Most companies didn’t decide to adopt AI. Their employees decided for them. Now leadership is catching up, and the gap between AI use and AI control has never been wider.
IBM’s 2026 Cost of a Data Breach Report found that 68% of breached organizations had no AI governance in place, up from 63% a year earlier. Among organizations that suffered an AI-related breach, 92% lacked proper AI access controls (Cycode on IBM’s report). Only about a third have strict approval processes for deploying AI tools.
Regulators are not waiting either. The EU AI Act’s transparency rules have applied since 2 August 2026. A short delay, the “Digital Omnibus”, moved the heavier high-risk obligations to 2 December 2027, but legal experts are clear that the requirements were postponed, not softened. Companies using AI in hiring, credit or education have about 13 months to get their house in order.
Here is how the EU AI Act timeline looks after the Digital Omnibus:
| Date | What applies |
|---|---|
| February 2025 | Banned AI practices, such as manipulative or social-scoring systems |
| August 2025 | Rules for general-purpose AI model providers |
| August 2026 | Transparency rules, such as telling people when they are dealing with AI |
| December 2027 | High-risk AI obligations, covering areas like hiring, credit and education |
| August 2028 | High-risk AI built into regulated products |
Even companies outside the EU feel this. Any business that offers AI-powered services to people in the EU falls within the Act’s reach, much as GDPR did.
Here are the five controls that separate companies using AI safely from companies hoping for the best, and a quick scorecard to see where yours stands.
Control 1: One front door with single sign-on
The problem. Microsoft Entra data shows 67% of employees reach generative AI tools through personal accounts. Every personal login is an account your company can’t see, can’t audit and can’t switch off when someone leaves.
What good looks like. All AI access runs through your existing identity provider. New hires get access on day one; leavers lose it the moment HR closes their account. IBM’s research consistently ranks identity and access management among the most effective ways to cut breach costs, yet 92% of organizations hit by an AI breach lacked proper AI access controls.
Ask yourself: if an employee left today, how many AI tools would still have their company data?
Control 2: Department-level policies, not one rule for everyone
The problem. A single company-wide rule is either too loose for legal and finance or too strict for engineering and marketing. So people work around it.
What good looks like. Each department gets the AI it needs, and nothing more. For example:
| Department | Models allowed | Tools and connectors | Monthly budget |
|---|---|---|---|
| Engineering | Coding-focused models from two providers | Code repository and docs search | Higher, usage-based |
| Marketing | Writing and image models | Brand asset library | Fixed |
| Legal and finance | One approved provider in an EU region | None by default | Low, reviewed monthly |
This is where platforms differ most. Enterprise plans such as ChatGPT Enterprise offer role-based access and workspace-level policies, while self-hosted workspaces such as Grengin let admins choose which AI providers and models each department can use and set department budgets. Whatever you evaluate, test whether policies can be set per team, not just company-wide.
Control 3: PII detection before the prompt leaves
The problem. Once a prompt is sent, you can’t take it back. And sensitive data goes into prompts constantly:
- 40% of files uploaded to generative AI tools contain personal data or payment card numbers.
- Cyberhaven found 39.7% of all data moving into AI tools is sensitive, and the average employee sends sensitive data to AI about once every three days.
- In BlackFog’s survey, 27% of employees using unapproved AI had entered employee data such as salaries, and 23% had shared financial statements or sales data.
- Netskope counts an average of 223 generative AI data-policy violations per organization every month.
None of this needs a bad actor. Picture a support agent pasting a customer complaint into a chatbot to draft a polite reply, card number and home address included. It takes four seconds and creates a reportable incident.
What good looks like. The AI platform scans every prompt and file before it goes to the model, and responds in proportion:
| What it finds | Example | Sensible response |
|---|---|---|
| Payment card or bank details | A card number in a complaint | Block and explain why |
| Government ID numbers | A passport or national ID number | Block |
| Contact details | Names, emails, phone numbers | Warn, or mask before sending |
| Credentials | API keys, passwords | Block and alert security |
The goal isn’t to stop people using AI. It’s to catch the four-second mistake before it becomes a four-month investigation. Look for detection built into the AI tool itself rather than a rule people have to remember: Microsoft 365 Copilot can apply existing Purview data-loss prevention policies, and some self-hosted workspaces, Grengin among them, include PII detection in the chat.
Control 4: An audit trail for every AI interaction
The problem. When something goes wrong, the first question is “what happened?” Without logs, nobody can answer. IBM found the average breach now takes 247 days to identify and contain, and that not knowing which applications hold your data adds about $201,000 to the cost of a breach. Not complying with regulations adds about the same again.
What good looks like. Every AI interaction is recorded: who asked, when, which model answered, and what files or tools were involved. Logs are tamper-evident, so no one can quietly edit them, and exportable, so they can flow into the security tools your team already uses.
