Picture three ordinary afternoons in a regulated firm. An AML analyst, mid way through an enhanced due diligence review, wants a client's source of funds documents summarised but the firm blocked ChatGPT on the corporate network months ago, so she simply photographs the documents on her screen and uploads them to ChatGPT on her personal phone. A developer at a crypto exchange drops a batch of flagged wallet addresses and transaction data into his own Claude account to debug a screening script. A player protection officer at an online casino feeds a VIP's account history and affordability data into Gemini to draft an intervention note. None of it is malicious, every one of them is trying to do their job as efficient as possible. But in each case regulated customer data has just left the building: to a personal account, with no record, no processing agreement behind it, and on some tiers a provider that may keep it to train its models. And nobody in compliance knows it happened.
Some regulated firms have embraced AI; others have tried to shut it down at the firewall. Either way, blocking at the network hasn't stopped it. It has simply pushed usage onto personal accounts, personal devices and home networks, where a corporate filter never reaches. Netskope's 2026 research found that nearly half of all generative AI users reach the tools through personal accounts that sit entirely outside enterprise controls, so "we've blocked ChatGPT" rarely means staff have stopped using it. More often it just means the remaining usage is the least visible kind. And a blanket ban is doubly self-defeating: the shadow usage continues regardless, while the real productivity gains; faster drafting, document review, analysis and coding are handed to competitors who worked out how to adopt AI safely. The firms that win are not the ones that say no, but the ones that can say yes, safely.
These scenarios aren't anomalies they're the everyday reality of how generative AI entered the enterprise. It didn't arrive through procurement; it arrived through a browser tab, faster than governance could keep up. That gap between adoption and control is where regulatory, data-protection and operational risk now lives, and most firms know it: Mimecast's 2026 State of Human Risk report found that 80% of organisations worry about data leaking through generative AI, yet 60% still have no specific strategy to address it and only 40% feel fully prepared.
The prize is real
It's worth being clear about what sits on the other side of the ledger, because it's the reason banning AI is the wrong instinct. AI is no longer a novelty in regulated industries — it is becoming core infrastructure. In iGaming, more than 80% of businesses now use generative AI, according to the 2026 State of AI in Gaming report from the UNLV International Gaming Institute and KPMG, and the most significant gains are proving to be operational rather than player-facing: demand forecasting, fraud detection and compliance itself, where high volume, error prone work like regulatory reporting, audit trail management, monitoring and data retention across jurisdictions is exactly what AI handles well. The same pattern holds across financial services and crypto: faster onboarding and KYC, sharper transaction monitoring, and skilled people freed from routine drafting to focus on genuine risk. Used well, AI doesn't just reduce cost, it prevents it, catching problems earlier and protecting expert time for the work that needs it. A firm that bans AI outright doesn't sidestep that race; it forfeits it. The goal is never to stop AI, but to make it safe to say yes.
The trouble is that the controls are lagging the adoption. In that same study, governance was the lowest-scoring dimension of AI maturity — 30 out of 100 — with only around one in five companies operating a dedicated AI governance role and most having no established practices at all. The appetite is there; the guardrails are not.
Why this is different when you're regulated
In a regulated environment the bar isn't "avoid a breach" — it's to demonstrate control and to evidence it.
Confidentiality. Customer PII, KYC and source of funds evidence, transaction data and intellectual property are exactly what gets pasted into public models and exactly what your GDPR and regulatory obligations exist to protect.
An obligation you may already be breaching. If you touch the EU market, the EU AI Act's AI-literacy duty has applied since February 2025, and the high-risk obligations; data governance, logging, human oversight — bite from August 2026. Non-compliance with those high risk obligations carries fines of up to €15M or 3% of global turnover; the headline €35M or 7% tier is reserved for the most serious breaches, such as the outright prohibited practices under Article 5. Affordability checks, player profiling, fraud detection and AML scoring can all fall in scope. You cannot evidence governance over usage you cannot see.
Operational resilience. Under DORA and NIS2, an unmanaged AI tool is an unmanaged ICT third party with access to your data and regulators are moving to treat the major AI and cloud providers as critical third parties, putting concentration and dependency squarely in scope.
Personal accountability. Under SMCR and the expectations of the FCA, GFSC, JFSC, MGA and their peers, named individuals answer for controls. Delegating a decision to an algorithm does not dilute that liability.
The goal isn't to stop AI. It's to make usage visible, governable and evidenced — without grinding teams to a halt.
The data-protection trap
A common assumption is: "We have a data processing agreement with the AI provider, so we're covered." It deserves a second look.
