The Hidden Cost of Disconnected AI in the Enterprise

One afternoon in customer operations

A customer writes in about a missing delivery. The operations lead opens her assistant and asks it to draft a reply. The draft is fine. Then she opens the order system to find the shipment, the carrier portal to see where it is, the CRM to check whether this customer has complained before, and the ticketing tool to log the case. She pastes each finding back into the assistant, asks for a redraft, and sends it. Twenty minutes, six windows, one email.

The assistant contributed the paragraph. Everything that made the paragraph correct came from her. Multiply that by every request in her queue and the assistant has saved her some typing and nothing else.

Where the cost actually sits

The 2025 version of this post quoted a figure for minutes lost per context switch and multiplied it across an organisation. We have removed it, because a number nobody can trace to a source is not a cost, it is a slide. The costs that are real are these:

• Licence spend that buys drafting. Enterprise assistant seats are paid for on the assumption that work changes. If the assistant cannot reach the systems the work happens in, the seat buys a better first draft and the process stays exactly as long.
• Decisions made on pasted context. Whatever the person pastes in is what the assistant knows. It does not know the order shipped an hour ago, or that the customer's contract has a service credit clause. The output is only as current as the copy-paste.
• Data leaving through the clipboard. Every paste is a small, unlogged export. Nobody approved it and nobody can reconstruct it. The security team is right to be uneasy about it, and the answer is not to ban the assistant.
• Parallel tooling. When the sanctioned assistant cannot do the job, teams find one that can, wired to whatever credentials they have. That is the shadow AI problem, and it is caused by the sanctioned tool having no reach.

What connected looks like

The same request, with the assistant able to reach the systems: "A customer says order 48213 is missing. Sort it out." The assistant reads the order, checks the carrier's tracking, sees the parcel is held at a depot, reads the customer's history, drafts the reply with the new delivery date, proposes a goodwill credit under the contract clause and logs the case. The operations lead reads the proposal, approves the credit, and sends. Two minutes, one window.

The mechanism is the Model Context Protocol: a system is exposed once as an MCP server and any assistant that speaks the protocol can discover and call its tools. Nothing about the sequence above was scripted. The assistant had four described tools and worked it out.

Why connected is not allowed by default

An assistant that can reach the order system, the CRM and the carrier portal has to be governed like a person who can. The protocol does not do that. It authenticates a client to a server; an authorised client then sees everything the server exposes to it; a read on one order and a read on every order are the same tool; proposing a credit and issuing one are both just calls. So the security team asks who may use which tool, on which records, with what approval, and with what proof, and a pilot built straight on the assistant cannot answer.

This is the real reason enterprises are stuck with disconnected AI. Not that connection is hard, but that ungoverned connection is unapprovable. The clipboard is the workaround, and the clipboard is worse.

What makes connection approvable

Each person gets one governed connector, assigned through the groups they already hold in the identity provider, carrying every tool and Skill they are entitled to into whichever assistant they use. Policies are evaluated on the actual arguments of a call. Calls that need a human are held for approval. Every call is logged with who, what, when and why, and cost is attributed per team. That is what Palma provides; the assistant and the systems stay as they are.

With that in place, the operations lead's assistant reaches exactly the four systems her group is entitled to, the credit waits for her approval, and the security team has a log instead of a clipboard.

How to size your own disconnected-AI cost

Do not multiply minutes. Take one queue (support, AP, onboarding, whatever has a backlog), count the systems a person touches to close one item, and count how many of those touches are copy-paste into and out of the assistant. That number, times the items per week, is the work the assistant is not doing. Then pick that queue as the first one to connect.

If you want a second opinion on which queue, book thirty minutes.