Beyond the productivity hack: what enterprise AI actually needs
Most AI agents are trained to be brilliant generalists. What businesses need is a specialist, one that knows their operations well enough to make a decision, not just describe the options.
By Carsten Thoma, president of Celonis
If six months ago was about the wonder of what artificial intelligence (AI) could do, the conversation today is about the work of making it happen. Every time I look at my phone, I have a notification about a new model that’s a million times better than the one launched ten minutes ago. We’ve all seen what tools like Claude Cowork can do, and we all have that one colleague who brags about how they’ve become ten times more productive (maybe it’s you…).
It’s exciting. But it’s not scaling. According to McKinsey's November 2025 State of AI survey, 88% of organisations are now using AI in some way. But in any given business function, no more than 10% report having scaled AI agents. There is a fundamental gap between a personal productivity hack and a robust enterprise solution.
AI models are brilliant generalists. They predict based on patterns they’ve learned from the internet, which means they know what an invoice is and how a supply chain works… generally. They’re trained on generic data, so they give you generic answers. But when it comes to your business, you don’t need a colleague who generally knows what’s going on. You need a specialist who makes smart decisions and gets the work done right.
A fortune cookie gives good advice that generally sounds nice. “The seeds you plant today will bloom tomorrow,” or “Believe in yourself and others will too.”
But I wouldn’t bet my business on a message from a fortune cookie. And I wouldn’t bet on AI that doesn’t have business context.
The high cost of AI’s operational blind spots Having AI guess at the reality of your operations is not just unreliable; it’s also inefficient and costly. When you ask an LLM to work out why an order is delayed, it will spend time and money trying to navigate complexity it doesn’t understand. Just working out the answer might cost you more than the solution. To be effective, your AI agents need to know your business inside and out: what’s happened in the past, what’s happening now and what should happen next. The same things that go through your human brain when you’re making a judgement call.
Data does not equal context Recently, we were working with a major manufacturer that had spent a lot of money putting AI agents in the hands of its supply-chain inventory-management team. They brought in data from its enterprise resource planning (ERP) systems, from its supply-chain management (SCM), from spreadsheets. And still, their agents weren’t able to make head or tail of it. The team asked the agent: “We’re seeing a spike in demand. Should we order more materials or risk potential stock-outs?” An important decision. Cost versus risk. Supply and demand. A tale as old as time. The agent responded: “A spike in demand can result in stock-outs. You may want to order more materials.” That’s not an answer. That’s a cop-out. But when we fed the agent true business context, it became a specialist. “Based on historical trends and seasonality, your exact stock-out risk is only 14%. So, I advise against ordering more materials at this time, and instead transferring them from your warehouse in Denmark, where they have excess stock. And I recommend shipping that stock by truck, since air freight is currently delayed in that region.” Now that’s an answer. That’s an agent you can trust to reason, decide and act autonomously for your business.
If you don't have a digital twin of your operation, complete with business context, it’s almost impossible to understand where to deploy AI and automation workflows!”
- Stephanie Bova, corporate vice-president and digital transformation officer, research & development, Novo Nordisk
Getting to context isn’t easy Providing this context is hard because it’s scattered across systems and applications, hidden in documentation, and bouncing around inside the heads of your people. And it only becomes harder when your vendors are building walls around your data. It’s not just data you need, but process data: an ontology that describes how work gets done in your business as a sequence of events. The steps, decisions, exceptions and interdependencies that make up every process. Nor is it just data and business knowledge. You also need intelligence: a way to enable agents to reason, predict, simulate what-if scenarios, weigh pros and cons, and make recommendations. And beyond data, business knowledge and intelligence, you need it all together in one place, secure yet openly accessible to your AI agents, and evolving dynamically as your business grows and changes every day. In short, it’s a tall order. A major challenge. But it’s one that we think is solvable. Because we have to solve it. If we don’t, AI will never fulfil its true potential, and we’ll all be stuck cracking open fortune cookies, hoping one of them has the answer we’re looking for. To stop relying on fortune cookies and start building a smarter business, download The AI-Driven, Composable Enterprise ebook today.
