Glossary. AI governance.
What is AI governance, and where does data come in?
AI governance is how an organization decides which AI it uses and on what data. This entry maps the EU AI Act, NIST AI RMF, ISO/IEC 42001 and GDPR to data.
By Dipal Panchal, co-founder and CEO · Updated November 8, 2026
Which AI, for what, on which data
AI governance is the set of policies, roles and records an organization uses to decide which AI systems it uses, for what, and on which data. Unlike AI security, which defends AI systems against misuse and attack, AI governance sets the rules and names who is accountable for them. This entry focuses on the data part, because an AI tool reaches data through the same accounts and shares people already have.
The NIST AI Risk Management Framework and ISO/IEC 42001 are voluntary. The EU AI Act is law in the European Union, with duties that depend on the risk level of each system, and GDPR applies whenever personal data is involved.
Where the frameworks touch data
The provisions that ask something about the data an AI system uses or reaches. Abridged; the linked text governs.
| Framework | Provision | What it asks about data |
|---|---|---|
| EU AI Act | EU AI Act Article 10, from December 2, 2027 (Annex III) and August 2, 2028 (Annex I) under Regulation (EU) 2026/1744 | Governed training, validation and testing data: relevant, representative, as error-free and complete as possible |
| EU AI Act | EU AI Act Article 26(4), (6), applying from the same dates | Deployers of high-risk systems keep relevant input data, where they control it, and the logs the system generates |
| EU AI Act | EU AI Act Article 4, as replaced by Regulation (EU) 2026/1744, Art. 1(5), in force from July 27, 2026 | Providers and deployers take measures to support AI literacy among staff and others using AI systems on their behalf |
| NIST AI RMF 1.0 | GOVERN 1.6 | Mechanisms to inventory AI systems, resourced by organizational risk priorities |
| NIST AI RMF 1.0 | MAP 4.1; MEASURE 2.10 | Documented approaches map risks of AI components, including third-party data or software; privacy risk is examined |
| ISO/IEC 42001:2023 | Annex A, controls A.7.2 to A.7.6 | Data for AI systems: acquisition, quality, provenance and preparation, defined and documented |
| GDPR and UK GDPR | GDPR Article 35 | A data protection impact assessment before processing likely to result in a high risk, including new technologies |
The data part, in four questions
The frameworks above reduce to four questions for each AI tool. Which accounts does it use: each user’s own access, or a service account of its own? What sensitive data can those accounts reach, through direct grants, group membership and sharing links? How is that data stored, in plain form or protected? And who decided the access was right, and why?
The first three are facts you can check in the systems. The fourth is the record a regulator or auditor asks for, and it means nothing without the first three.
Five AI accounts and what they reach
Accounts in one company, reviewed on November 7, 2026. Rows 2 to 4 are service accounts that AI tools run as.
| Identity | Role or group | Object | Privilege | Last used | Decision |
|---|---|---|---|---|---|
| Assistant users (each user’s own access) | Members of the HR site | HR site > Investigations library | Edit, inherited from the site | Not measured | Narrow the site membership before the assistant is switched on |
| svc_rag_indexer | kb_reader | Policies library and Finance library | Read | November 6, 2026 | Remove the Finance library; the index answers policy questions only |
| support-agent-bot | crm_ro | crm.contact and crm.payment_cards | SELECT | November 7, 2026 | Drop the payment table; the bot needs contact and ticket data |
| ml_training_job | warehouse_ml | analytics.customers_full | SELECT | October 2, 2026 | Replace with a de-identified extract before the next training run |
| AI feature inside the CRM | Managed by the vendor | Every CRM record, inside the application | Read within the app | Not measured | Excluded from the grant review; assessed under the vendor contract |
Govern the reach before the model
“Where we need to pay more attention as an organization” is the question of “what kind of data are we feeding to it,” said Dipal Panchal, Sylox’s co-founder and CEO, in a May 2026 interview. Before any AI tool goes live, his first check is what data it can reach and with what permission: full values, or protected ones.
Our rule: no AI tool goes live until each of its accounts has an owner, a list of the sensitive data it can reach and a decision on that reach. Rows 2 to 4 each fail one of those today, which is why each carries a change.
What an access map doesn’t answer
An access map doesn’t show what an assistant was asked or what it answered, which needs the tool’s own logs. It doesn’t judge whether a model is accurate, fair or fit for its use, which is the risk review the frameworks describe.
And it can’t see inside a vendor’s feature, like row 5, where the vendor’s application decides what its AI reads. That access is governed by the contract and the vendor’s own assessment.
IRIS maps who can reach sensitive data in supported sources, and the accounts AI tools use are identities in that map like any other. Capabilities vary by edition.
The sensitive data an AI tool’s account can reach: IRIS produces part of this, for supported sources that expose permission data.
How that data is stored: IRIS produces part of this, checked on sampled values, once the protection check is activated on your install.
An inventory of AI tools, prompts or outputs (your web proxy or CASB logs show which tools staff use): Not produced by IRIS.
Mapping findings to AI regulations: Not produced by IRIS.
Explore AI governanceFill in the four questions for one tool
Pick the AI tool closest to going live and write the four answers down: its accounts, the sensitive data each can reach, how that data is stored, and who decided. The check passes when every account has all four, and none reaches data its use doesn’t need.
Tools nobody approved are shadow AI, and they belong in the same register. Before an assistant reads your files, work through the checklist for AI assistant oversharing.
Official sources
- NIST AI Risk Management Framework (AI RMF 1.0): GOVERN 1.6, MAP 4.1, MEASURE 2.10, NIST. Opened November 8, 2026.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1(5) and the amended Article 113, European Parliament and Council. Opened November 8, 2026.
- ISO/IEC 42001:2023, Annex A, controls A.7.2 to A.7.6 (data for AI systems), ISO. Opened November 7, 2026.
From AI tools to the data they reach
What is data access governance?
The step under AI governance: every path from an account, an AI tool’s included, to sensitive data.
What is shadow AI?
The AI use nobody approved, which governance has to find before it can rule on it.
What is oversharing?
The next step before an assistant goes live: files that reach more people, and so more prompts, than they should.
What teams ask about AI governance
Yes, in some cases. EU AI Act Article 2 covers providers placing AI systems on the EU market, and providers and deployers outside the EU where the system’s output is used in the EU.
A named executive for the policy, with risk, legal and compliance for approvals. The data part needs the data owners and security, because they hold the accounts and the grants.
It is the data part of AI governance: the accounts, reach and storage covered above. Treat it as one program with one register.
Protect your brand. Start with your data.
Bring the accounts your AI tools use. We’ll show what sensitive data they can reach, in supported sources.