Your Vendors Are Adding AI Features You Never Approved

Between January and March 2026, Salesforce, ServiceNow, Microsoft, and SAP all shipped major AI feature updates to their existing products. Enabled by default in most cases. Your procurement team approved these vendors years ago. Your security team reviewed them during renewal. But nobody reviewed the AI models now processing your customer data, employee records, and financial transactions. This is not the shadow AI problem. Shadow AI is employees going rogue with ChatGPT. This is worse. This is your approved, contracted, budget-line-item vendors quietly changing how your data gets processed, and nobody in your organization has a line of sight into it.

The Scope of the Problem Is Bigger Than You Think

I ran a quick inventory at a mid-size financial services firm last quarter. They had 47 SaaS vendors under active contract. 31 of those vendors had added AI features in the previous 12 months. Of those 31, the company had formally reviewed exactly 4. The other 27 were running AI capabilities that touched production data with zero governance review.

This is not an edge case. Gartner reported that 72% of enterprise SaaS vendors added generative AI features to existing products in 2025. Most of them shipped those features as opt-out, not opt-in. Meaning unless someone on your team actively disabled the feature, it was running.

Here is what that looks like in practice. Your HR platform starts using AI to screen resumes. Your CRM starts generating customer summaries from call transcripts. Your IT service desk starts auto-categorizing tickets using a large language model. Each one of these changes how your data flows, where it gets sent, and who (or what) processes it. Each one carries risk. And each one happened without a governance review.

The contract you signed two years ago gave the vendor permission to update their product. That clause was written for bug fixes and UI changes. It was not written for the vendor to start feeding your data into a model trained on their entire customer base.

Why Your Current Vendor Management Process Misses This

Traditional vendor management reviews happen at procurement and renewal. Maybe annually for critical vendors. The review covers uptime SLAs, data residency, SOC 2 compliance, and pricing. It does not cover what models the vendor is running, where inference happens, whether your data trains their models, or what happens to AI-generated outputs.

I spent 20 years in banking where we had model risk management frameworks that would make most companies cry. SR 11-7, the Federal Reserve's guidance on model risk, forced us to inventory every model, validate it independently, and monitor it continuously. That framework exists because regulators learned the hard way in 2008 that unmonitored models create systemic risk.

Most companies outside banking have nothing close to this. They have vendor management. They have security reviews. They have compliance checklists. None of those were designed to catch a vendor embedding a GPT-4 class model into your workflow overnight.

The gap sits between three teams. Procurement owns the contract but does not understand AI risk. Security owns the data flow but does not track feature releases. IT owns the platform administration but does not have a governance mandate. Nobody owns the question: did our vendor just change the computational logic processing our data?

The Vendor AI Feature Review Framework

Here is the framework I have been implementing with clients. It is not heavy. It adds roughly 2 hours per vendor per quarter for your top-tier vendors, and a simple questionnaire for the rest. The goal is not to block AI features. The goal is to know they exist and make a conscious decision about them.

Step one: build a vendor AI feature inventory. This is a spreadsheet, not a platform. Three columns that matter. Vendor name. AI features currently enabled. Date the feature was detected. Your IT admin team can populate this in a week by reviewing vendor release notes and checking admin consoles for AI toggles. At the firm I mentioned earlier, this took one analyst four days to complete for all 47 vendors.

Step two: classify each AI feature into three risk tiers. Tier 1 is AI features that process personally identifiable information, financial data, or health data. These get a full review before they stay enabled. Tier 2 is AI features that process business data like tickets, projects, or internal communications. These get a lightweight review within 30 days. Tier 3 is AI features that process no sensitive data, like UI suggestions or formatting tools. These get logged but do not require review.

Step three: for Tier 1 and Tier 2 features, send the vendor a standardized AI disclosure questionnaire. I use 12 questions. The five that matter most are: Does our data train your model or any shared model? Where does inference happen geographically? Can we opt out without losing core product functionality? What is your data retention policy for AI inputs and outputs? Do you use sub-processors for AI inference, and if so, who?

Step four: make the enable/disable decision at the right level. Tier 1 decisions go to your AI governance committee or equivalent leadership body. Tier 2 decisions go to the business owner of that vendor relationship plus your security team. Tier 3 decisions stay with the IT admin. This takes the bottleneck out. You are not routing every AI toggle through the C-suite. You are routing the ones that carry real risk.

How to Implement This Without Slowing Down the Business

The number one objection I hear is that this will slow adoption. That teams will lose access to features that make them more productive. Here is the reality. Of the 27 unreviewed AI features at that financial services firm, 19 passed review in under a week. 5 required vendor clarification that took another two weeks. 3 got disabled because the vendor confirmed that customer data was being used to train shared models. Those 3 were the entire point of the exercise.

Speed comes from having the process ready before the next feature ships. Set up a shared inbox or Slack channel where IT admins flag new AI features as they appear in vendor release notes. Most major SaaS vendors publish release notes monthly. Assign one person on your security or governance team to triage those flags weekly. That is 30 minutes a week in steady state.

Put a clause in your next vendor renewal that requires 30 days advance written notice before enabling any AI feature that processes customer or employee data. I have gotten this clause accepted by 8 of the last 10 vendors I have negotiated with. The two that refused were smaller vendors with less negotiating leverage on their side, and they eventually agreed to a 15-day notice window. This one clause changes the entire dynamic from reactive to proactive.

Finally, tie this to your existing change management process. If your organization tracks changes to production systems, vendor AI feature activations are production changes. They should show up in the same change log, get the same risk classification, and follow the same approval paths. You do not need a new process. You need to expand the definition of what counts as a change.

One more practical point. Document the decisions you make and why. When a regulator, auditor, or board member asks what AI is running in your environment, you want the answer to be a spreadsheet, not a shrug. I have seen audit findings issued specifically because a company could not answer the question: what AI models are processing our customer data? That finding cost one firm six figures in remediation work. The inventory I am describing would have prevented it entirely.

Actionable Takeaway