You Wouldn’t Hand a New Hire the Keys to Everything on Day One. So Why Are We Doing It With AI?

I was twenty-two the first time someone handed me a login to an enterprise software system and walked off.

It was my first week on the technical help desk at IBM. Nobody sat me down with a manual titled “Here’s what you’re actually allowed to touch.” I was given access to a managed group inbox for client support tickets, a CRM, headset, and an unspoken assumption that I’d work out the boundaries myself. It took me an afternoon and a Severity 1 ticket to learn something that’s stuck with me for twenty years now: access and judgement are not necessarily mutually exclusive. One gets handed to you on your first day. The other you earn, usually through hard-lived experiences.

I bring this up because most companies rolling out agentic AI right now are taking the exact approach I experienced in that first week. Except the ticket isn’t a support case to remotely configure a router or reset an access point – it’s an active purchase order worth hundreds of thousands; or a maintenance plan for critical equipment. A decision that used to require someone with decades of experience working, is now made in seconds by a system that never hesitates.

We built the capability first. We’re only just getting around to writing the rules to ensure the guardrails are bolted down.

Adoption Was Never Going to Be the Hard Part

For a while, “adoption” was the only number that mattered. It’s what shows up in a board deck. It’s a clean percentage, a line that goes up and to the right, a slide that makes everyone in the room feel like the company is moving. Governance, by comparison, doesn’t photograph well. It looks like a slower rollout. A harder conversation. A committee somebody has to sit on.

So capability got the budget, and judgement got treated as something we’d figure out later.

The numbers back this up more starkly than I expected. By Deloitte’s most recent count, close to three-quarters of large organisations plan to be running agentic AI within two years. But only one in five say they currently have a governance model mature enough to handle what that means once it’s live. That’s not a small gap to be closing quietly in the background. That’s most of the enterprise world about to hand a new hire the keys before anyone’s written the handbook.

I’ve sat in enough rooms now with CIOs and operations leaders across mining and energy to notice the same pattern repeating: the pilot goes well, the demo impresses leadership, and then someone asks the one question nobody prepared for, “Okay, but who’s accountable if we get this wrong?” The room goes quiet, because the honest answer is usually “We haven’t fully worked that part out yet.”

An AI That Suggests Something isn’t the Same Animal as the AI that Just Goes Out and Does It

I think people flatten this distinction more than they should. One system tells you what it thinks you should do. The other one doesn’t wait to be asked.

A recommendation can be wrong and mostly harmless. Someone reviews it, catches the error, moves on. An action can’t be walked back the same way. If an agentic system places the order, reroutes the workflow, or overrides a threshold on its own authority, the mistake isn’t a suggestion that got ignored. It’s already happened.

And agentic systems, by their nature, don’t hesitate the way a person does. A new employee who isn’t sure will pause, ask someone, and flag it up the chain. That instinct is built from years of learning where the edges are. An AI agent has no equivalent instinct unless someone has deliberately built one in. Left to its own defaults, it treats “I’m technically able to” as “I should,” without the flicker of doubt that might have saved a person from the same mistake.

In the industries I spend my time across, this is not an abstract risk. A single automated decision, made confidently but without the right boundaries, can misroute a safety-critical part, or quietly compound into a cost that nobody notices until it shows up in a much bigger number months later. Not because the model was bad, but because nobody had defined exactly where its authority was supposed to stop.

What We Actually Built Around This Problem

This isn’t a position we’ve arrived at from the outside looking in. It shaped how we built our own platform, and it slowed us down in places where we know it would have been faster to ship without deploying these guardrails and investing the time to develop our context engine.

Every AI-generated recommendation our platform, MiLi, produces – a reorder trigger, a flagged exception, a stock strategy recommended change – is traceable back to why it happened, the moment it happens, and to the person who was accountable for that decision. We went through the process of formally achieving SOC 2 certification, and one of the principles that runs through that whole framework is one I think more AI vendors should be forced to answer to: access should be limited to exactly what’s needed, and nothing more. It’s the same idea as not handing the new hire the master key on day one. You give access as experience and aptitude is earned, not as a default setting.

That’s also where we draw the line between what our system is allowed to recommend and what it’s allowed to execute unsupervised. It’s not a philosophical stance. It’s a design decision, and in my experience it’s the single most consequential decision a company makes when it deploys agentic AI – even more consequential than which model they license or how many agents they deploy.

The Companies That Get This Right Won’t Be the Loudest Ones

They won’t be the ones with the flashiest agent demo on stage at a conference. They’ll be the companies whose AI can explain, in plain language, exactly why it made every decision it made last quarter – to a regulator, a customer, or a board member who wants a straight answer.

Adoption was never going to be the part that separated the companies that win from the ones that end up a cautionary case study. That’s the real shift happening right now, whether most leadership teams have named it yet or not: from AI adoption to AI accountability. Accountability doesn’t show up on a dashboard. It shows up in the room after something’s already gone wrong, when everyone’s trying to work out whose name is behind the decision. A lot of the industry is still writing that chapter after the system’s already live.

I’d rather have that conversation with my team before it becomes somebody’s postmortem than after. If you’re watching an agentic system get more capable by the month and quietly wondering who’s actually accountable for what it does next, don’t wait for the board to ask first. Have that conversation now.

Sladen Moses is the Co-Founder and Chief Executive Officer of Lexin, an AI-powered SaaS company helping asset-intensive organisations unlock working capital, reduce operational risk and transform indirect material supply chains.  

With more than two decades of experience spanning IBM and high-growth enterprise software businesses Syniti and Utopia Global, Sladen has built and scaled technology companies through international expansion, capital raises and successful exits.

Today, he leads Lexin’s global growth strategy, partnering with some of the world’s largest mining, energy and manufacturing organisations to deliver measurable commercial outcomes through enterprise AI.