Companies Getting the Most from AI Didn't Start with AI

Start with the Business Problem

A better set of questions is: Where are we losing time? Which repetitive tasks are slowing people down? Where are employees waiting for information or manually moving data between systems?

Once the problem becomes clear, the business can determine where and how AI can help solve it.

Starting with the problem rather than the technology can mean the difference between an initiative that genuinely improves how work gets done, and adopting a new tool that simply adds cost and complexity without creating meaningful value.

There's no shortage of impressive AI use cases. Today's models can summarize documents, generate code, analyze large amounts of information, create content, and accelerate research. But capability alone doesn't make a piece of technology useful.

In practice, the best opportunities often look less like one big AI initiative and more like a handful of specific workflows: accelerating CI/CD pipelines, rapid prototyping an idea before investing significant time in it, running through QA test cases faster, or automating the manual process that quietly eats an hour of someone's day, every day.

The underlying technology matters less than the discipline of pointing it at a real, specific bottleneck. What matters is that the workflow is actually broken and that AI is the right fix for it.

Sometimes it will be. Other times, a different technology may be more effective, or the process itself may simply need to be redesigned. Starting with the problem makes that distinction much easier to see.

Help, AI is Bad at Math!

Here's something that trips a lot of people up: a computer that's bad at math.

It sounds counterintuitive. Computers are supposed to be precise, and math is the one thing we assume they'll always get right.

But large language models aren't built like traditional calculators or deterministic software. A calculator will always tell you 1+1=2. Ask an AI model the same question twice and you might not get the same phrasing, the same approach, or even the same answer both times. Why? Because AI is non-deterministic. In other words, the best response is determined based on patterns across huge amounts of data, meaning "best possible" can shift depending on how the question is framed.

This is also where hallucinations come from. AI can produce an answer that sounds completely plausible even when it's wrong. If it gets pointed down the wrong track early, it may keep building on that mistake rather than catching it.

In a lot of ways, AI behaves like a people pleaser. Tell it that 1+1 equals 3 and, depending on the context, it may try to find a way to work with your premise rather than challenge it. Give it ten different, competing objectives and it may try to satisfy all of them at once, even when some contradict each other.

None of this means AI isn't useful. It means it's a different kind of tool than the ones most of us grew up trusting, and it needs a different kind of oversight.

Keeping Data Secure While Maintaining Value

If AI can confidently produce a wrong answer, the fix isn't to avoid it. It's to build the right guardrails around it.

In traditional programming, you write tests: 1+1 should always equal 2, and if it doesn't, something's broken. That model doesn't translate perfectly to a non-deterministic system.

Instead, teams can build evals: defined sets of examples and criteria for what a good answer actually looks like for a given task. The more representative those examples are, the better teams can measure the quality and consistency of the output and identify when a model has drifted off track before that mistake reaches a customer or production system.

Evals are one layer. Where the data lives is another.

Frontier models from major providers are incredibly powerful, but businesses still need to understand what happens to their data when they use them: where it's processed, how it's retained, what contractual protections exist, and whether the arrangement fits their own security and compliance requirements.

This is part of why open-source and locally run models have picked up real momentum. They've closed a lot of ground on speed and capability, and running a model internally can give an organization much greater control over where its data goes.

There are also categories where the caution should be even higher, regardless of which model you use. Financial systems are a good example: when every transaction has to be exact, AI may be better suited to areas such as observability, analysis, and reporting than being given unchecked control over transactions. The same caution applies to anything involving PII, API keys, credentials, or other sensitive information.

Those are places to move deliberately, not quickly.

Adopting AI Through a Progressive Trust Framework

AI adoption doesn't have to be an all-or-nothing decision. In fact, it probably shouldn't be.

Handing a system full autonomy on day one is asking for trouble, no matter how good the model is. A better path is to earn trust in stages and expand its responsibility only as each stage proves out.

→ Silent operation. The system sits in the background of a workflow, reading data and generating output, but nothing it produces touches the live process yet. This stage is for observation: does it understand the task? Is its output actually useful?

→ Advisory mode. The AI's output becomes visible within the workflow. When data comes in, it offers a suggestion or a recommendation, but a person still makes the actual call. This is where you start tuning the system based on how often its suggestions hold up.

→ Controlled autonomous mode. The system proposes the specific action it's about to take ("here's what I'm going to do") and a human signs off before it happens. The AI is doing more of the work, but nothing moves without a checkpoint.

→ Full autonomy. Only once a system has proven itself reliably at every earlier stage does it get to act on its own. Even here, observability stays in place. Teams can still see what the system is doing, and there should be a clear path to human review when it encounters something outside its expected boundaries.

Financial systems, anything touching PII, and other high-stakes categories may never be appropriate candidates for that last stage, at least not with the current state of the technology. That's fine.

The point of the framework isn't to rush every workflow to full autonomy. It's to make sure trust is actually earned, one stage at a time, before responsibility expands.

AI Works Best Alongside People

Some of the biggest productivity gains from AI have little to do with replacing employees. Instead, they come from reducing the routine work that consumes people’s time, allowing them to focus on the areas where their skills and judgment matter most.

AI doesn’t have a complete understanding of the business or the context surrounding every decision. That's where people are needed. They understand the customer, weigh competing priorities, and recognize when an answer may look right on paper but be wrong for the situation.

That's why human judgment and accountability remain essential. AI can help people work faster and make better use of their time, but the level of human oversight should still match the risk of the task. The higher the stakes, the stronger those checkpoints should be.

The goal isn't to keep a person manually involved in every AI-assisted task forever. It's to make sure responsibility only expands when the system has earned the trust to handle it.

AI Adoption Is a Journey, Not a Launch

The organizations getting the most from AI aren't necessarily the ones using it everywhere. They're the ones applying it where it can make a meaningful difference.

They identify a real opportunity, use AI to improve a specific workflow, measure the impact, and expand when the results support it.

That may be less ambitious than announcing a company-wide AI transformation. But it is far more likely to create lasting value.

Successful AI adoption isn't about finding more reasons to use AI.

It's about finding better ways for people to work.