The AI Investments That Don't Compound

Two years in, the pattern is becoming visible. Businesses that adopted AI tools in 2022 and 2023 are splitting into two groups, and the dividing line has almost nothing to do with the technology they chose.

One group has real leverage, decisions that are faster, better-informed, and consistently made at the right level of the organization. The other group has subscriptions. Sometimes expensive ones. And a quiet, lingering question about whether any of it was worth it.

I work with businesses on ERP integration and AI deployment in real commercial operations. The pattern I keep seeing isn't about platform choice or budget or technical readiness. It's about where the starting point was. When you start with a tool, the default path into AI adoption usually looks like a vendor demo, a compelling use case from a competitor, or pressure from the board to "do something with AI." The procurement happens. The tool gets deployed. Someone is assigned to champion it. And then, six months later, a small fraction of the team uses it regularly, the ROI calculation is murky, and nobody quite wants to say the quiet part out loud.

Starting with a tool is the default because tools are concrete. You can show them in a demo. You can get a quote. You can check a box. The problem is that tools are answers, and you haven't asked a question yet.

When you lead with a tool, the implicit question becomes: "How do we use this?" That question will always find an answer, because people are resourceful and no one wants to admit a purchase was wasted. So the tool gets pointed at whatever work is most visible. Usually automating tasks, reports, summaries, data entry. And sometimes that's fine.

But tasks that are visible are not the same as tasks that are valuable. And automating a process that's inefficient or unnecessary doesn't fix it. It just runs it faster. I've seen businesses automate workflows that existed only because their ERP was configured wrong six years ago. The automation didn't solve the problem. It cemented it.

When you start with a question, the businesses getting real compounding value from AI started somewhere different. Not with "what can this tool do" but with "where does a decision get made in this business, and would better information change that decision?"

Most organizations don't have a clear map of where decisions actually happen, not in org charts, but in practice. Who decides which leads get followed up with urgency? Who decides when to adjust pricing? Who decides whether a job is worth quoting? These are high-impact moments that often run on incomplete information, intuition, and whatever was in the last email thread.

When AI is applied to those moments not to replace the decision, but to improve the information available to the person making it, the value compounds. A better-qualified lead doesn't just improve one deal. It changes how the sales team allocates its time. Better job costing data doesn't just help one quote. It recalibrates the pricing model. The effect isn't linear.

I worked with two businesses in commercial operations, both of which implemented AI-assisted quoting in roughly the same period. One automated the quoting process itself. Faster turnaround, consistent formatting, reduced admin time. The other stepped back and asked what actually determined whether a quote was worth writing in the first place. They used AI to improve lead qualification, analyzing inquiry signals, job characteristics, and historical win rates before a quote ever got produced.

Eighteen months later, the first business had a faster quoting process. The second had improved their quote-to-close rate by restructuring where they focused attention. The first invested in speed. The second invested in judgment. Only one of those compounds.



The question worth asking before any AI investment isn't "what will this automate." It's: if the quality of this decision improved by thirty percent, what would change downstream in the business?

If you can answer that clearly, you probably have the right starting point. If you can't, the tool can wait.