Skip to content
Guides

Shadow AI: roll out AI assistants safely at work

Shadow AI is staff using ChatGPT for work outside approved tools. Use SSO, an audit log and budgets, not a ban alone.

Last checked
19 November 2026
Reading time
15 min
Security rollout
Download as Markdown

A Sluis guide for security officers and CISOs. Last checked 19 November 2026.

Shadow AI is staff using ChatGPT for work outside approved tools. Use SSO, an audit log and budgets, not a ban alone. This guide sets out the five controls a CISO needs for a safe ChatGPT rollout, what the Dutch data protection authority has said about chatbot data breaches, how to keep a prompt log from becoming an employee monitoring problem, and a 90 day rollout plan with owners and evidence.

In short

  • Banning does not work on its own. People use chatbots because they save time. Provide an approved route that is easier, faster and already signed in, so staff have less reason to use a personal account.
  • Five controls carry the rollout: a policy on what may be entered, approved tools behind SSO, logging and audit, budgets and model allow-lists, and data protection before the prompt leaves your environment.
  • The Dutch Autoriteit Persoonsgegevens (AP) treats personal data pasted into a chatbot against the employer's rules as a data breach. You need an incident process that can decide on notification within the 72 hours of Article 33 GDPR.
  • A prompt log is itself processing of employees' personal data and can be employee monitoring. Log metadata and policy decisions by default, restrict who can read content, set a retention period and tell staff. In the Netherlands, involve the works council and check whether a DPIA is mandatory.
  • Under Article 4 of the AI Act, as amended in July 2026, you must take measures to support AI literacy of your staff. See our guide to the AI Act for organisations that use AI.

Why shadow AI happens and what the regulator says

Employees open a chatbot in a browser tab because it answers in seconds, while the official route is a ticket and a wait. Nobody intends a data breach. The risk comes from what goes into the prompt: a customer export, a draft letter about a named colleague, a block of code with a key in it.

On 6 August 2024 the AP published a warning based on breach notifications it had received. Its main points:

  • The AP had received several reports of data breaches because employees shared personal data of, for example, patients or customers with an AI chatbot. In one case an employee of a GP practice entered patients' medical data, against the agreed rules. In another, an employee of a telecom company entered a file containing customer addresses.
  • Employees often use chatbots on their own initiative and against agreements with the employer. If personal data was entered, that is a data breach.
  • Where chatbot use is part of the organisation's policy, it is not a data breach, but it is often not legally permitted. Organisations must prevent both situations.
  • Most companies behind chatbots store all data that is entered, often without the person typing it realising this.
  • The AP's advice: make clear agreements with staff on whether chatbots may be used and which data may and may not be entered, and consider arranging with the provider that it does not store the input. Where a breach does happen, notifying the AP and the people affected is in many cases mandatory.

The second point is what makes shadow AI a security issue and not only a policy issue. The point of an approved route is to bring processing under the organisation's control; approval itself is not a lawful basis. You still need a basis, a processor agreement under Article 28 GDPR and, often, a DPIA.

The controls, the risks they address and who owns them

ControlRisk it addressesWhat good looks likeWho owns itHow Sluis supports it
Policy on what may be enteredCustomer, patient, HR and secret data pasted into any modelOne page: approved tools, forbidden data categories by name, how to report a mistake without blameCISO with the privacy officer and legalOwners choose the models members may use; detected personal data is replaced before dispatch. Your organisation writes and approves the policy
Approved tools with SSOPersonal accounts, no offboarding, no visibilityEvery employee can sign in with the company identity on day one; leavers lose access automaticallyIT and identity teamSign in with Microsoft, Google, Apple or your own OpenID Connect provider (Keycloak, Okta, Entra, ADFS 2016 or later, authentik); SAML only through an OpenID Connect broker; SCIM 2.0 provisioning; owners can require two-factor and single sign-on
Logging and auditNo record of which route was used; no evidence after an incidentMetadata and policy decisions per request; content readable by few; retention set and documentedSecurity operations, with the privacy officerHash-chained audit trail verifiable offline, one audit row per request; audit log in the console
Budgets and model allow-listsRunaway spend; unvetted models and providersA monthly cap per person and per team; a short list of approved models with a defaultFinance with the platform teamMonthly euro budget per member and blocking budgets per workload; owners choose the full catalogue or a restricted set with a default
Data protection before the prompt leavesPersonal data and secrets reaching the model providerIdentifiers replaced before sending; residency enforced; limits of detection knownPrivacy officer with the platform teamDetected personal data and secrets replaced by reversible typed «TOKEN» placeholders and restored on the way back; default EU-only residency, other jurisdictions refused with 403 before anything is sent. Detection is not perfect and does not by itself establish GDPR compliance
Detect shadow use on the networkUnknown volume of unapproved toolsAggregated reports of traffic to AI services by department, reviewed monthlySecurity operationsDone with your existing proxy, DNS filtering, CASB or endpoint tooling. Sluis does not detect or block other AI sites
Incident processBreach not recognised, notified late or not at allA reporting channel, a triage owner, a 72 hour clock, a breach registerPrivacy officer or DPOAn audit row per request on the approved route. Retained content, if enabled, supports investigation; there is no Sluis record for shadow use
AI literacyStaff trusting wrong output; staff not knowing the rulesShort role based briefings and a record of who attendedHR or learning team, with the CISONone beyond the product itself

Sluis is a managed product from a young company, 7Lab B.V. in Amsterdam, with an ISO 27001 certification. It is one way to provide the approved route. The controls in the first column apply whatever tool you choose.

