Your Accounting Software's AI Feature Was Added Last Quarter. Nobody Audited It.
Shadow AI Proliferation in Legacy Accounting Platforms
---
Last quarter, your accounting software updated. Somewhere in that update was an AI feature, enabled by default. Your clients' financial data started flowing through it. No procurement ticket was raised. No GDPR review triggered. No data processing agreement was updated. The software was already approved, so nobody looked at what the update added.
AI features that arrive via software update don't knock. They let themselves in.
The Governance Gate That Wasn't There
Every organization that manages AI tools responsibly has a process: new tool request, legal review, GDPR Article 28 assessment — which governs how companies choose and oversee third-party processors of personal data — DPA update, IT security sign-off. The process works when someone is making a deliberate decision to adopt a new tool.
Software vendors found the side door. When Sage adds an AI invoice categorization feature to an existing Sage 50 update, no IT ticket opens. No procurement review triggers. The software is already approved, already running, already processing client financial data. The AI feature is an enhancement to the approved product — and enhancements, by standard IT governance logic, aren't new products.
You approved the software. The vendor added the AI. Nobody approved the AI — because it arrived inside something you'd already approved.
Major accounting software platforms added AI features to their existing products in 2023-2024: Microsoft embedded Copilot capabilities into Office 365 and Dynamics, available to any organization with an existing Microsoft 365 subscription. Sage added AI-powered invoice matching and automated categorization to Sage 50 and Sage Business Cloud. Cegid integrated AI features for automated journal entries. QuickBooks added AI-assisted categorization and smart reporting. Each was positioned as an enhancement. Each arrived via standard software update. None required a separate procurement decision.
The Numbers That Define the Exposure
89% of enterprise AI usage generates no logs, no SSO connection, and no oversight visibility, according to LayerX research from 2025. That figure isn't describing employees using unauthorized tools on personal devices. It's describing what happens when AI functions activate within approved enterprise software: the processing occurs, but no governance-accessible record is created in the firm's systems.
97% of organizations that experienced AI-related security incidents had zero access controls governing AI usage at the time of the incident, according to IBM and Kiteworks research from 2025. The connection: default-enabled AI features don't get access controls, because nobody made a deliberate decision to deploy them.
Connect the sequence: Cegid releases a quarterly update — AI invoice categorization is enabled by default — client invoices process through it — the AI routes queries to Cegid's cloud infrastructure — no new data processing agreement was signed for this — the firm's DPO has no record of the processing activity — the CNIL inquires about AI processing of client financial data. At the inquiry, the firm's IT team can show the software was approved. They cannot show the AI feature was.
What GDPR Actually Requires
GDPR Article 28 requires that when a data controller — the accounting firm — adds a sub-processor to its data processing chain, the processor — the accounting software vendor — must notify the controller first. The controller must then evaluate the new sub-processor and update data processing agreements accordingly. This obligation exists regardless of how the sub-processor was added: deliberate choice or embedded software update.
A software update that activates a new AI feature routing financial data through a new cloud provider is a material change in processing. Under GDPR precedent established post-Schrems II — the 2020 Court of Justice ruling that invalidated Privacy Shield over surveillance law incompatibility — any material change in data processing requires review of transfer safeguards, sub-processor chains, and jurisdictional exposure.
Most accounting software release notes don't include GDPR impact statements for AI features added. They don't disclose which AI providers are used. They don't flag which features involve data leaving the EU. Firms relying on release notes to identify governance-relevant changes are working with incomplete information by design.
The Personal Liability Element
Partners at accounting firms carry professional liability for data protection failures in client engagements. GDPR enforcement can target individuals — not just organizations — in cases of deliberate or negligent non-compliance. If a data protection authority investigation finds that a firm processed client financial data through an AI feature for 18 months without covering it in a data processing agreement, the "we didn't know it was running" argument is not a defense. It is an admission that governance controls were insufficient.
The GDPR violation from an unauthorized AI feature isn't triggered when the feature is discovered — it is triggered when the feature first processed personal data. Retroactive governance doesn't fix it. The processing that already occurred cannot be undone. The only question at that point is whether the firm can demonstrate they had adequate controls and were misled by the vendor, or whether the controls simply weren't in place.
NIS2 — the EU directive requiring organizations to manage cybersecurity risks across their supply chains, effective since 2024 — further requires organizations to assess security risks from software updates — including material changes in what existing software does. Under NIS2, a software update that adds AI data processing is a supply chain change event, not a routine patch. Most accounting firms haven't adjusted their patch management processes to classify AI feature additions separately.
The Symmetry Test
Symmetry test: if an employee installed a new AI tool on their work laptop and used it to process client financial data, the firm would treat it as a serious shadow AI incident — policy violation, potential GDPR breach, disciplinary action, client notification assessment. When the vendor installs an equivalent AI feature via software update, the same data flows through AI the same way, but nobody opens an incident ticket. The data doesn't know how the AI got there.
Vendors gain AI adoption metrics that justify their investment and satisfy investor expectations without the friction of opt-in requirements. A default-enabled feature reaches 100% of the installed base immediately. An opt-in feature reaches whoever clicks through a notification and evaluates it. Commercial incentives and governance requirements run in opposite directions on this question, and commercial incentives are winning.
Deployment of AI features via default-enabled software updates is a governance bypass mechanism. Not a malicious one — vendors believe they're delivering value. But the consequence is identical to employees using unauthorized tools: financial data enters an AI system without prior assessment, without DPA coverage, without the firm having exercised its role as data controller.
What the Audit Looks Like
Firms that haven't reviewed their accounting software for AI features enabled in the last 18 months are almost certainly processing client financial data through systems they haven't audited. The audit is straightforward and takes less time than most compliance tasks: review release notes for every major accounting platform in use — search the notes for "AI," "intelligent," "smart," "automated," "machine learning," and "Copilot" — then check for each AI feature whether a GDPR Article 28 assessment was completed and whether the current data processing agreement covers it.
DORA — the EU Digital Operational Resilience Act, which took full effect in January 2025 for financial entities — requires firms to maintain registers of all ICT tools including material updates to existing tools that change data processing scope. An AI feature added to accounting software via update is a material change under DORA's framework. Firms subject to DORA that haven't maintained this register for the past year have a documentation gap that DORA auditors will find.
Stralevo doesn't update into something different without disclosure. Every AI processing capability in Stralevo was designed into the product from the start — documented before any client data was processed, covered in data processing agreements from the initial deployment. When Stralevo adds new capabilities, clients know what changes and why, because the governance model requires it. AI that was built as AI from day one operates differently from AI that arrived via a Tuesday update.
Before the Next Quarterly Update Cycle
Leading accounting firms have begun formalizing a quarterly AI feature audit: reviewing release notes from every approved software vendor for AI additions, checking DPA coverage for anything new found, updating sub-processor lists. The process takes 30 minutes per platform per quarter. The alternative — discovering a gap during a DPA investigation — takes considerably longer and carries consequences the quarterly audit doesn't.
Right now, your accounting software may be running AI features that weren't active six months ago. Your clients' financial data is going through them. Your data processing agreements may not mention them. The question isn't whether your software vendor followed their own update procedures correctly — they almost certainly did. The question is whether your governance framework treats "software already approved" as a permanent license for whatever the software becomes.
Ahead of the next update cycle, document which AI features are currently active in each accounting platform your firm operates. Check whether each is covered in your current data processing agreements. If any aren't, the conversation with your vendor and your DPO is a straightforward one. Having it now, proactively, is categorically different from having it after a data protection authority inquiry. The firms that formalized this process in 2025 are the ones who won't be explaining 2024's update cycle to regulators in 2027.