Picture a mid-sized firm in the final round with its largest prospective client. Procurement sends over a security questionnaire, and this year it has a new section: list every third party and AI system that will process our data. The COO starts working down the list. The CRM added an AI assistant in a product update last spring. The meeting platform began generating AI summaries by default. The support desk now routes tickets through a chatbot built on another company’s model. Nobody approved any of it as “AI.” It arrived as a feature, and now nobody can say where the data goes.

This is not a hypothetical problem. In April 2026, Vercel, a cloud platform used by thousands of development teams, disclosed a breach that started at a third-party AI tool. A Vercel employee had connected Context.ai, an AI agent tool, to the company’s Google Workspace account with broad permissions. Attackers who had compromised Context.ai used that connection to reach Vercel’s internal systems and a limited subset of customer credentials. Vercel’s customers never chose Context.ai, and most had likely never heard of it. The exposure came through an AI tool their vendor’s staff was using.

It is also not a one-off. In August 2025, attackers used stolen credentials from Drift, an AI chat agent that Salesloft sold and many companies had connected to their Salesforce CRM, to pull data out of those customers’ Salesforce environments. Google’s threat intelligence team said more than 700 organizations were potentially affected. In both cases, the businesses that were hurt did nothing wrong inside their own walls.

The numbers point the same direction. The Verizon 2026 Data Breach Investigations Report found that breaches involving a third party grew 60% in a single year and now account for 48% of all confirmed breaches. Two editions earlier, that figure was 15%.

That changes the AI data privacy conversation for leadership. The question is no longer only “What are our employees uploading to AI?” It is “Where is AI already processing our data, across every tool and vendor we rely on?” If your largest client asked tomorrow which of your vendors use AI on their data, could you answer from documents you actually hold?

Why This Can’t Wait: What It Costs When You Find Out Late

The gap between what a business thinks is happening to its data and what is actually happening widens with every vendor product update. Finding that gap late is far more expensive than looking for it now.

Start with breach cost. IBM’s 2026 Cost of a Data Breach Report puts the average U.S. breach at $11.5 million, more than double the global average, and supply chain compromise remains one of the costliest factors. The same report found that more than 20% of organizations had experienced a breach targeting an AI model or application. AI is no longer an emerging category of risk. It is already showing up in breach data.

The cost often arrives before any breach. Enterprise buyers and government primes now ask about AI and subprocessors in their security questionnaires. A company that cannot answer slows its own sales cycle or loses the contract outright. Insurers are moving the same way, adding AI questions to renewal applications and introducing AI exclusion language in several policy lines. A business that cannot describe how its vendors use AI may learn at renewal that it is paying more for less coverage.

Most importantly, responsibility does not transfer. When a vendor or its AI provider is breached, your company is still the one notifying clients, answering regulators, and filing the insurance claim. Every month without visibility is another month of vendor updates the business has not reviewed.

Deliberate AI Sharing vs. Passive AI Processing

Most AI programs are built to address only one of two very different kinds of AI data privacy risk.

Deliberate AI Sharing

Deliberate sharing happens when employees knowingly use AI. They upload contracts into AI tools, enter client details into chatbots, use consumer AI accounts for company work, or adopt unapproved apps, a problem often called shadow AI. This is a real risk, but it is manageable with familiar controls: an AI acceptable use policy, training, approved tools, data loss prevention, site blocking, and oversight.

Passive AI Processing

Passive processing is AI embedded in software the business already approved. Think meeting transcription and summaries, CRM conversation and sentiment analysis, inbox summaries, “ask an agent” chatbots, support case summarization, and enterprise search. Employees may never intentionally send data to AI. It simply becomes part of the normal workflow, which is what makes it so hard to see and control.

A real example shows how this plays out. In August 2025, a federal lawsuit seeking class-action status accused Otter.ai of having its AI notetaker join and record work meetings without the consent of every participant, and of using those conversations to improve its AI. A similar suit has since been filed against another AI notetaker, Granola. The business lesson is direct: when a vendor’s notetaker joins a call with your client, your client’s confidential pricing or legal discussion is now inside a third party’s system, and your client never agreed to that.

