The context tax: the AI use case hiding in plain sight

You’re reading The AI Directive — practical AI intelligence for business leaders. Subscribe here if someone forwarded this.

Ask an experienced colleague to resolve a customer escalation, prepare a supplier decision or explain why a project has slipped. They rarely begin with the decision itself. They begin by reconstructing what happened across email, meeting notes, a CRM record, a spreadsheet, a service ticket and somebody’s memory. One source is current, another incomplete, and two appear to contradict each other.

I think of this as the context tax: the repeated cost of gathering, checking and organising the information required before useful judgement can begin.

Microsoft’s 2023 Work Trend Index found that 62% of respondents struggled with too much time spent searching for information. Its 2025 telemetry found that Microsoft 365 users were interrupted, on average, every two minutes, while 52% of leaders described their work as chaotic and fragmented.

The opportunity is not simply better search. It is helping someone reach a reliable understanding of the situation faster.

The obvious AI answer is to summarise the material. That can help, but a smooth paragraph may hide which source is current, where accounts disagree, what remains unknown and whether a conclusion is evidence or inference.

For material work, the useful output is closer to a decision-ready context pack. It should show:

AI is not being asked to sound informed. It is being asked to assemble the conditions in which a human can become informed quickly. Microsoft Research recently made a similar distinction in its work on memory for AI agents: long interaction records could overwhelm the system, while structured, reusable and decision-relevant knowledge improved performance.

The same principle applies to organisations. More documents are not the answer if the person still has to work out which ones matter.

The context tax appears in many forms.

A customer escalation needs the order history, previous complaints, promises and delivery evidence before anybody decides how to respond. A project review needs the last decision, current milestones, unresolved dependencies and changes since the previous meeting. A finance close needs evidence, mismatches and ownership assembled before an exception can be cleared.

These processes share a useful pattern: skilled people repeatedly gather scattered evidence before they can apply judgement. That is a stronger starting point than asking a room of managers to brainstorm where AI might help. Look for work where the same questions recur:

If those questions consume experienced time every week, there may be a worthwhile use case.

Context assembly sounds low risk, but it still needs design.

The AI needs authorised access to the right sources, rules for recency and identity, and a way to cite what it used. It should distinguish fact from inference and flag missing or conflicting evidence rather than quietly choosing a convenient version.

The human role remains explicit. In a customer case, AI may reconstruct the history and identify the unresolved promise; a manager still decides the remedy. In a supplier review, it may highlight uncertainty; the commercial owner still decides whether the risk is acceptable.

Anthropic’s recent study of its own AI-heavy workplace illustrates the other side of the same issue. When decisions were made in meetings without a durable note, agents lacked the context needed to update their goals and assumptions. The technology was capable; the organisational context was missing.

A context use case should be measured against the work it replaces, not the quality of the demonstration. How long does it take an experienced person to reach a usable understanding? How often is evidence missed? How much rework or delay comes from stale or conflicting information?

Then test whether the AI-supported version improves those measures without weakening confidence in the evidence.

The result may not be a dramatic autonomous agent. It may be a reliable briefing assembled in five minutes instead of an hour, with the source trail intact and the uncertainty visible.

That is less theatrical than asking AI to make the decision. It is also much closer to business value.

Find one recurring workflow where an experienced person spends the first part of the job reconstructing what is going on.

List the sources they check, the conflicts they resolve and the evidence they need before they can act. Then measure the time between starting the search and reaching a decision-ready view.

That is your context tax. It may also be your next useful AI use case.

This is the first of three articles looking at practical patterns for finding AI value in real work.

Next: using AI to prepare difficult cases so scarce experts can focus their judgement where it matters.

See you Tuesday.
— Toby