A rollout plan for 30, 60 and 90 days

Use three decision points: by day 30, policy and supplier assessment; by day 60, a pilot with tested controls; by day 90, organisation-wide access and a review. These are suggested planning targets, not legal deadlines. Do not start a personal-data pilot before the required assessment.

WeekActionOwnerEvidence
1 to 2Measure current shadow use from proxy, DNS or CASB data, aggregated by department and service, not by personSecurity operationsList of AI services in use and rough volumes
2 to 3Write the one page AI use policy: approved tools, forbidden data, reporting routeCISO, privacy officer, legalPolicy version 1, approved by management
3 to 4Select the approved route, sign the processor agreement, start the DPIA, ask the works council for consent on the logging rulesPrivacy officer or DPO, HRSigned agreement, DPIA draft, consent request
5 to 6Pilot with two or three teams: SSO, model allow-list, per member budget, data protection on, retention period setIT and platform teamConfiguration export, sample audit rows, pilot DPIA section
7 to 8Role based literacy briefings; tabletop test of the incident process with one chatbot breach scenarioHR or learning team, securityAttendance record, tabletop notes with timings
9 to 10Roll out to all staff through SCIM; announce that personal chatbot accounts are no longer allowed for work dataIT, communicationsProvisioning count, intranet notice, updated policy
11 to 13 (day 90)Review: shadow traffic trend, budget use, refused requests, incidents; update the DPIA; set the review cycleCISOReview report and next review date

The AP's 2026 tool asks for a periodic reassessment with named owners. Day 90 is the first one, not the last.

Audit logs that help and audit logs that create a new problem

Security teams want every prompt logged. That instinct can create the next problem. A log of what each named employee typed into an assistant is processing of employees' personal data, and it can be used to watch how they work.

Two Dutch rules apply directly.

  • DPIA. The AP's list of processing operations that require a DPIA (Staatscourant 2019, nr. 64418, of 27 November 2019) includes, at point 11 "Controle werknemers", large scale processing of personal data and/or systematic monitoring of employees' activities. Its examples include checking e-mail and internet use. A searchable log of every prompt per employee comes close to those examples, so assume a DPIA is required unless your analysis shows otherwise.
  • Works council. Article 27, paragraph 1, of the Wet op de ondernemingsraden requires the works council's consent for a decision to adopt, amend or withdraw (k) a regulation on the processing and protection of personal data of people working in the enterprise, and (l) a regulation on facilities that are aimed at or suitable for observing or checking their attendance, behaviour or performance. A prompt log is suitable for that, whatever its purpose.

What a defensible log looks like:

  1. Metadata and policy decisions by default. Who, when, which model, cost, and what the policy did: values tokenized, request blocked, budget reached. That answers most security and finance questions.
  2. Restrict who can read content. If you keep prompt content, limit reading to a small named group for incident investigation, and record when they do.
  3. Set retention. Pick a period you can justify in the DPIA. In Sluis, retained content expires after 30 days by default. You can store the tokens instead of the original values, or turn content retention off and keep only the metadata ledger, which is never altered.
  4. Tell staff. Say what is logged, who can see it, for how long and why, in the AI policy and in the privacy notice for employees.

Budgets and model allow-lists

A budget is the simplest control against both cost surprises and quiet misuse. In Sluis Workspace an owner sets a monthly euro budget per member; blank means no cap per member. Teams, agents and applications get a workload with its own allowed models, rate limits and a daily, monthly or total budget. When a blocking budget is reached, calls are refused with 402 before they reach a provider; in notify mode they continue. A model outside the workload's allow-list is refused with 403.

The model list is a security decision as much as a cost decision: each model is a provider, a jurisdiction and a set of terms. Start with a restricted set and a default, and add models on request. Our guide on controlling LLM costs covers budget sizing.

Data protection before the prompt leaves

The AP's 2026 tool opens with the advice to examine whether generative AI can be used without personal data at all. The GDPR points the same way: Article 25 names pseudonymisation as an example of data protection by design, and Article 32 lists pseudonymisation and encryption among the security measures to consider. A technical layer helps: Sluis replaces detected values with typed placeholders such as «EMAIL_1» or «IBAN_1», sends only the placeholders to the model, and restores the originals in the answer. It ships 61 built in detectors, 29 for personal data and 32 for secrets and credentials, plus custom terms.

This is not perfect. Pattern detectors miss names and descriptions such as "the colleague on long term sick leave". Treat placeholders as a reduction of exposure, not as permission to paste anything. Our guide on keeping personal data out of AI prompts shows measured detection rates and the legal effect of pseudonymisation.

