Many artificial intelligence-enabled features inside enterprise resource planning, customer relationship management and sales platforms depend on third-party large language models that distributors neither selected nor control. Understanding those dependencies now can help prevent unexpected operational disruptions later.
Fill rates, inside sales productivity, pricing decisions, and customer service. That’s what you were managing last week. That’s the job.
But something happened recently that deserves your attention because it has direct implications for how your operation runs.
The U.S. government ordered a major AI model provider to restrict access to one of its models over national security concerns. The directive arrived late on a Friday. Because cloud providers cannot easily separate foreign nationals from domestic users, the company did the only thing it could do: It shut the model off for everyone. Live sessions ended in errors. New requests were quietly routed to an older, less capable system. Users were never asked. The provider made the decision and executed it automatically.
Here’s the part distributors should pay attention to: You don’t use that model directly. Almost nobody does.
But you use Copilot. Your customer relationship management platform added an AI assistant. Your quoting platform drafts proposals. Your software vendor launched a chatbot.
Many of those features rely on a large language model underneath, the same category of technology that was just shut off. If that model becomes unavailable, you’re not the one who chose to depend on it. Your vendor did. You may not realize there’s a problem until something stops working. And because the rest of the software may continue functioning normally, the root cause can be difficult to identify.
That’s the real issue: The dependency is hidden one layer down.
The AI tools spreading fast through distribution are not standalone AI subscriptions. They’re business applications with AI features embedded inside them. An enterprise resource planning module that drafts purchase orders. A customer service bot that answers questions at 2 a.m. A natural language search box on a dashboard.
In most cases, the AI is a feature, not the product itself. As a result, few distributors stop to ask what is powering it.
It’s also important to be precise about where the risk exists.
Not every AI feature in your software carries this exposure. Your demand forecasting system may rely on statistical models that have been in place for years. Your pricing engine may use machine learning that predates the current generative AI wave entirely. Those systems are unaffected by the type of disruption described above.
The exposure is concentrated in generative AI features, the tools that write, summarize, chat, and answer questions in plain language. Those capabilities typically depend on one of a small number of large language model providers.
And for those features, your vendor made the decision.
They may have built their application around a single provider. They may switch providers without informing customers. You don’t know unless you ask.
Distributors manage this type of risk every day in their physical supply chains. You know which products are sole sourced. You know which suppliers lack a viable backup. You would never allow a critical stock-keeping unit (SKU) to depend on a single supplier without understanding the risks.
Yet that’s exactly what many distributors are doing with AI-enabled software today because nobody asked the right questions when the software was implemented.
Here are the questions worth asking now.
Which features use a large language model, and whose model is it?
Separate generative AI features from everything else. The assistant and chatbot rely on a large language model. Forecasting and pricing algorithms do not. For every feature that depends on a large language model, determine whether the vendor relies on a single provider or multiple providers.
If the model is open source, which model is it? What is the country of origin of the model?
Open source is not a single thing. Open weights (Meta’s Llama, Mistral) give a vendor real optionality; open documentation without infrastructure access gives them a marketing bullet point. Also, country of origin matters: models from China-domiciled organizations carry different regulatory exposure than European or American alternatives.
If the model is their own trained model, is your data used to train it?
Ask this specifically: does my data improve your model, and does that improvement benefit other customers? Vendor contracts routinely permit training use under “service improvement” language broad enough to include your pricing logic, purchasing patterns, and margin structure. If the answer is yes, you are funding their product development with your operational intelligence—potentially including the model performance your competitors will use against you next year. Consider a data processing addendum that explicitly excludes training use, or price that contribution accordingly.
What happens if that model becomes unavailable?
Can the vendor switch to another model, or does the feature stop working entirely? Vendors that have thought through this scenario will have an answer. Vendors that have not are effectively transferring their architectural risk to you.
How quickly can you recover, and how will you know there’s a problem?
If the underlying model goes dark, are you facing an outage measured in hours, days, or weeks? Will the vendor proactively notify customers, or will your sales team discover the quoting assistant has stopped working in the middle of a shift?
What’s your exposure if you lose the tool for 60 days?
This question is for your team, not your vendor. For every AI-enabled application performing meaningful work, understand what breaks and identify the manual fallback process.
The good news is that most AI-enabled software is not yet mission critical. That gives distributors a window to ask these questions while the answers are still relatively inexpensive.
That window will not stay open forever.
As AI becomes more deeply embedded in distributor operations, these tools will move closer to the same level of importance as enterprise resource planning systems, procurement platforms, and key supplier relationships. Most distributors won’t think about these dependencies until a disruption occurs.
The ones that ask the questions now will be better prepared when someone else’s model suddenly goes dark.
Many AI-enabled features inside enterprise resource planning, customer relationship management and sales platforms depend on third-party large language models that distributors neither selected nor control. Understanding those dependencies now can help prevent unexpected operational disruptions later.
Share this article:
