Directing AI instead of just prompting it

The philosopher has set aside his quill pen for a moment and is now, oddly enough, holding a management journal in his lap. He looks equal parts amused and startled. "They've just arrived at something I was trying to say in article two," he mutters, pointing at a headline: Stop Prompting AI. Start Directing It.

Researchers Jennifer Sloan and Vern L. Glaser recently published a research-based article in MIT Sloan Management Review that draws a distinction worth pausing on – between having a conversation with an AI and directing it. And it turns out to sit directly in the continuation of everything we've already discussed in this series.

Conversation versus direction

When you "prompt" an AI, you're in a conversation: you ask, it answers, you evaluate, you ask again. You hold the thread yourself, and the exchange only exists as long as the window is open. That's exactly the situation we described in the context article – you are the model's System 2, the one who constantly has to fill in the context the model lacks.

An agentic AI works differently. Instead of holding a conversation, you configure a system that then acts independently over time, with access to more material and the ability to sustain an analytical thread across entire data sets – something a single prompt window can never manage. That's not the same as always being right: agentic systems can also drift off course, miss relevant information, or find correlations that aren't really there. But the ability to hold a long investigation together is new, and worth understanding on its own terms.

The researchers identify three things you're effectively configuring when building such a system:

Context – what the agent actually has access to (databases, documents, records), and unlike a one-off prompt, that access persists.
Capabilities – what the agent can actually do, not just say: run analyses, compare data sets, switch between tools without waiting for the next instruction.
Orientation – what the agent is asked to direct its attention toward. Not an instruction about what to produce, but a purpose and direction that shapes how it meets whatever it actually finds in the data.

That last point, orientation, is the piece missing from our own context list in the previous article. We talked about instruction, history, material, role, tools, and goal – but not specifically where attention should be directed once the agent gets to explore on its own. It's a nuance worth adding: the same material, given a different orientation, produces different discoveries. Not because the question changes, but because the system is redesigned.

Directing doesn't mean giving the system unlimited freedom, either. A director also sets the boundaries of the scene: which sources may be used, which actions require approval before they're taken, how results should be verified, and when the work should stop. Without those boundaries, it isn't direction – it's letting something loose and hoping for the best.

Four ways to create friction

The core of the article is that real insight rarely comes from a single, well-aimed question. It emerges from friction – from putting things in contact that are normally kept apart. The researchers describe four ways to deliberately create that friction, and all four feel familiar from our own series, just in a more hands-on, business-facing form:

Multiple lenses at once. Let several well-reasoned interpretive frameworks collide against the same material simultaneously, and let the contradictions between them point toward what no single framework would have found alone. This is the exact same mechanic as our "risks versus opportunities" example from the Kahneman article – scaled up to running four strategic frameworks in parallel on the same data and letting the friction between them surface a question nobody was asking.

Surfacing silences. Compare what actually exists in the material with what gets said out loud about it, and treat the gap as meaningful in itself. This is pragmatics in its most concrete form – what's not said carries as much information as what is. An AI isn't automatically less blind to an organization's blind spots than the organization itself; it's only as good as the material and perspectives it actually gets access to. But used well, with the right material, it can make the gap visible in a way that's harder for someone who's sat in the room for twenty years.

Bridging levels. A symptom rarely appears at the same level as its cause. Tracing a problem up and down through an organization – from individual decisions to overarching patterns and back – is exactly the kind of relationship map an ontology exists to describe.

Stress-testing categories. Compare your existing classifications against how reality actually behaves, and let the divergence be the finding. This is straight out of our "customer" and "address" exercise from the ontology article, just applied to real rejected deliveries instead of a hypothetical customer registry – same principle: the map (the categories) rarely matches the terrain (reality) perfectly, and the gap itself is where the insight lives.

Rule one: configure for discovery, not answers

The most counterintuitive advice in the article might also be the most important: the instinct, especially for someone used to ordering clear deliverables, is to tell the agent what to find. That's exactly the wrong move. An agent configured to confirm a hypothesis runs a real risk of doing exactly that – and in doing so, misses everything it could have discovered if it had been allowed to explore freely within a clear direction instead of toward a predetermined answer.

It's the human's job to set what the agent should pay attention to. Not what it should conclude.

The philosopher nods in agreement

"You see," says the philosopher, setting down the journal, "this is the same thing I've been on about the whole time. First you need to know what the words mean. Then you need to know which map of concepts you're even navigating. Then you need to know where on the map you're standing. And now, it turns out, you also need to decide where you're asking the system to look – otherwise it just looks where you already knew it would, and you learn nothing new."

Semantics, ontology, context, orientation. Four layers of the same old insight: the clearer you are about what you're actually wondering, the more AI can actually help you discover what you didn't know you were wondering.

And that very question – what you should actually be directing your attention toward – brings us right back to what we've already promised to talk about: how data actually tells a story, how it becomes worth something, and what happens when we give agents their own keys and a ticking clock.