The four-second leak: why AI governance is a dealmaking problem

These are the questions clients are now putting to their advisors about AI in dealmaking. They sound simple. For many firms, they are surprisingly hard to answer. We hear some version of them in almost every conversation we have with advisors.

Clients used to ask whether their advisors were using AI. Now they want proof it is being used safely. The firms that can answer all three questions clearly, and evidence those answers, will be the ones trusted with the next mandate.

The problem with shadow AI

Shadow AI is the use of AI tools outside a firm’s approved, secure systems, usually without IT, compliance or the client knowing. It might be a draft agreement dropped into a consumer chatbot for a quick summary, or a schedule run through a free translation tool before a bidder call. It is rarely a deliberate breach. It is capable people under time pressure reaching for the fastest tool available.

What makes it so dangerous is what cannot be seen. Once confidential information leaves the deal environment, there is no log and no audit trail. The firm cannot say where that information is now stored, who can access it, or whether it is being retained or used to train someone else’s model.

Because nothing visibly goes wrong in the moment, the risks of shadow AI are rarely reported.

Many firms have responded to shadow AI with policies such as not uploading client documents into public AI tools. This is a good first step, but it often relies on individuals choosing the slower option late at night when a tight deadline is approaching. Over enough deals, the deadline usually wins. That is not a failure of discipline. It is a design problem.

Make the secure option the fastest one

At Ansarada we have facilitated more than 60,000 critical transactions across 170 countries and are trusted by all of the Big Four, top legal firms, investment banks and 85 of the ASX 100. Across that work, what we hear from advisors is consistent. The controls a firm puts in place are rarely what worries them. The workarounds people reach for when those controls slow them down are.

The answer is not more rules. It is giving advisors AI that is fast enough to remove the temptation and built so the three questions are answered before anyone asks them.

That is the thinking behind AiDA, the Ansarada Intelligent Dynamic Assistant. AiDA runs inside the data room. Nothing leaves the controlled deal environment. Only the specific, permission-filtered content needed to answer a question is used. The advisor still gets the summary and the document stays where it belongs. The Ansarada platform carries ISO 27001 certification and SOC 2 attestation, and aligns with ISO 42001, the international standard for AI management systems.

It takes about four seconds to paste a clause into a public chatbot. Asking AiDA takes the same four seconds, and the document never leaves the room.

Where does the data live?

This question no longer stops at the source documents. It now extends to everything AI produces along the way, including OCR outputs and extracted entities.

For firms under the scope of the GDPR, Saudi Arabia’s Personal Data Protection Law, the UAE’s data protection regime or mandates that require data to remain in-country, this is where AI governance stops being a theoretical concern and becomes a commercial one. Data ending up in the wrong jurisdiction can breach the law or the tender conditions, with consequences ranging from fines to losing the next deal.

At Ansarada , AI content, including OCR outputs and extracted entities, stays in the storage region appointed for the data room. That region is chosen before the room is created and cannot be moved mid-transaction.

Who can see it?

AiDA works within the permissions already set in the secure data room environment. Answers are limited to each user’s room role, Q&A role and team permissions, so it cannot surface a document to someone who is not entitled to open it.

AiDA also answers only from the content in the data room. It does not search the internet or draw on outside information. That protects confidentiality and keeps hallucination risk low, which matters most when an answer could end up informing a disclosure.

It is read-only. It finds and explains, but it does not edit, move, share or delete. AI participation also remains the customer’s decision. Administrators can revoke AiDA access for a specific team, and account owners can request that it be disabled across the whole data room.

Does the AI model learn from it?

When it comes to what an AI model can learn, clients and regulators are becoming more wary of blanket reassurance. They want to know exactly what a model learns and from what. Advisors should be asking their technology providers the same question.

At Ansarada, we can answer it precisely. Customer document content is never used to train any model, and prompts and responses are not retained or logged once a request completes. Our contracts with AI providers include zero-data-retention commitments. Even the Bidder Engagement Score , our predictive model, was developed on anonymised historical behavioural data, such as how often bidders log in and how much of the room they have viewed. It never reads what the documents say, and no confidential or strategic information can be extracted from it. By day seven, it identifies which bidders are serious with 97% accuracy.

That is the level of specificity advisors can take to their own clients, and they should expect the same from every tool they use.

Defensible. By design.

AI governance is not a compliance layer bolted on after the fact. In dealmaking, it is a design decision. Shadow AI leaves no trace. Governed AI leaves an auditable one.

Firms that can say where the data lives, who can see it and whether the model learns from it, before the client asks, will not just avoid the awkward conversation. They are better placed to win the mandate.