What the AP expects before you roll out

On 13 July 2026 the AP published a practical tool, Hulpmiddel inzet generatieve AI & de AVG, for organisations that deploy generative AI, alongside a broader guidance document. On the same page it published its advice after a prior consultation by the municipality of Haarlemmermeer on Microsoft 365 Copilot, concluding that the intended processing would infringe the GDPR. The tool works through the principles of Article 5 GDPR. Its expectations, in short:

  • Purpose limitation. Define a specific purpose. "Supporting employees" or "making a process more efficient" is too general. Prevent use for other purposes.
  • Authorisation. If the assistant can reach existing files, access must follow employees' roles. Not everyone may see everything, including through the AI system.
  • Data minimisation. Limit the files and applications the system can reach, and set a policy on which categories of personal data users may enter.
  • Storage limitation. Know and control where inputs, outputs, metadata, diagnostic and telemetry data are stored and for how long. Keeping data because it is convenient is quickly unlawful, and a vendor's silence on retention does not lift your obligation.
  • DPIA, also for a pilot. A pilot can process personal data too, so a DPIA is needed. A generic reference DPIA from another organisation is not enough. High residual risk means prior consultation of the AP under Article 36 GDPR.
  • Wrong outputs. Make employees aware that answers can be wrong, and give them the knowledge and time to check them.

The tool itself states that it does not replace a DPIA. Our guide on the DPIA for ChatGPT and AI assistants for employees walks through one.

Worked examples

SituationWhat the policy saysControl that catches itResidual risk
A sales rep pastes a customer list with names, e-mail addresses and phone numbers into a personal chatbot account to write follow-up mailsCustomer data only in the approved tool; no bulk exports in any promptOn the shadow route only network tooling and training. On the approved route, e-mail addresses and phone numbers become placeholdersNames and free text context can slip through; if it happened on the shadow route, assess it as a breach under Article 33
An HR advisor drafts a warning letter that mentions an employee's absence and health situationHealth data and HR case files do not go into external modelsCustom terms for case numbers, block mode for detected values that must never leave, a restricted model listDescriptions of illness in plain words are hard to detect; keep this use out of scope in the DPIA unless strong measures are in place
A developer pastes a configuration file with a cloud access key to debug an errorSecrets never go into a prompt; use the approved coding routeSecret detectors (AWS keys, private keys, .env dumps, generic high entropy tokens) replace or block the valueA key already pasted into a shadow tool must be rotated; custom secret formats may need a custom term
A marketing team works in free personal chatbot accounts and uploads an event attendee listPersonal accounts are not allowed for work data; the team uses the approved WorkspaceSSO and SCIM so everyone has an account on day one; a per member budget; network reports measure remaining shadow trafficDo not assume data already entered under consumer terms has been deleted; ask for deletion where possible and record the decision

When a prompt becomes an incident

Give staff one reporting channel and ask them to report immediately, without copying the sensitive prompt into an open ticket. Record the service, account type, time, data categories and approximate number of people affected. Stop further sharing, restrict access, ask the provider for deletion where possible and rotate exposed secrets.

The privacy officer or DPO assesses the risk. Article 33 GDPR requires notification to the supervisory authority without undue delay and, where feasible, within 72 hours after awareness, unless the breach is unlikely to create a risk to people's rights and freedoms. Do not wait for a complete investigation before making that decision. Article 33(5) requires you to document the facts, effects and remedial action for any personal data breach. Record the decision not to notify too. Test this process with a synthetic customer list before rollout.

Checklist

  • One page AI use policy published, with forbidden data categories named
  • Approved route available to every employee through SSO, with SCIM provisioning
  • Model allow-list and default set; monthly budget per person and per team
  • Data protection on; detection limits documented
  • Logging rules agreed: metadata by default, content readers named, retention set
  • Works council consent requested under article 27 WOR where the log or rules fall under it
  • DPIA done, including for the pilot; processor agreement signed
  • Incident process tested, with a 72 hour notification decision and a breach register
  • Literacy briefings held and recorded
  • Shadow traffic measured with your network tooling before and after rollout
  • Periodic reassessment scheduled with a named owner

Questions

Can we just block ChatGPT on the network?

You can, with your proxy or DNS filter, and for unmanaged consumer accounts that is often sensible. On its own it moves the problem to personal phones. The AP's warning describes staff using chatbots on their own initiative. Offer the approved route first, then block the alternatives.

Do we need works council consent to log prompts?

In the Netherlands, very likely. Article 27(1)(k) WOR covers rules on processing employees' personal data, and (l) covers facilities suitable for checking behaviour or performance. A prompt log meets the second test even when monitoring is not the purpose.

Does Sluis stop shadow AI?

No. Sluis does not detect or block other AI services; that remains a job for your proxy, DNS, CASB or endpoint tooling. Sluis provides the approved route: SSO, model allow-lists, per member and workload budgets, data protection before dispatch and a hash-chained audit trail. Pricing is usage-based with no seat fees; see pricing.

Sources

All guides
Next guideDPIA for employee use of ChatGPT under the GDPR