Copilot Without Guardrails: The New Document Risk

A junior lawyer is asked to draft a sales agreement for a new strategic client. They open a blank Word document and ask Copilot to ‘draft a Master Services Agreement for a UK customer, based on our recent deals’. Copilot obliges. It finds three different MSAs on SharePoint, blending a high-value software licence from 2019, a low-value services agreement from 2021, and an internal draft with placeholder text. The result is a confident, well-formatted, and completely incorrect document. The liability clauses are from the wrong jurisdiction and the payment terms are disadvantageous.

This is the new document risk. Generative AI inside Microsoft 365 does not access the wider internet for its answers; it draws from your own tenant. It learns from every Word document, PowerPoint presentation, and Teams message you have ever created. If your digital estate is a mess of outdated drafts, inconsistent branding, and rogue clauses, Copilot will dutifully learn from it. It will amplify this chaos into plausible, authoritative, and dangerously wrong outputs.

Garbage In, Polished Garbage Out

CIOs and their IT teams have spent years managing the *location* and *security* of documents, but not necessarily their *content*. The strategy was to provide a secure container—SharePoint, Teams—and leave content creation to the business. This approach was manageable, if inefficient, before AI.

Copilot for Microsoft 365 changes this dynamic entirely. It turns your entire corpus of unstructured data into a knowledge base. The AI does not distinguish between a final, approved contract and a draft from five years ago. It sees both as equally valid sources. When a user asks it to create a new document, it synthesises information from this mixed-quality pool.

The result is not just a productivity gain; it is the industrialisation of your company’s existing bad habits. Outdated fee schedules, retired brand assets, and non-compliant legal language are recycled into new documents at speed. The AI-generated output looks correct, which makes it far more dangerous than a visibly flawed, copy-and-pasted document from a human user.

The Hallucination You Are Liable For

Much has been written about AI ‘hallucinations’, where a model generates factually incorrect information. In the context of enterprise document creation, this risk is more subtle. Copilot is not inventing facts in a vacuum; it is reassembling your own information in novel, and often incorrect, ways.

Consider a quarterly financial disclosure created in PowerPoint. An analyst might ask Copilot to ‘summarise our Q3 performance based on the last board pack’. If the most recent file the AI finds is an incomplete draft, it may present draft numbers as final. It might pull a risk statement from last year’s report and place it in this year’s deck without context. The output is not a wild fabrication, but a misleading collage of your own data.

Critically, your organisation owns this output. When a flawed contract is sent to a client, or an inaccurate board pack is shared with directors, the defence that ‘the AI did it’ is not a defence. The failure is a lack of internal control over the tools used to produce legally and commercially significant documents.

Templates are the True Guardrail

Many organisations are reaching for familiar tools to mitigate this risk, such as Data Loss Prevention (DLP) policies or new content filters. These are inadequate. A DLP policy can stop a user from emailing a document containing a credit card number, but it cannot judge whether a liability clause in a contract is the correct one for a given deal size or jurisdiction. The risk is semantic, not a simple data leak.

The only effective guardrail is to control the source material. Rather than policing the output, you must govern the input. This means ensuring that when Copilot or a human user creates a high-stakes document, they start from a managed, approved, and properly structured template.

A governed template layer provides the structure that both humans and AI need to operate safely. It defines the safe playing field for document creation.

• Centrally managed templates ensure that every new MSA, SOW, or board pack starts from the correct, approved version.
• A shared clause library, integrated into Word and Outlook, provides pre-approved language for legal, commercial, and compliance-sensitive content that can be inserted into these templates.
• Protected sections and clear placeholders in the template prevent the AI or user from overwriting critical terms or legal boilerplate.
• Embedded metadata automatically classifies the document from inception, driving downstream workflows and ensuring proper record-keeping.
• Branding, disclosures, and signatory blocks are kept current by the central service, removing this burden from individual users.

The CIO’s Path to AI Governance

For the CIO, governing Copilot is not an application-level problem; it is a strategic infrastructure challenge. It requires a fundamental shift in how you view the corporate content estate. The era of treating SharePoint as a loose collection of user-managed folders is over. To prepare for AI, IT must lead the move to a structured, governed content layer.

This is not about blocking the tool. Restricting Copilot to a small group of ‘safe’ users neuters the productivity promise of the entire Microsoft 365 platform and creates a workforce of information have-nots. The goal must be to enable the entire organisation to use these powerful new tools safely.

The CIO’s role is to provide the tracks on which the AI engine can run. This means taking ownership of the template and content asset problem. It means procuring and implementing a platform, like Kameleon, that can manage high-value document templates, clauses, and brand assets centrally and deliver them into the Microsoft 365 applications where work happens.

Regulatory Scrutiny is Not Far Away

Regulators are beginning to address the use of AI in business processes. The EU AI Act, for instance, proposes obligations for organisations deploying ‘high-risk’ AI systems. While Copilot itself may not be designated high-risk, using it to generate regulated documents like contracts, financial reports, or employment offers could easily fall under such scrutiny.

Under these emerging regulations, organisations will need to demonstrate that they have appropriate human oversight and robust governance in place for AI-assisted workflows. A demonstrable system of governed templates and centrally managed content provides a clear, auditable answer to this requirement. It shows that you are not letting the AI operate unguided, but are proactively shaping its output by controlling its inputs.

Generative AI is a powerful engine with the potential to drive significant productivity gains. But an engine without a steering mechanism or a defined road will inevitably crash. It is IT’s responsibility to build the roads.