Take Anthropic. It does offer a DPA, but only within its commercial terms, covering the contracted enterprise products (Claude for Work and the API), with Standard Contractual Clauses and a UK transfer addendum. That is exactly what you want, for the instance it covers. The trap is twofold.
First, the DPA covers only the enterprise tenant you contracted. The moment an employee uses a personal account, free or paid Pro, on consumer terms with no Article 28 relationship, and on some tiers the provider's right to retain or train on inputs, you have permitted an uncontrolled transfer of personal, possibly special-category, data you cannot see or retrieve. Second, even where the DPA applies, it is a paper control: it constrains what the processor may do, but does nothing to stop your own people over-sharing categories you never meant to send. Contractual cover is not data minimisation, and a regulator examining a breach will ask about both.
Which is why being able to see and scan what is actually uploaded is the control that makes the paperwork real.
What effective control actually looks like
Visibility is where most firms fall short: Microsoft's 2026 threat intelligence found that 73% of organisations had detected unauthorised AI use on their networks, but only 28% had any comprehensive way to monitor or control it. Detecting the problem is not the same as governing it. Practical AI control tackles that in order — see it, control access, then protect the data:
Visibility. Discover every AI tool in use, distinguish personal accounts from licensed enterprise instances, and show where your data is actually going.
Account-level access control. Rather than blocking a tool outright, permit your contracted enterprise tenant while blocking personal accounts — free and paid — so usage flows through the instance your DPA actually covers. Enforcement should run on the device itself, so it follows employees onto any network and covers both browsers and native desktop apps — not just managed machines sitting behind the office gateway.
On-device data-loss prevention. Inspect content inline and stop sensitive data — PII, PCI, PHI, IP — before it leaves the endpoint, without routing your most sensitive material through a third-party cloud just to inspect it.
Together these turn "we have a DPA" from a hopeful statement into a demonstrable control: enterprise-only enforcement brings usage under the contract, and inline scanning enforces minimisation regardless.
But technology is only half the answer. A control is only defensible if it sits on a written AI acceptable use policy; one that defines which tools and accounts are approved, which data must never leave the building, who is accountable for AI use under SMCR, and how staff are trained to meet their AI-literacy obligations, with clear procedures for review and for handling an incident when something does slip through. Policy without enforcement is a document nobody follows; enforcement without policy is a control you cannot explain to a regulator. Regulated firms need both, working as one.
"But don't we already have Microsoft Purview?"
If you run Microsoft 365, possibly — and Purview governs Microsoft's own AI, Microsoft 365 Copilot, well. The danger is that it quietly narrows your AI strategy to a single lane: govern Microsoft AI, block everything else. Yet the tools your people increasingly want — Claude for Work, Claude Cowork, Gemini Enterprise, ChatGPT Enterprise — sit outside that estate and are often the better choice for the job. Purview's reach into them is gated behind E5 licensing, leans on browser extensions that never see native desktop apps (Claude Cowork and Claude Code among them), and can't permit your enterprise tenant while blocking personal logins. That leaves a false choice between Microsoft-only AI and ungoverned shadow AI. A vendor-neutral, device-level control removes it — letting you safely adopt the best enterprise AI, on any network, whoever makes it.
Where Endida comes in
This is the gap Endida closes. As a pure-play, AI-enabled cyber, fraud and compliance partner built for regulated industries — iGaming, crypto, fintech and financial services — Endida has the experience, the capability and the solutions to bring shadow AI under control and wrap it in the policy, procedures, training and evidencing regulators expect: drafting and enforcing the AI acceptable-use policy, discovering and controlling AI usage, enforcing enterprise-only access, preventing sensitive data from leaving your endpoints, and turning all of it into a defensible, auditable posture aligned to your obligations — FCA, GFSC, JFSC, MGA, VARA, the MiCA framework, the EU AI Act — and mapped to a recognised standard such as ISO/IEC 42001 where you need to demonstrate maturity to regulators and counterparties.
And the problem is growing, not shrinking. Today it is an employee pasting data into a chat window; next it is autonomous AI agents reading, retrieving and transmitting data with far less human oversight — something both the ICO and the FCA are already examining. The control you build now is the foundation that keeps that next wave governable.
The firms that handle this well move from "we think staff are using AI" to "we can see exactly who is using what, we govern it, and we can prove it" — without slowing anyone down. If you want to understand where shadow AI sits in your organisation today, and bring it under control without killing productivity, speak to Endida: https://endida.com/endida-ai-governance.html
Photo by Igor Omilaev on Unsplash