Five Questions Every Business Leader Must Ask an AI Vendor Before Signing

Sales demos for artificial intelligence tools are engineered to impress. You see a pristine workflow where customer records summarise themselves, emails draft in seconds, and spreadsheets analyse their own trends without a single formula error. Then comes the annual subscription contract, often framed as a modest per-seat investment that seems negligible compared to the promised productivity gains.

If you run a business with 10 to 250 staff, you cannot afford to buy software based on a rehearsed demo. AI procurement carries operational and legal risks that traditional software never introduced. When a system actively processes, interprets, and generates information using your internal data, the stakes shift from standard software management to fundamental business governance.

Before you commit your budget or hand over access to your operational files, pause the procurement process. Demand clear, written answers to these five questions.

1. Is our data used to train your models, and where does it live?

Many software vendors have bolted AI capabilities onto their existing platforms by calling external application programming interfaces (APIs) provided by companies like OpenAI, Anthropic, or Google. Others host open-source models on their own cloud infrastructure. In either scenario, you need to understand the lifecycle of your company data.

Ask the vendor to walk you through the data flow step by step:

• Does our input data, file content, or prompt history ever enter a training pool for your public or shared models?
• Are our prompts and documents processed in a dedicated, isolated tenant, or are they mixed with other customers in a shared environment?
• In which geographic region are the processing servers located, and do they comply with UK GDPR or relevant local data residency laws?
• How long are prompt logs retained on your servers, who has access to them, and when are they permanently purged?

A vendor may tell you verbally that they take privacy seriously. That is not enough. Ask them to point directly to the clauses in their Data Processing Addendum (DPA) and Terms of Service that legally prohibit them from using your commercial data for machine learning improvement. If those protections require an expensive enterprise-tier upgrade, you need to know that now.

2. What is the total cost of ownership beyond the base licence?

Hidden costs in AI deployments are frequent and frustrating. In traditional software, a user licence typically covers unlimited standard usage. With generative AI, every query consumes computational power, which means vendors often structure their pricing to pass those compute costs back to you.

Clarify these commercial details before agreeing to an annual contract:

• Does the per-user fee include unlimited access, or is it capped by credits, tokens, or query volumes?
• What happens when a team member hits their monthly limit - does the tool stop working, or does your card get charged automatic overage rates?
• Are there mandatory onboarding, setup, or data preparation fees required before we can use the software effectively?
• Does customer support include timely human intervention, or are smaller accounts pushed to an automated chatbot?

For smaller businesses, predictability is essential. A software licence that looks attractive at £25 per user per month can quickly become unsustainable if your team incurs variable usage penalties or requires thousands of pounds in custom configuration just to connect to your existing databases.

3. How does the tool respect our existing permission structures?

This is one of the most dangerous blind spots for growing organisations. Generative search tools index large volumes of company data so users can ask questions across SharePoint, Google Drive, or your CRM. However, if the AI tool does not strictly mirror your access control policies, it will inadvertently expose sensitive files.

Put these questions to their technical team:

• How does your platform integrate with our identity provider, such as Microsoft Entra ID or Google Workspace?
• If an employee does not have permission to view a folder containing payroll data, management meeting minutes, or redundancy plans, will the AI exclude that information from their search queries?
• Can administrators run audit reports showing which files were indexed and which users accessed specific data through the AI?

Tools like Microsoft Copilot are designed to respect the permissions already established in your Microsoft 365 environment. However, this also means that if your internal permissions are currently untidy - such as an open company-wide folder containing sensitive contracts - the AI will find and surface that information to anyone who asks. You need to know whether the vendor helps you identify permission vulnerabilities before deployment.

4. Who carries the liability when the model produces errors or hallucinations?

All large language models can produce inaccurate information with complete confidence. In a sales or internal advisory context, an inaccurate number in a customer quote or a false statement in a compliance check can cause genuine financial and reputational harm.

Review the vendor's liability terms carefully:

• Does your service level agreement (SLA) guarantee output accuracy, or does your contract disclaim all liability for errors, hallucinations, and omissions?
• Do you provide an intellectual property indemnity clause that protects our business if the model generates copyrighted material that leads to a dispute?
• How do you notify customers when you update or change the underlying foundation model, and how do you ensure updates do not break our established workflows?

Almost every software vendor includes clauses that place the responsibility for verifying outputs entirely on the customer. While this is standard industry practice, acknowledging it upfront forces you to design internal review steps. If an AI tool drafts your customer proposals, you must ensure your staff understand that human review is a non-negotiable step before anything leaves the building.

5. How do we extract our data and workflows if we decide to leave?

Vendor lock-in happens quickly in AI adoption. Once your team spends six months building custom prompts, standard operating procedures, and automated workflows inside a specific tool, switching providers becomes costly and disruptive.

Ensure you have a clean exit route:

• If we cancel our subscription, in what format can we export our data, custom prompts, and workflow configurations?
• Can we extract our audit logs and user activity records for compliance purposes?
• How quickly is our data permanently deleted from your systems following termination, and will you supply written confirmation of destruction?

If a vendor relies on proprietary formats that prevent you from migrating your assets elsewhere, you are handing them significant pricing power when renewal time arrives.

Next steps: Test before you invest

Never sign an enterprise-wide, multi-year AI contract on the strength of a presentation. The most effective way to protect your business is to insist on a paid, time-limited pilot with a small, cross-functional group of 5 to 10 staff.

Define clear success metrics before the trial begins - such as reducing routine administrative time by two hours per person per week, or accelerating customer response times without sacrificing quality. Use the pilot to verify the vendor's claims about security, permission handling, and real-world costs. If the vendor refuses to support a pilot or pressures you into an immediate annual commitment, view that reluctance as a warning sign and look elsewhere.