checklist with scorecard

The PDF version includes all 24 questions with room for three providers, the weights and the scoring, ready to print or fill in. Leave us your work email and the download starts instantly.