The model is rarely the risk
Dipal Panchal, co-founder and CEO, puts the risk in the data an AI tool can reach and the access around it, not in the model.
AI access. Data access review.
AI assistants and agents read data through accounts you already granted. List those accounts, and IRIS shows the sensitive data each one can reach.
Dipal Panchal, co-founder and CEO, puts the risk in the data an AI tool can reach and the access around it, not in the model.
A service account set up for a nightly report can be handed to a new assistant, and every grant it holds comes with it.
Dipal Panchal notes that many companies give enterprise AI licenses to every employee, who may then upload spreadsheets and PDFs holding company data without the company knowing.
Four AI accounts, one rejected and Copilot excluded. Last-used dates come from sources that record activity; on PostgreSQL, activity is tracked per database account unless audit logging is turned on.
| Identity | Role or group | Object | Privilege | Last used | Decision |
|---|---|---|---|---|---|
| svc_support_bot (support assistant) | support_readers | crm.person (email, mobile) | SELECT | November 7, 2026 | Narrow to a view without mobile numbers. Reviewed by the CRM owner |
| svc_rag_indexer (document search assistant) | Direct grant | cbs.customer_master (PAN, date of birth) | SELECT | October 30, 2026 | Revoke by November 22: it answers product questions and needs no customer IDs |
| svc_fraud_model (model training pipeline) | ml_feature_readers | cards.card_accounts (card number) | SELECT | November 5, 2026 | Keep: the protection check, activated on this install, found the sampled card numbers masked |
| svc_analytics_agent (warehouse agent) | analytics_writers | analytics.customer_snapshot (date of birth) | SELECT, INSERT | Per database account only: audit logging is off | Downgrade to SELECT: the agent has no reason to write |
| svc_etl_nightly | etl_writers | crm.person (email) | INSERT, UPDATE | November 8, 2026 | Not an AI account: it runs the nightly load, so the normal access review covers it |
| Microsoft Copilot | Each user’s own permissions | SharePoint and OneDrive files | Not an account IRIS maps | Not applicable | Excluded: it answers with each user’s access, so the review moves to the files’ sharing |
Before an AI assistant or agent goes live, Dipal Panchal’s first question is what data it can reach, at what permission, and whether it reads the values in full. The table above answers it one account at a time, with the reviewer’s decision beside each.
Our rule for a rollout: no assistant goes live on an account that can write, or that reaches restricted data stored in plain form. Each exception gets a named owner and a review date before the launch, written into the review sheet.
For each account you list, the sensitive tables and files it can reach, and whether it holds admin rights over them.
Whether the sensitive values it can reach are stored in plain form, from a check on sampled values that runs once it is activated on your install.
List the accounts your AI tools run as, then run one scan of the sources they read: the how-it-works page follows that scan, and the access review page shows the sheet your reviewers sign.
An account an AI tool runs as is mapped like any other account, from the permissions each source exposes.
Links each service account to the sensitive data it can reach, from table and view grants on databases and sharing on OneDrive and Google Drive.
Flags admin rights on sensitive data, which an agent shouldn’t hold unless administering that data is its job.
Checks sampled values for encryption, masking or hashing, once the check is activated on your install, and flags sensitive values stored in plain form.
Reviewers keep the access, or mark it for downgrade or revocation, with a reason; your team makes the change in the source, and a re-scan records the new grants.
Capabilities vary by edition.
IRIS doesn’t discover AI tools or agents, and keeps no inventory of models. It doesn’t read prompts or outputs, it can’t see what staff upload to an AI service, and it blocks nothing.
Nor does it connect to Microsoft Copilot (formerly Microsoft 365 Copilot). Copilot surfaces only the data each user already has permission to view, as Microsoft documents it, so its review is a review of who can open which files.
No. Your team lists the accounts its AI tools run as, and IRIS maps what each of those accounts can reach.
Partly. Copilot answers with each user’s own access, so the work is the sharing on your files. IRIS flags OneDrive and Google Drive files with sensitive data that anyone with the link can open.
No. IRIS reads the data sources you connect and the permissions on them, and has no view of prompts or uploads.
No. Duties on providers of high-risk AI systems, such as Article 10 of the EU AI Act on training data, aren’t mapped, and Article 10 applies from December 2, 2027 for Annex III systems. Cite each scan’s dated access record in your own assessment.
Bring the list of accounts your AI tools run as. We’ll show what each one can reach today.