Data access is not organisational understanding
Access to organisational systems cannot capture all the history, context and situated judgement needed for consequential work.
There is an attractive idea behind many AI-agent projects. Connect the agent to enough information and it will acquire the context required to do the work. Give it access to the CRM, SharePoint, email, policies, procedures and transaction history. Surely it now understands the organisation? Not necessarily. Organisations contain a great deal that is never completely represented in their information systems. People remember the history behind a difficult customer relationship. They know what happened the last time an unusual order was accepted. They know that two departments interpret the same rule differently.
They know that the stated priority this month is not necessarily what the managing director is actually worried about this week. Much of this knowledge is acquired through meetings, conversations, experience and participation in organisational life. Data access is not organisational understanding. This becomes important when we delegate work requiring judgement. We can describe objectives, rules, policies and known exceptions to an AI agent. What we cannot necessarily describe in advance is everything an experienced person would notice when confronted with a particular situation. That does not mean AI cannot perform consequential work.
It means delegation needs boundaries. Where situations are predictable and outcomes can be specified clearly, substantial delegation may be entirely sensible. Where circumstances are ambiguous, unusual or consequential, we may want the agent to stop and involve somebody who understands the wider context. This is why I think escalation deserves more attention in the design of agentic systems. It is not merely what happens when the AI “fails”. It is the point at which the organisation decides that the available representation of the situation is no longer sufficient for autonomous action. The challenge is not giving AI access to everything. It is knowing when access to everything we have recorded is still not enough.