AI and business data
Before you connect AI to business data.
An AI tool that can read your inbox, documents or customer records can save time, and it can also see more than it should. Answer these questions before you give any tool access, and start with the smallest access that still works.
Map the data
Know what the tool will see.
01
Which sources
List every source the tool would read or change: mailboxes, shared drives, the CRM, the online shop, accounting, chat history.
02
Which kind of data
Note whether the sources contain personal data, customer data, contracts, prices, passwords or other confidential information.
03
Who it belongs to
Some data belongs to your clients or partners and comes with contractual limits on who may process it.
If personal data is involved, the GDPR applies. Ask your legal adviser or data protection officer to review the use before the pilot starts, not after.
The tool and its settings
Read the terms of the plan you actually use.
- Check whether your inputs may be used to train or improve the provider’s models, and how to switch that off. Business plans and consumer plans often differ.
- Check how long inputs, outputs and logs are kept, and whether you can delete them.
- Check where the data is processed and stored, and whether the provider offers a data processing agreement.
- Prefer a company account managed by your business over personal accounts of individual employees.
Want help applying this? See our AI process automation service or tell us about your project
Access
Give the smallest access that still works.
01
Read before write
Start with read-only access. Allow the tool to send, change or delete only after the pilot shows it works as expected.
02
A separate account
Connect the tool through a dedicated service account or app, not through a person’s full account, so access can be limited and withdrawn without side effects.
03
Narrow scope
Limit access to the folders, labels or records the process needs, not the whole mailbox or drive.
04
Keys and tokens
Store API keys in a password manager or secret store, never in documents or chat messages, and know who can revoke them.
Specific risks
What is different about AI tools.
Instructions hidden in content. A tool that reads emails or web pages can be misled by text written to look like an instruction, for example a message asking it to forward data. This is known as prompt injection. Do not let a tool that reads outside content take actions without human approval.
Confident mistakes. Language models can produce plausible answers that are wrong. Keep a person in the loop wherever an error would reach a customer, change a record or cost money.
Data in outputs. Summaries and drafts can repeat confidential details. Decide who may see the results and where they are stored.
Checklist
Before you grant access.
- Is there a written list of the sources and the kinds of data the tool will see?
- Has someone responsible for data protection reviewed the use of personal data?
- Are model training, retention and storage location settings checked for your plan?
- Does the tool use a company account or a dedicated service account?
- Is access read-only and limited to what the process needs?
- Where is the human approval step before anything is sent or changed?
- Are actions and errors logged, and who reviews them?
- Who can switch the tool off and revoke its access, and how quickly?
If you have not chosen the process yet, start with Which business process should you automate first?
Planning an AI pilot?
Our AI process automation work starts with the data and access questions, before any tool is connected.