AI knows everything... except how your business works. Here’s how to fix it.

HERO IMAGE

Devavrat Shah, MIT professor and Celonis Chief Scientist, explains why context matters, what happens when AI gets your business’s context wrong, and how to make smarter decisions.

Summary

  • Devavrat Shah, MIT professor and Chief Scientist, Enterprise AI at Celonis, joined MIT Technology Review’s Business Lab podcast to share his insights about Enterprise AI and operational context.
  • Your AI might be a genius. But it can still make terrible decisions. Without operational context, even powerful agents can make mistakes that cost companies dearly.
  • Give your AI agents a job description. See how the Celonis Context Model helps agents understand their roles and work within the rules of your business.
  • Fewer spreadsheets, more brainpower: AI could take on repetitive work and give employees more time to focus on strategic decisions.
  • Don't overthink your first move. Start with one decision worth improving, find where an agent could help, and measure whether it's actually delivering value.

Ask AI to explain quantum physics, and it’ll give you a full-blown lecture. Now ask it why your company’s invoices keep getting stuck in approvals, and it starts hallucinating. Would you trust it to make decisions without knowing what's going on behind the scenes?

Devavrat Shah, Chief Scientist, Enterprise AI at Celonis and MIT professor, shares his insight. In a recent episode of MIT Technology Review’s Business Lab podcast, he joined host Megan Tatum to discuss why business context matters for AI and what it could mean for the future of enterprise operations.

→ Listen to the full podcast: Making the case for context-aware AI.

An AI agent with little business context. What could go wrong?

Quite a lot, apparently. Shah believes one of AI's biggest opportunities is bringing what he calls decision intelligence to enterprises. Every company makes decisions through processes. AI agents could take on specific roles in those processes, helping people decide faster.

But as Shah puts it, "You can't just do that in a vacuum."

He explains that even a model with extraordinary knowledge could spend enormous computing resources figuring out the context behind a simple task. And when important information is missing, the consequences can get expensive.

"We'll miss some important context, important constraint, important guardrail, and then we'll make a mistake."

Shah shares a comparison: An AI recommending a disappointing restaurant might ruin your dinner. An AI producing a flawed financial forecast could, in his example, contribute to a company's downfall. That's an expensive lesson to learn the hard way. And exactly why Shah believes context matters so much.

Give AI a model of your world

So, how do you make sure your AI agents actually understand the business they're working for?

Shah shares that Celonis has spent years helping companies figure out how their processes actually run. Now imagine giving AI agents access to that same understanding, so they know where their job begins and what needs to happen next.

Enter the Celonis Context Model. Shah explains how it helps agents get a handle on the processes they're joining, with three capabilities working together:

  • Connect to workflow tools and their event logs to identify the parts of a process where agents could help.
  • Bring AI into specific roles in what Shah calls a "structured, guardrailed manner."
  • Connect with existing agentic infrastructure so outputs can be reviewed by people before they're approved or automated.

Worried your company's data isn't exactly in pristine condition? Shah has some reassuring news.

"They don't need to be complete. They don't need to be [structured] in the cleanest possible manner," he says of the systems already in place.

The Context Model is designed to connect to those systems and make sense of the information they generate. Of course, Shah emphasizes that proper information security checks still need to happen before connecting new software to enterprise systems.

Once agents understand the processes they're working in, companies can start asking: Could AI help us do this whole thing differently?

Fewer spreadsheets, more thinking

How much of the average workday is dedicated to coordinating with coworkers? Shah sees a huge opportunity for AI agents to take on work that currently slows down decision-making.

"You're taking decisions in time, you can plan much better," he explains.

There's only so much thinking you can do when half your day disappears into chasing updates and keeping work moving. Shah sees AI agents taking on more of that repetitive work, leaving employees with some much-needed breathing room to focus on decisions that actually need their attention.

Shah describes the potential shift as "a lot less repetitive, but a lot more strategic work."

That doesn't mean employees can simply hand everything over to AI and call it a day. Shah stresses the importance of educating people to work with agents and understand their outputs. He also expects new kinds of work to emerge, even if nobody can say exactly what all those jobs will look like yet.

Start with one decision

Figuring out where to begin doesn't have to be a workout for the old noggin. Shah suggests picking a business decision you'd like to improve and taking a closer look at how it's made.

Once you've got that figured out, here's what he recommends:

  • Find where things go haywire. Where does the process get complicated? That's where an agent might come in handy. Then, see whether the agent is actually helping. As Shah puts it, "let's measure how much speed are we getting, how much value are we getting."
  • Give agents ground rules. Nobody wants an AI agent going rogue. Shah explains that for each role, "we can tell what it can or cannot do very clearly."
  • Make sure people know what they're looking at. Employees need to understand what agents produce before they can confidently act on those outputs. Shah emphasizes the importance of educating people as AI takes on more work.

You don't need to hand over an entire department to AI on day one. Find one decision worth improving and measure what happens.

Planning that doesn't take an eternity

Shah sees potential for AI in the planning processes that consume enormous amounts of time across an enterprise.

Take financial planning or workforce planning. These are business-critical decisions, yet the processes behind them can stretch across lengthy planning cycles. With AI supporting the people making those decisions, Shah suggests they could happen considerably faster.

"Now, maybe it should not take that much time," he says.

Imagine wrapping up a planning process in weeks instead of months. Shah doesn't promise that timeline, but he does see an opportunity for AI to help people reach important decisions sooner.

Of course, those agents still need to understand the business they're helping. Even the smartest AI model can't be expected to know the operational reality of every enterprise on the planet (yet).

So, before handing your next big business decision to an AI agent, make sure it knows what it's getting itself into. Shah's advice is a useful reminder that even the smartest AI needs to understand how your company actually works.