AI at Work: What's Ready, What's Risky, and What Should Wait?

AI is moving fast, and most businesses feel the pressure to keep up. But moving fast and moving well aren't always the same thing. That tension sat at the center of our recent brainstorming session on AI adoption, where we talked through what's working today, what's risky, and what's probably worth waiting on.

We didn't land on a simple crawl, walk, or run answer. What came out of the conversation was more useful: the right pace depends on the process in front of you.

Start with the problem, not the AI

A lot of AI projects begin with "Where can we use AI?"

A better starting point is "What problem are we trying to solve, and is AI actually the right fit for it?"

AI tends to deliver the most value when the problem is clearly defined, the data behind it is reliable, the outcome is measurable, and someone can check the output. 

Often, that isn't the flashiest use case. It's the report someone builds by hand every week, the stack of documents that needs sorting, or the repetitive task eating up hours of a skilled person's day.

And sometimes the answer isn't AI at all. If a process is structured, rule-based, and repetitive, like moving data from one system to another or submitting filings to a court portal, traditional RPA often handles it faster, cheaper, and more predictably.

AI earns its place where the rules run out: reading unstructured documents, interpreting messy correspondence, or making judgment calls. Many of the best results come from combining the two.

Treat AI like a new hire

Gerald Paulraj framed it in a way that stuck with the room: "Think of AI as a Digital Employee or a Junior Intern."

It's a helpful way to look at it. You wouldn't hand a new intern the keys to every system on their first day.

You'd give them a clear role, the tools they need, and someone to check their work. AI deserves the same approach.

The more it can access and act on by itself, the more permissions, monitoring, and guardrails matter.

This becomes even more important as AI shifts from generating answers to taking actions. An assistant that drafts an email is one thing. An agent that logs into a system, makes changes, and messages customers is a very different level of responsibility.

This is where RPA plays a quiet but important role. In many setups, AI decides what should happen and RPA bots carry it out, working within defined permissions, following set steps, and leaving an audit trail of every action. It's a practical way to give AI real capabilities without giving up control.

Not every process deserves the same autonomy

Shantanu Gangal offered a useful lens: "Agentic AI is three-dimensional axis. Value they deliver, frequency of use, and level of compliance exposure."

A low-risk internal task can usually be automated with confidence.

A customer-facing or regulated process is another story.

So beyond asking "Can AI do this?", it's worth asking "What happens if AI gets this wrong?"

Don't wait for perfect

Chris Ball raised a fair point: "If humans aren't 100% accurate, why expect AI to be?"

Accuracy still matters, but perfection can't be the bar.

AI might get a process 90% or even 99% of the way there, and whether that's good enough depends on the work.

For some internal tasks, a person can review and fix the output in seconds. For others, that last 1% could mean a compliance issue or an unhappy customer.

The more practical comparison is how AI stacks up against the current process, and what happens when something goes wrong.

That's why back-office work is often a smart place to start. It's easier to control, audit, and review, which gives teams room to learn how AI behaves before using it in more sensitive areas.

Voice AI is a good example. It handles structured conversations well, but open-ended calls are still tricky, especially in regulated collections, where specific disclosures have to be made every time. Starting with lower-risk or inbound calls is a more sensible way in.

Keep moving, but know where you're going

Chris Taylor summed up the pace of it all: "I feel like I need to sprint – and that mile never ends. Keep advancing. Keep experimenting. Keep finding better AI tools."

Waiting for the technology to settle down isn't really an option. But moving forward doesn't mean giving AI free rein. It means experimenting, measuring results, and expanding autonomy where the process can handle it.

So, crawl, walk, or run?

It depends on where you're running.

Some processes are ready for AI today.

Some need a human in the loop.

Some need stronger controls first, and some simply aren't ready yet.

The companies that come out ahead probably won't be the ones that adopted AI the fastest. They'll be the ones that got good at matching the right level of AI autonomy to the right problem. The goal isn't to automate everything. It's to know what should be handled by rule-based automation, what can be assisted by AI, and what should stay human.