How to choose an AI provider or workflow platform for a regulated firm
Eight things to check before a brokerage, agency or lender approves an AI tool or workflow platform: training, retention, security reports, data location, firm control, connections, audit logs and breach terms.
Before a regulated firm approves an AI provider or workflow platform, it should confirm in the provider’s terms and data processing agreement that its data isn’t used for training, how long data is kept, which independent security reports exist, where data is stored, that the firm controls models and access, that connections start read-only, that activity is logged, and how fast a breach is reported.
An AI policy names which tools are approved. This is how to decide which ones make the list. Check what each provider commits to in its terms, data processing agreement and security reports. Marketing pages describe intentions; contracts are what you can hold a provider to.
The eight checks
- Training. The terms or data processing agreement say your inputs and outputs aren’t used to train the provider’s models. Check for exceptions, such as feedback users choose to send. Anthropic, for example, says its commercial products don’t train on customer data by default, except feedback users send (Anthropic’s privacy centre).
- Retention. How long prompts, files and outputs are kept, and whether the firm can shorten it or turn retention off.
- Independent security reports. A SOC 2 Type II report or ISO 27001 certification the provider will share with you. These are audits by an outside firm, not the provider describing itself.
- Where data goes. Which country data is stored and processed in, and the list of other companies (sub-processors) the provider passes it to.
- Firm control. The firm decides which models are allowed, people sign in through accounts the firm controls, and access can be removed the day someone leaves.
- Connections. For workflows that connect AI to email, a CRM or a loan system: read-only by default, only the fields needed, and an approval for each action that changes a record or sends something.
- Audit log. Who ran what, what the AI accessed, and what was sent, kept long enough for your record-keeping rules.
- Breach terms. How quickly the provider must tell you about an incident, and what it will tell you.
Why workflows need more checks than chat
A chat tool sees only what a person types into it. A workflow connects AI to your systems and can run without someone watching each step. That makes the connection the main risk: what it can read, what it can change, and whether a person approves before anything reaches a client. What to check before connecting AI to loan data.
Questions to send a provider
- Is our data used to train any model, including for feedback? Where does your agreement say so?
- How long do you keep our data, and can we set that to zero?
- Can you share your latest SOC 2 Type II report or ISO 27001 certificate?
- Where is our data stored and processed, and who are your sub-processors?
- Can we restrict which models our people use, and remove access centrally?
- What does your audit log record, and for how long?
- How quickly will you notify us of a breach?
How Broker AI fits
Broker AI works with the AI provider the firm approves. It checks which AI tool is connecting on every request and refuses unapproved ones, records who accepted the firm’s AI policy, checks client-facing drafts against the firm’s rules, and keeps the record. It never stores client data. Add the checklist to your policy with the free AI use policy template.
General information, not legal or compliance advice. Rules differ by state, province and regulator; check with your compliance team or counsel.
Questions people ask
- Do AI providers train on business data?
- Many business and enterprise plans say in their terms that they don’t by default, often with exceptions such as feedback users choose to send. Consumer plans usually follow different terms. Check each provider’s agreement.
- What is a SOC 2 Type II report?
- A report by an independent auditor on whether a provider’s security controls worked over a period of time, usually six to twelve months. Ask for the latest one before approving a provider.
- Should an AI workflow have write access to our systems?
- Start read-only, with only the fields it needs. Add actions that change records or send messages one at a time, each behind an approval and logged.
- Is this legal advice?
- General information, not legal or compliance advice. Rules differ by state, province and regulator; check with your compliance team or counsel.
By DisruptionIO, the team behind Broker AI. Published October 28, 2026.