This also matters for the EU AI Act. Companies deploying high-risk AI systems will need to keep the logs those systems generate, generally for at least six months, once the high-risk rules apply in December 2027. Building the habit now is far cheaper than retrofitting it later.
Control 5: Budgets and usage analytics
The problem. AI spending is easy to waste and hard to see. Independent benchmarks show the same model can cost nearly eight times more at its maximum effort setting for the same result, and a single runaway session can burn millions of tokens. Spread across hundreds of employees, small leaks become large bills.
What good looks like. Each department has a budget, and admins can see usage by user, team and model. That shows which teams get real value, which settings waste money, and where to invest next.
Several platforms now cover both controls. Microsoft 365 Copilot includes audit logging at no extra cost, Claude Enterprise exposes audit data through a Compliance API, and self-hosted workspaces such as Grengin keep exportable, tamper-evident logs alongside department budgets and usage analytics in your own cloud.
The 60-second AI governance scorecard
Give your company one point for each “yes”.
- Does every employee reach AI through company single sign-on, with no personal accounts?
- Can different departments have different models, tools and limits?
- Is sensitive data like card numbers and IDs caught before a prompt is sent?
- Could you export a full log of last month’s AI activity within an hour?
- Do you know what each department spent on AI last month, and on which models?
5 points: governed. You’re ahead of most companies. Keep reviewing quarterly.
3 to 4 points: exposed in places. Fix the gaps before December 2027 brings the EU’s high-risk rules into force.
0 to 2 points: running on trust. You are in the majority, and in IBM’s data that majority is the one paying for breaches.
A 90-day plan to get from 0 to 5
You don’t need a year-long programme. Most companies can put all five controls in place in a quarter if they go in the right order.
Days 1 to 30: close the front door. Pick one approved AI tool, connect it to your single sign-on and publish a one-page usage policy. Announce it with a simple message: here is the approved tool, here is what never goes into a prompt. Start blocking personal-account access only after the approved option is live.
Days 31 to 60: set the guardrails. Turn on PII detection, starting in warning mode so you learn what people actually paste before you start blocking. Agree department-level rules with team leads, beginning with the most sensitive teams: legal, finance and HR.
Days 61 to 90: prove it works. Connect audit logs to your security tools, set department budgets and run your first monthly usage review. Then re-take the scorecard above, and share the result with leadership.
5 mistakes that quietly undo AI governance
- Buying the tool without the controls. Many AI products keep single sign-on, audit logs or data residency for their top tier. Check which plan includes what before you sign.
- Logging everything, reviewing nothing. An audit trail nobody reads only proves the breach happened. Assign an owner and a monthly review.
- Blocking so hard people give up. PII rules that stop legitimate work send employees straight back to their personal accounts. Tune the rules using warning mode first.
- Writing the policy, skipping the training. In a May 2026 Gallup poll, only 25% of US employees said their organization had shared a clear AI plan. A policy people haven’t heard about doesn’t exist.
- Treating governance as a one-off project. New models and tools arrive every few weeks. Put the scorecard on the calendar every quarter.
How the main options handle the five controls
No single product wins on every control, and the best fit depends on what you already run. Based on each vendor’s published documentation:
- ChatGPT Enterprise: single sign-on with SAML and automatic user provisioning, role-based access, workspace policies and audit logs. A natural fit for companies standardising on OpenAI.
- Microsoft 365 Copilot: reuses Purview sensitivity labels, data-loss prevention and audit logging, at $30 per user per month. A natural fit for companies already deep in Microsoft 365.
- Claude Enterprise: customer-managed encryption keys and a Compliance API for audit access. A natural fit for teams standardising on Anthropic’s models.
- Self-hosted open source (Grengin, LibreChat and others): runs in your own cloud and connects to several AI providers. Grengin, for instance, lists single sign-on, per-department model access and budgets, PII detection and tamper-evident audit logs. A natural fit for teams that want to own their data and can run the software themselves.
Two questions decide how far you can trust any of these controls: where they run, and whether you can verify them. SaaS plans run the controls in the vendor’s cloud and ask you to trust its documentation. Self-hosted, open-source tools run inside your own boundary and let your security team read the code, at the cost of operating it yourselves. Score each option on the five controls above before you sign anything.
Don’t let the deadline find you first
AI rules, risks and tools are moving faster than most governance teams can track. We turn the latest research, regulations and real incidents into short, practical guides you can act on the same week.
Follow us so the next guide reaches you well before the next deadline does.
What was your score out of five? Share it in the comments, along with the control your team finds hardest to put in place. We’ll answer the most common challenges in an upcoming post.
If your CISO, IT lead or compliance officer hasn’t seen this scorecard yet, forward it to them. It takes 60 seconds to read and might start the most useful meeting of their quarter
