Five years ago, I designed and implemented a blockchain network. The dream was to connect a large number of entities and create an ecosystem with a single source of truth, shared transactions, and none of the clunky enterprise integrations. The technology worked. But the world was not ready yet. The sponsoring organisation ended up running the nodes on behalf of the network participants as well, and exposing APIs to them. We had built a decentralised network, and then centralised it so that people could use it.

I have been thinking about that project while watching how agentic and AI-enabled solutions are progressing. They have good traction in the consumer domain. My view is that consumers are validating the answers from ChatGPT and Gemini themselves. Every user has quietly become an always-on human-in-the-loop. They read the answer, check it against what they know, and decide whether to act on it. The same is true of the newer autonomous agents that write code or respond to personal events. When something goes wrong, people take the responsibility themselves.

That’s not the case with businesses. There are social and legal responsibilities, and someone has to be accountable for every decision. Adding a human to the loop often makes the process slower, sometimes slower than the process it was meant to improve. So what is stopping AI from becoming a runaway success in the enterprise, compared to its public adoption?

As is often said, context is what makes generative AI effective. Building a good context for AI to operate in is the real challenge in many organisations. We tend to associate context with data and data quality. But that’s not all of it.

Process discipline is a key aspect of building a good organisational context. If we have taken an action and did not update the system in time, or left the decision in an email or a chat thread, then from the AI’s perspective it never happened. The context is stale. People cope with this because they were in the meeting, or overheard the conversation, or simply know whom to ask. An agent does not. The alternative is to write another AI agent to read the emails and update the system, which is not a solution, but a workaround.

Process discipline comes in many forms - standard process definitions, operating procedures, process automation, application integration and performance metrics. Together they make sure that the systems reflect what actually happened in the business, and when. None of this is new. We have talked about it for decades under different names. What is new is that the cost of not having it has gone up. A human colleague fills the gaps with experience and a phone call. An agent fills the gaps with a confident guess.

The next aspect of context is ontology. We comprehend data quite differently from an AI agent, and it’s not just about meaning or semantics. As soon as we read something or notice a change, we instinctively wire that detail to everything related to it, and draw correlations and causations without thinking. A delayed shipment is not just a status change. We know which store is waiting for it, which promotion depends on it, and which supplier was late the last time. Not so with agents. They see the record, not the web around it.

So we need to give agents the ontological graph - the structure and the interconnections. A taxonomy alone is not enough. It tells the agent that a smartphone is a kind of phone, and a phone is a kind of electronics product, but it does not tell it that this product is supplied by that vendor, sold in these stores and covered by that warranty. Nor is it enough to put all the data and documents into a knowledge base and let the agent search through it. Retrieval finds the text; it does not find the relationships. We need to build a network of knowledge - a knowledge graph - with the ontology as its blueprint. And we need to keep it up to date, which brings us back to process discipline.

A few days back Google released Google Agent, an idea to introduce AI as co-workers. Progress in AI is blazing ahead with new models and capabilities every few weeks. The ecosystem does not move at that speed. As with the blockchain network, the technology may arrive before the organisations are ready for it. It requires us to step back and prepare the environment for AI agents - discipline in how we run our processes, and structure in how we describe our world. Maybe AI can help with that, to start with.