Internal AI policies govern employees. They do not govern vendors. A company can have a strong AI policy and still have its most sensitive client conversations processed by AI it never evaluated.

Your Software Stack Is Becoming an AI Supply Chain

The vendor relationship used to be simple: your company and a software provider. Today, the chain often runs from your company to the SaaS provider, then to an AI model provider, then to a cloud provider, and then to additional subprocessors. Each layer is a place where data can be retained, logged, reused, or exposed, and your business has a contract with only the first one.

Vendors add AI in three main ways, and each carries a different kind of risk. Some build it in-house. That is usually the most transparent option, but it can still be a “black box” to the customer. Others buy it from another AI company, plugging that company’s model into their product. Your data then flows to a fourth party you never vetted, possibly outside the United States. A third group assembles products quickly with AI coding tools, sometimes called “vibe coding,” and may not have the discipline to secure and operate a cloud product well. This risk is not limited to startups. Any vendor rushing AI features to market can cut corners.

Many vendors also do not clearly disclose their subprocessors, so it is often unclear whose hands your data passes through or where those parties are based. That is not a technical detail. It determines what your business can truthfully tell its own clients.

Leaders should be able to answer a few questions for every critical vendor. Where does our data travel, and which third parties and AI models touch it? Is it retained or logged, for how long, and where is it processed? Is it used for model training or “product improvement”? Does our original risk assessment still describe today’s product? A practical first step is to pull the terms of service and data processing agreement for your five most critical vendors and check whether subprocessors, including AI providers, are disclosed. If they are not, ask for the list in writing. This is the core of AI third party risk management, and it applies to every vendor, not just the newest ones.

It’s Not Just the Tools You Use: How Your Vendors Use AI

Vendor AI risk is not limited to features your team can switch on or off. Vendors are also using AI tools inside their own operations to handle your data. Their support teams summarize your tickets with AI assistants. Their account managers run AI notetakers on calls with you. Their developers use AI coding tools on systems that hold your records, and their analysts may drop customer data into AI tools to build reports. Your business may never touch any of these tools and still carry the consequences.

The Vercel incident is the clearest example. Vercel’s customers did not use Context.ai. A Vercel employee did, and that single connection gave attackers a path toward customer credentials.

This kind of use should be covered in writing, not left to assumption. The places to look are the vendor’s master agreement, data processing agreement, privacy policy, and terms of service. Each should state whether the vendor’s staff may use AI tools with customer data, which AI providers are involved, and whether that data can be retained or used for training. Even if your business never uses the tool itself, the vendor should be having this conversation with you. A vendor that cannot explain how its own teams use AI with client data probably has not thought about it.

When you raise the topic, ask whether the vendor’s employees or contractors use AI tools with your data and which ones, whether that use is governed by an internal AI policy they will summarize for you, whether the AI providers behind those tools are listed as subprocessors, whether their privacy policy or terms allow your data to be entered into AI tools or used for training, and whether they will notify you before any of that changes. For a complete review process, see our guide to AI vendor due diligence.

A vendor’s AI use is part of the vendor relationship, whether it shows up in the product or only in the vendor’s back office, and the agreement should say so.

The Contract You Signed May Not Describe Today’s Product

Vendors change after the signature. They add AI features and new AI providers, update privacy policies, change retention practices, expand integrations, add model-training provisions, and turn new AI features on by default. The contract and risk review your business completed at purchase can quietly stop describing reality.

Recent history shows how this happens. In 2023, Zoom updated its terms in a way widely read as allowing customer content to train AI, then revised them after public backlash. In 2024, Slack drew criticism when customers learned their data could be used for machine learning unless an administrator emailed to opt out. In 2025, WeTransfer reversed a terms change after users objected to language about using files to improve machine learning. In each case, the change was reversed only because the public noticed. Most terms updates are never noticed at all.

Leadership should ask a direct question: do we understand the risk profile of our vendors as they are today, not as they were the day we signed? If the answer is uncertain, that is the signal to conduct a risk assessment.

What Control Do You Actually Have Once Data Is With a Vendor?

A mid-sized business will not rewrite a large vendor’s standard terms. But it has more control than most leaders assume, and that control sits in a few specific places. For every critical vendor, review whether you have a model-training opt-out confirmed in the contract rather than a help-center article, opt-out rights for new AI features and default settings, data deletion rights including at contract end, clear retention terms, and defined breach notification timing and contacts.

