The AI Data Dilemma: Four Questions Every Business Leader Should Be Able to Answer
Updated: Sep 25
Most of what you read about AI and data is about companies that build AI models: where they sourced their training data, what they were licensed to use, and what happens in court when the answer is unclear. Very little of it is about companies like yours using data from various sources with you AI tools.
Here is why you should pay attention anyway. Every one of those disputes turns on a single question: can you show where the data came from, and what you were allowed to do with it? That question does not stop at the companies that build AI models. It gets asked of every company that uses one. And on the day somebody asks you, you will answer it out of whatever systems you happen to have.
Intellectual property attorney Kevin Thompson and Scalesology CEO Ken Scales cover this topic from both the legal and the operational side in a free 60-minute online webinar, The AI Data Dilemma: Balancing Innovation, Privacy, and Legal Risk
You can view the video of the webinar here:
How Does AI Actually Reach Your Data?
You are almost certainly not training a model. You are pointing models that already exist at data you already have. That happens through four paths, and most companies have all four running right now.
Tools your people chose. Employees adopt AI tools because they are trying to work faster, and they are right that it works. Customer data, financial reports, and contracts get pasted into those tools, usually against no written rule, because there is no written rule.
Features your vendors enabled. AI appears inside the CRM, ERP, HR, or support system you already license. Nobody at your company decided to turn it on. A vendor shipped a release, a terms of service was amended, or a toggle defaulted to on.
Retrieval over your documents. AI assistants that read SharePoint, shared drives, and knowledge bases inherit whatever permissions those stores already have. If your shared drive has been accumulating since 2014, the AI tool did not create that problem. It made the problem searchable in plain English by anyone who asks.
Integrations nobody owns. Connectors copy data continuously between applications, built at different times by different people for different reasons. This is the path that creates the most exposure and gets the least attention.
Not one of these four has a single owner in most mid-market companies. IT owns some of it, department heads own some of it, procurement signed some of it, and a fair amount of it nobody owns at all. That is not a discipline problem. It is a visibility problem, and visibility problems are solvable.
Where Does Consent Get Lost?
Consent is rarely lost at the moment you collect it. Most companies handle that part correctly. It gets lost downstream, and it happens in three places.
Vendor claims you cannot verify. A vendor tells you they hold rights to their training data and that your data is never used for training. Both statements may be true. The problem is that you have no practical way to check either one, and the assurance usually lives in a sales email rather than the contract.
Your own people, acting in good faith. Finance pastes an aging report into a chat tool. Sales uploads the customer list to clean up territories. Nobody acted maliciously and nobody broke a written rule. The employees doing this are usually your most motivated ones.
Movement nobody reviewed. You connect the CRM to the marketing platform, the ERP to the reporting tool, the ticket system to a support application. Those connections run continuously. Then a vendor adds an AI feature to one of those downstream systems, and customer data collected for one purpose sits inside a tool nobody evaluated for that purpose.
None of these requires bad intent. All of them require visibility the business does not currently have.
The Four Questions
Provenance sounds like a word for museums and lawsuits. For a mid-market business it comes down to four questions. Score yourself honestly.
Where does our sensitive data actually live? Systems, stores, data classes, and an accountable owner for each.
Which AI tools can reach it, and by what path? Users, permissions, embedded vendor features, and integrations.
What leaves, and where does it go? Destinations, subprocessors, retention, and evidence of deletion.
Who approved that access or movement, and when? A decision record, not a recollection.
In the assessments we run, most companies can answer one of these four. Occasionally two. The other three have to be reconstructed manually, every time someone asks, and they are out of date a month later.
That is the real problem. Not carelessness. The answers exist nowhere except in the memories of people who have day jobs.
A Policy Is Not a Control
Here is the distinction that decides whether your AI use policy protects your business or simply documents it.
A policy says do not put customer data in unapproved AI tools. It is circulated, acknowledged, and filed. It depends on every employee remembering it while under deadline pressure, and it is very difficult to verify after the fact.
A control means they cannot. Approved-tool access with a usable alternative. Rules that catch sensitive prompts and uploads. Permissions and integration allowlists scoped by purpose and owner. A review step when a vendor switches on a new AI feature. Logs, approvals, and exceptions captured as they happen.
If you have the first without the second, what you have produced is a document proving you knew about the risk. For every line in your AI use policy, name the control that makes it real. Where no control exists, either accept that the line is aspirational or do not write the line. Six enforceable rules are worth more than thirty that nobody can verify.
Where Should You Start?
The same four moves that govern data for companies building AI work just as well for a company using it. Apply them to the business you already run rather than to a training set.
Source. Know where your sensitive data originates and where it sits today. This is the step people skip because they assume they already know, and they almost never do.
Track. Follow it across every integration, connector, and AI feature that can reach it. This is the step that surprises people, because most of those paths were built by someone who has since left.
Verify. Check what a tool is allowed to touch before anyone switches it on. Vendor terms, feature intake, and configuration review. This is a standing habit, not a one-time clearance.
Deploy. Turn it on with controls in place, and keep the record of who approved what. Captured as you go, not rebuilt later.
Each move makes the next one possible. Skip Source and Track, and Verify is guesswork.
None of this is about slowing down. Most companies we talk to have an AI project stuck somewhere, and it is usually stuck because somebody sensible asked a question nobody could answer. The inventory is not what slows you down. Not having it is what has already slowed you down.
Frequently Asked Questions
Do we need to worry about AI training data if we are not building AI? Not in the way a model developer does. What matters for your business is the deployment side: what your tools can reach, what leaves your systems, and whether you can produce a record of it. Questions about legal exposure should go to your counsel.
Is a security audit the same as this exercise? No. A security review asks whether your data is protected from people who should not have it. This asks whether you know where it is, where it goes, and what is allowed to touch it. Related, but not the same, and the second one is usually missing.
Should we just block AI tools? Blocking without providing an approved alternative produces shadow usage on personal devices, which is worse because it is invisible. Give people a sanctioned path.
How long does this take? For a company in the $10 to $40 million range, an assessment of systems, data flows, and AI paths is measured in weeks, not quarters. The most common reason companies do not start is that they assume it is enormous.
Scalesology: Build the Foundation Before You Scale AI
AI is an amplifier. Point it at a governed data foundation and it accelerates good decisions. Point it at data nobody has mapped, with permissions nobody has reviewed, moving through connections nobody owns, and it amplifies that just as efficiently. The order matters. Foundation first, then AI.
At Scalesology, we help mid-market companies answer those four questions and keep them answered. That starts with a Business Systems, Data and AI Readiness Assessment covering your systems, data flows, automation opportunities, and AI readiness, and continues through application integration and information security work that puts real controls behind your policies. For companies that want the data movement itself governed in one place, Syncology provides a managed integration layer with AI access controls and audit logging, so the record is produced as data moves rather than reconstructed afterward.
Want to go deeper?
Join Kevin Thompson and Ken Scales on Wednesday, September 23 at 12:00 PM CDT for The AI Data Dilemma: Balancing Innovation, Privacy, and Legal Risk, a free 60-minute session covering the legal landscape and the operational playbook, with a roundtable and open Q&A at the end.
Ready to find out how many of the four you can answer? Contact us today, it is time to scale your business with the right data insights and technology.