There’s no shortage of interesting things we can build. Figuring out which ones are actually useful is a different question.
AI has made that conversation a lot more interesting. The capabilities are moving quickly, the possibilities are real, and there’s plenty worth exploring.
“We need AI” isn’t really a requirement.
It’s easy to understand how companies get here.
AI is showing up in CRMs, marketing platforms, productivity software, customer service tools, development environments, and just about every other category of business software.
And businesses are getting real value from it.
A March 2026 Goldman Sachs survey of small businesses found that 76% were already using AI. Among those using it, 93% reported a positive impact and 84% cited increased efficiency or productivity.
But there was a much bigger gap between using AI and actually integrating it into the way the business operates.
Only 14% said AI was fully integrated into core operations. Businesses also cited data privacy and security concerns, lack of technical expertise, and difficulty choosing the right tools as common obstacles.
That distinction is important.
Giving employees access to ChatGPT, enabling an AI feature in the CRM, or buying another AI-enabled product may all be useful.
But none of those things tells us what problem we’re trying to fix.
Start with the friction
If someone tells me they want AI in their CRM, I’d rather start with questions like:
- Where are people spending time on repetitive work?
- What keeps falling through the cracks?
- What information is difficult to find?
- Where are people entering the same information more than once?
- What decisions take longer than they should?
- What do employees repeatedly have to search for, summarize, compare, or rewrite?
- What would actually be better six months from now if we solved this?
Once we understand that, AI may absolutely be part of the answer.
But we may also discover something much simpler.
If sales follow-up is inconsistent, maybe the first answer is workflow automation.
If information is spread across three systems, maybe it’s an integration problem.
If the CRM is full of duplicate or incomplete data, putting AI on top of it doesn’t make the underlying data better.
If people aren’t using the system because the process itself doesn’t make sense, an AI assistant probably isn’t going to fix that either.
On the other hand, if the problem involves finding patterns across a large amount of information, summarizing activity, matching candidates or customers, understanding unstructured information, or reducing repetitive knowledge work, now we may have a genuinely useful AI problem to solve.
AI itself also covers a pretty wide range now. The answer could be a simple copilot helping a person with a task, an AI-enabled workflow operating behind the scenes, or an agent that can work through multiple steps with some degree of autonomy.
The more capable the technology becomes, the more important it is to understand what we’re actually asking it to do.
Building agents reinforced that lesson for me
I’ve been experimenting with AI agents in controlled development environments and looking at where they can realistically help with technical work.
The technology is interesting, but one of the bigger lessons has been that the model itself is only part of the solution.
Agents can accomplish significant work, and agentic systems are becoming capable of sustained, multi-step work. That progress makes the right environment—including context, permissions, verification, and recovery—more important, not less, if they are going to produce the intended result consistently.
An agent may need access to files, source code, business data, email, or other systems to do useful work. That means privacy, permissions, supervision, approvals, and recovery become part of the implementation — not something you figure out later.
In other words, you still have to understand the process.
Putting a more capable AI model in front of a poorly defined problem doesn’t magically turn it into a well-defined one.
This lines up with broader research as well. A 2026 survey from Deloitte’s AI Pulse Check series found that 48% of respondents said their organizations had introduced AI without redesigning the workflows or roles around it. Only 12% reported redesign at scale with a new operating model behind it.
That doesn’t mean AI isn’t producing meaningful results. But the larger opportunity appears to come when companies rethink how the work gets done instead of simply adding AI to the process they already have.
Sometimes the right answer will be AI
None of this is an argument against AI.
I’m spending a fair amount of time experimenting with it because I think there are some very real opportunities here.
But I don’t think the goal should be to find somewhere to put AI.
The goal should be to find the places in the business where technology can remove friction, improve a decision, reduce repetitive work, connect information, or make something practical that wasn’t practical before.
Then decide what technology belongs there.
Sometimes that’s AI.
Sometimes it’s an integration.
Sometimes it’s automation.
Sometimes it’s replacing a platform that no longer fits.
Sometimes it’s a relatively small change to something you already have.
And sometimes the answer is to leave the technology alone and fix the process.
For me, the better place to start is with where the business is running into friction. Once we understand that, it becomes much easier to decide what technology might actually help.
So when someone says:
“We need AI.”
I’m increasingly inclined to respond with another question:
What would you like it to fix?
That seems like a much better place to start.
And I’m curious what other businesses are seeing:
What problem would you actually want AI to solve?
Working through a technology decision?
If you're considering AI or automation and want to understand where it could help, get in touch.
Get in touch