Depending on your industry, regulations may require more than this baseline, including data residency, subprocessor notice, audit rights, and data export. Healthcare, for example, carries strict limits on sending U.S. patients’ health data outside the country, and government contractors handling controlled unclassified information face similar constraints. Leaders need to know which rules apply to their own industry.

When a vendor’s AI practices become unacceptable, the options run roughly in order of effort. You can turn off or restrict the AI feature in the admin console, limit which data flows into the tool, negotiate a contract amendment or move to a plan tier with stronger terms, add compensating controls, and, as a last resort, plan an exit. That last option is hard. Critical vendors are embedded in business processes, hold years of historical data, connect to other systems, and require retraining to replace. Large migrations disrupt the business. That switching cost is exactly why the review needs to happen before a crisis, while you still have negotiating room and time to plan.

A Framework for Evaluating AI Vendor Risk

eSilo uses its AI Decision Framework to evaluate any vendor that uses AI. It looks at four areas.

Data Sensitivity: What Data Am I Sharing?

This is the most important question. The more sensitive the data, the higher the bar for any AI tool that processes it. Paying for an AI subscription does not make a tool safe for every kind of data, especially confidential or regulated information. Look at what each vendor can reach: customer data, financial data, legal documents, employee records, regulated information such as PII, PHI, and CUI, intellectual property, and strategic plans.

AI Exposure: Where Does the Data Go?

Built-in AI is the most common and usually the least transparent. Leaders should know whether a vendor uses third-party models or open and self-hosted ones, which AI subprocessors are involved, how long data is retained, and whether the vendor’s own staff uses AI tools with your data even when the product itself has no AI features. Even when a vendor promises your data will not train its model, you cannot assume your interactions are not being used for product improvement.

Organizational Control: Who Controls the Model?

Consider the technical controls and settings your business can manage itself, the contractual protections available, including opt-outs, deletion, and audit rights, and whether the vendor must notify you of changes and incidents.

Vendor Dependency: What Happens When Something Goes Wrong?

Weigh the impact of a breach or outage, the cost and complexity of moving your data elsewhere, the business interruption a switch would cause, and whether real alternatives exist. The highest-risk profile combines sensitive data, limited control or transparency, high AI exposure, and high switching costs. Any vendor that matches all four belongs at the top of your review list.

Where to Start: A 30-Day Plan for Leadership

In week one, list the vendors that touch your most sensitive data. Start with the 10 to 15 platforms that hold client, financial, or employee information, and note which ones have added AI features since you signed.

In week two, check the admin settings. Look for AI features that were switched on by default, such as meeting summaries, notetakers, and CRM assistants, and decide whether each should stay on and for which data.

In week three, pull the paper. Gather terms of service, data processing agreements, privacy policies, and subprocessor lists, and ask each critical vendor in writing how its own staff uses AI with your data. Flag any vendor that will not disclose its AI providers or reserves the right to use your data for training or product improvement.

In week four, assign ownership and decide. Name an owner for each critical vendor, rank vendors using the framework above, and decide which need contract changes, configuration changes, or a deeper assessment.

Most companies stall in week three. The answers require reading contracts and technical documentation with a critical eye, and that work rarely belongs to any one person today.

Questions Leadership Should Be Asking

  • Which of our critical vendors have added AI capabilities since our last review?
  • Which vendors use third-party AI, and can we identify their AI subprocessors?
  • Do our vendors’ agreements, privacy policies, and terms of service address how their own employees use AI with our data?
  • Do our contracts address AI, model training, and opt-out rights?
  • Are vendors required to notify us when AI-related processing changes, and what happens if we reject new AI terms?
  • Which vendors would be hardest to replace, and what compensating controls exist in the meantime?
  • Who owns AI vendor risk inside our company, and how often are vendors reassessed?
  • Could we answer a client’s third-party AI questionnaire today from documents we hold?

