Can staff use AI to summarise contracts covered by a confidentiality agreement?

Staff can use AI to summarise a confidential contract only when the intended processing is permitted by the agreement and your organisation has approved the specific tool, account and handling arrangements. If either point is unclear, do not upload the contract. A subscription receipt is not permission from the party that disclosed the information.

Confidentiality is about more than whether a model trains on your document. An external service may receive the text, retain it, make it accessible to support staff or process it through other suppliers. Those activities can matter under your contract even when model training is disabled. Equally, not every external processor is automatically prohibited: some agreements permit defined service providers under suitable obligations. Read the wording rather than assuming either answer.

The UK Intellectual Property Office's NDA guidance gives a useful starting point. It says an NDA should specify a permitted purpose and warns against assuming conversations with advisers are automatically confidential. An AI supplier should not be treated as a solicitor or other professional adviser simply because it helps you understand a document.

Separate two decisions: are we allowed to process this information through this supplier, and can we trust the resulting summary? Passing the first test does not make the summary legally reliable. Passing the second does not repair an unauthorised disclosure.

Start with the agreement covering the information, not just the document you want summarised. A supplier contract may sit under an earlier NDA, tender conditions or a client's engagement terms. The confidentiality clause may also protect the existence of negotiations, prices and commercial schedules rather than only paragraphs marked confidential.

Identify the permitted purpose, permitted recipients and any restrictions on copying, subcontracting, overseas processing or particular technology. Check whether outside service providers are allowed, whether equivalent confidentiality obligations are required, and whether prior written consent is necessary. A clause allowing disclosure to employees does not necessarily extend to an external AI platform.

The Intellectual Property Office names accountants, banks and business coaches among advisers businesses may approach, but still tells businesses to ensure disclosures are made in confidence. Its guidance also notes that confidentiality periods of three or five years are common, while some information may need protection indefinitely. These are examples from the guidance, not standard terms that replace your signed agreement.

Consider a hypothetical UK engineering firm reviewing a confidential subcontract. Permission to share the document with its solicitor does not establish permission to upload it to an unrelated AI supplier. The business should have the relevant contract owner assess the recipient wording and obtain written permission if required. It should not ask the same unapproved chatbot to decide whether uploading the document was allowed.

For ambiguous wording or material commercial exposure, a solicitor's focused review is usually more useful than trying to infer permission from the AI supplier's marketing page. Record the conclusion and the limits of the approval so staff do not turn one exception into a general rule.

Approval needs to cover the actual service being used. Microsoft Copilot, ChatGPT and Gemini include different products and account arrangements. A tool's brand name alone does not tell you which contractual protections apply. Nor does a familiar office-software subscription automatically approve a new add-on, agent or browser extension.

Ask who receives the uploaded text, which subprocessors are involved, where processing occurs, and who can access inputs and outputs. Establish whether prompts or files can be used for training, how long content is retained, how deletion works and what security or abuse-monitoring exceptions apply. Check the contractual commitments alongside the available settings. Record the relevant service and terms rather than relying on a screenshot of a general privacy statement.

The National Cyber Security Centre's guidance on public language models recommends not putting sensitive information into public-model queries and stresses understanding the terms and privacy policy. Published in 2023, it is an enduring caution, not evidence of the current retention settings of every business product. Verify those separately before approval.

A self-hosted model can reduce external disclosure, but it is not an automatic exemption. Your hosting company, support arrangement, logging service or document-conversion tool may still receive information. Internal permissions and confidentiality restrictions still apply. The simpler alternative may be an authorised person reading the contract in an existing controlled document system.

Keep summaries in the same approved environment as the source. Otherwise the upload may be controlled while the output quietly appears in a widely shared chat or personal notes account.

Removing names helps only when it removes the protected information you do not need. It is not a reliable permission shortcut. A combination of unusual pricing, delivery dates, project descriptions and product specifications can identify a deal without naming either party. Commercial secrets can remain confidential even when no individual is identifiable.

If an approved task needs only one clause, use that clause rather than a complete contract and every attachment. Remove unnecessary signatures, contact details, bank information and unrelated schedules. Check the actual extracted text: comments, tracked changes, file properties and hidden spreadsheet cells may contain material that is not obvious on the first page. Use an approved local redaction process, not an unapproved online converter.

Personal information creates a separate data-protection assessment. Identify the purpose and appropriate lawful basis, assess necessity, address transparency and determine the supplier's role and required contractual arrangements. Consider international transfers and whether a data protection impact assessment is required. Confidentiality permission does not settle those questions, and UK GDPR compliance does not by itself authorise disclosure under an NDA.

The ICO's AI and data protection guidance covers accountability, lawfulness, transparency and accuracy. The ICO currently flags that this guidance is under review following the Data (Use and Access) Act, so check the latest guidance for the proposed use. A short document is not necessarily low risk: one paragraph can contain highly sensitive personal or commercial information.

Use a written approval boundary: which contract types, which service and account, which users, which purpose and which data categories. Name the person who owns that decision. Give staff an obvious route for exceptions instead of expecting each employee to interpret confidentiality law under time pressure.

Test with a fictional contract before introducing real confidential information. Ask for a factual extraction with clause references: parties, payment timetable, renewal dates, notice requirements and stated responsibilities. Require the system to say when a point is missing or uncertain. Do not let it silently convert a summary into a recommendation to sign, waive rights or accept a liability.

A reviewer must compare important points with the original contract, including definitions, exclusions, schedules and amendments. A plausible sentence can reverse the effect of an exception or miss a notice deadline. Clause references make checking easier, but the references themselves must also be verified. Keep the original authoritative and label the summary as an aid, not a replacement.

For example, an agreement with a £25,000 payment milestone could be summarised as payable on delivery when the wording actually requires written acceptance. This is a hypothetical illustration, not a reported case. The difference matters to cash flow even though the summary looks tidy.

Record the approval, document version, authorised recipient and reviewer without duplicating the confidential document in an unrestricted activity log. Review the workflow if the provider, terms or connected tools change. For wider supplier questions, our guide to red flags in an AI agency contract covers a different but related buying decision.

Do not use an external AI summary when the agreement prohibits the proposed disclosure, required consent has not been obtained or you cannot establish how the supplier handles the information. The same applies when you cannot restrict access or cannot provide a competent reviewer. An urgent deadline does not remove those conditions.

Use particular care with legal advice, litigation material, sensitive negotiations, employment disputes and transaction documents. Legal professional privilege is a separate issue from ordinary commercial confidentiality. Get advice from the responsible solicitor before processing privileged material through a new service. Do not assume that putting a confidentiality label on a chatbot conversation preserves privilege.

AI can be a poor investment for a handful of short contracts. An agreed manual checklist may cost less overall than supplier assessment, permissions, testing and review. For a repeatable volume of similar documents, a controlled summary service may be worthwhile, but measure the entire process rather than the seconds spent generating text.

As illustrative arithmetic, saving 20 minutes at an internal time value of £45 an hour gives £15 of potential capacity per contract. If checking takes another 15 minutes, that falls to £3.75 before software and governance costs. Those figures are hypothetical, not market rates, guaranteed savings or a reason to weaken confidentiality controls.

The practical answer is permission first, controlled processing second and human verification throughout. If you want help assessing whether a proposed contract-summary workflow fits your business, start with a non-confidential description of the task. You do not need to send the contracts to begin that conversation.