Here they come. You’ve seen them and heard them. Yesterday’s digital transformation experts are today’s AI-powered software sales reps.
I should know. I’m one of them.
Twenty years ago, we sold document management, digital workflow, ERP, MPS, and whatever else fit under the heading transformation. Grand promises: increased efficiency, shorter processes, lower costs, and more revenue.
Sometimes it worked, and sometimes the customer spent 18 months digitizing the same lousy process they had been using on paper. Often, the project simply faded into the background.
Now AI is getting the same treatment.
AI-powered CRM. AI-driven analytics. AI-enabled service desks. A chatbot in the corner; suddenly the old product has learned to think.
Yeah, sure. Right.
But here’s the promise and the irony. AI actually delivers.
It’s happening right now. Sales reps use an LLM to research an account; the good ones rehearse the next meeting before lunch. A service manager drops three years of tickets into a model and finds a pattern nobody noticed.
Ten times yesterday’s output? Yes. I’ve seen it. It really works.
Unless the corporation arrives.
A committee forms. Policies are written. Somebody buys six months of “AI training.” Data must be “cleaned.” Legal needs a look. IT needs another look. The consultant needs more discovery.
By the time AI reaches the sales rep, technician, dispatcher, or account manager, it has been wrapped in enough policies, rules, and procedures to choke an Indigo. The dumbification of the LLM.
It’s the way. The old way. Take a new technology and force it to behave inside aged processes.
That’s how companies automate bad decisions.
The marketing doesn’t help.
“AI-powered” often means one task leaves the application, visits an outside model, and comes back. “Predictive” often means the software recognizes a pattern in last year’s numbers. “Built for your industry” often means somebody uploaded a stack of PDFs.
None of that means the product has no value. It just means the way it’s being described can be vague or misleading. Go figure.
Your defense? Ask the oldest question in sales.
What does it do?
Not “How does it do it?” Nobody cares about transformers, harnesses, APIs, or MCPs.
Does it reduce the time it takes to create an accurate proposal? Route a service call? Find a pricing mistake? If the answer is “better decisions,” you still do not have an answer.
And, while the salesperson explains what disappears, ask what remains.
Who preps the data, and who checks the answer? Who fixes it when it’s wrong? Who takes the call when the customer says, “Your AI service chat-thing did something stupid?”
We asked the same questions during document management projects, MPS rollouts, ERP installations, and every digital transformation that arrived carrying a three-ring binder and a promise to change the world.
The binder is gone, but the questions are still relevant.
The House of Cards and AI
Keep asking and another detail appears: the new AI-enhanced application uses several models.
Claude handles one task. Gemini handles another. OpenAI handles a third. Perplexity does research. Grok may be wandering around somewhere back in the toner aisle.
The explanation sounds advanced: each model has strengths, so the system sends each task to the model best suited for the job.
I get it, and I don’t like it.
I would rather hear that the application runs primarily on one model than listen to a description of five models stitched together to create one AI engine.
The experts will disagree. Fine.
To me, a collage of LLMs creates more dependencies, more pricing structures, more security terms, more updates, and more places for something to break.
This is not theoretical. Google’s Gemini deprecation table lists dozens of model endpoints already shut down or scheduled for shutdown. OpenAI was expected to shut down its Assistants API on August 26, 2026, forcing developers to migrate applications built on it. A retirement may not kill the application, but somebody has to replace the model, reconnect the system, and test it all over again.
A connection that worked on Monday gets patched on Wednesday and fails on Friday.
The screen looks the same. Underneath, several companies are making decisions the seller may not control.
Ask one question: When a model or API changes, who owns the problem?
That answer tells you more than the demo.
So, what should you, the copier dealer, do?
Here is the part that will bother the committee and all the freshly minted “AI experts” out there.
For now, do nothing.
Do not build an AI committee. Do not launch a corporate program. Do not clean every database, schedule six months of training, or hire consultants to measure a return nobody has defined.
That does not mean ignore AI.
Pick one frontier model. Pay for it. Use it every day against work already sitting on the desk.
Then turn employees loose on the problems they already understand.
The gains will show up before the committee finishes defining them. That will happen faster than another formal program built around cleansing, implementation, retraining, re-evaluation, and a consultant who keeps discovering more consulting.
Give people the tools. Watch what they do.
And if management insists on training, train curiosity. The prompt class can wait.
AI Columnist Greg Walters
About the columnist: Greg Walters is a writer, analyst, speaker, and longtime technology operator who has led managed print and IT initiatives for several organizations. He is a founding member (and past president) of the Managed Print Services Association (MPSA) and creator of The Death of the Copier blog. Walters also co-founded the Cricket Continuum, which works to convert office technology dealers into robotics resellers and service partners. Future columns will continue to focus on artificial intelligence in the copier dealer community. If you have curiosity about the subject, drop us a line, and we’ll see if we can come up with something.