How eSilo Helps

eSilo gives leadership visibility into vendor AI risk and a prioritized plan to address it, so your business can keep using the tools it depends on without inheriting risks it never agreed to. eSilo does not resell vendor products, so its recommendations are not shaped by a sales relationship. Every eSilo advisor brings 15+ years of experience in large enterprise or federal government environments, where they designed, implemented, and operated the vendor risk controls that regulators, insurers, and enterprise clients now expect from mid-sized companies.

eSilo helps leadership:

  • Identify where AI is operating across the technology stack, including features vendors switched on by default
  • Pinpoint where sensitive and client data meets vendor AI and fourth-party providers
  • Confirm how vendors’ own teams use AI with company data and whether agreements cover it
  • Review contract and data terms and identify what to request at renewal
  • Prioritize vendor risk by business impact and build a vendor risk process the business can maintain
  • Prepare documentation to answer client questionnaires, insurer questions, and audits

eSilo has helped businesses protect their data since 2002, with an average client relationship of more than 10 years, Best of Florida recognition from 2022 through 2025, and WOSB, MBE, and WBE certifications. This work draws on eSilo’s AI strategy and governance and cybersecurity compliance consulting services.

Frequently Asked Questions

If we pay for an enterprise or business plan, does that mean our vendor won’t use our data for AI?

Not necessarily. Paid tiers often carry stronger data terms than free versions, but “no training” commitments vary, and many vendors still reserve the right to use interactions for product improvement. The reliable answer is in the contract and data processing agreement, not on a marketing page. Confirm it in writing for every vendor that touches client data.

How do we find out which of our vendors are already using AI on our data?

Start with vendors that hold client, financial, or employee data. Check admin consoles for AI features enabled by default, review release notes and recent terms updates, and request subprocessor lists that include AI model providers. This is rarely one person’s job today, which is why it gets skipped.

If our vendor’s AI provider has a breach, who is responsible to our clients?

Responsibility does not transfer just because a vendor held the data. Your business generally remains the one notifying clients, answering regulators, and filing the insurance claim, as many Salesloft Drift customers learned. Review notification, cooperation, and liability terms with legal counsel before renewal.

Can we turn off AI features in software we already use?

Often yes, at least partly, through admin settings, and it is one of the fastest wins available. But some AI processing happens on the vendor’s side and cannot be switched off, and some settings change with product updates. Document which features are on, which are off, and who reviews them after each major update.

What is fourth-party risk, and why should a CFO care about it?

Fourth-party risk is the risk that comes from your vendor’s vendors, such as the AI company your CRM provider uses. Your business has no contract with that party but still carries the consequences if it fails, and with nearly half of breaches now involving a third party, that exposure is financial.

What should we ask for when a vendor contract comes up for renewal?

Ask for a model-training opt-out in writing, retention limits, deletion at contract end, breach notification timing, advance notice of new AI subprocessors, and the right to opt out of new AI features. Renewal is when you have the most leverage.

If we never use a vendor’s AI features, are we still exposed?

Yes, potentially. Vendors increasingly use AI tools in their own operations, such as support, account management, and development, and those tools can touch your data whether or not you turn on an AI feature. The Vercel incident is a clear example. Ask each critical vendor how its staff uses AI with your data, and confirm the agreement, privacy policy, and terms of service address it.

How quickly can we get a clear picture of our vendor AI exposure?

Most businesses can identify their highest-risk vendors within a few weeks with a structured approach. eSilo’s Risk and Resilience Diagnostic is a three-week engagement that produces a prioritized roadmap.

Conclusion: Your Biggest AI Exposure May Sit With Your Vendors

Internal AI policies do not govern what vendors do with your data, and the largest AI data privacy exposure may sit with your vendors rather than your staff. Every product update a vendor ships without review widens the gap between what your business believes is happening to its data and what is actually happening.

Three questions sum it up. Do we know which vendors are processing our data with AI, in their products and in their own operations? Do we understand our rights if their practices change? Which vendors are we too dependent on to replace easily? Strong AI third party risk management starts with being able to answer all three.

If those answers are not clear today, contact eSilo to start a Risk and Resilience Diagnostic. This structured three-week engagement maps where AI and third-party vendors touch your company’s data, identifies the highest-risk gaps, and delivers a prioritized roadmap. Get the answers before your next client questionnaire, insurance renewal, or vendor contract renewal, not after.