What Is Actually Happening to the Work?

Suppose a dashboard tells you that your account managers spend a large part of their week in the CRM.

What should you do with that information?

They might be recording useful customer context. They might be searching for information that should already be available. They might be correcting records, duplicating updates or documenting decisions that need to be retained.

The number describes where activity occurred. The management question is what that activity was doing for the business.

That distinction becomes more important as humans and AI share the work. A completed agent task may be useful progress. It may also be the beginning of another person’s checking, correction or exception handling. You need to understand the surrounding work to know.

Anthropic’s September 2026 economic scenarios make a useful distinction between the work AI can affect, how widely it is used, and the productivity gains that follow. The authors note that company-level adoption figures overstate the share of work actually done with AI: an adopter may use it on only some tasks, and only some occasions. [3]

Inside a company, that distinction becomes a practical management question. Which work has changed, what new work has appeared around it, and has the whole arrangement improved? An adoption figure cannot answer those questions on its own.

The process people actually perform

A formal workflow is a useful starting point. It defines the intended sequence, responsibilities and controls. But real work includes the adjustments people make when information is missing, customers behave unexpectedly or systems fail to fit the situation.

Organisation researchers Martha Feldman and Brian Pentland distinguish the general understanding of a routine from its performance by particular people in particular circumstances. Formal documents can express parts of that understanding; they do not contain the whole routine. The way people perform it can also change the routine over time. [1]

That is a useful way to think about an operating model. It is enacted through decisions, conversations, workarounds and handoffs, as well as formal steps.

A process might say that an AI agent prepares a recommendation and a person approves it. In practice, the person may reconstruct the evidence, check another application and rewrite the recommendation before approving it. The recorded approval happened. So did a substantial amount of work that the process description failed to anticipate.

Equally, the review may be light, useful and exactly what makes safe delegation possible. The distinction cannot be settled by counting approvals.

Deep understanding means connecting the evidence

Different sources reveal different parts of the operation.

Application activity can show where work is concentrated. Workflow records can show sequence and waiting. Agent traces can reveal repeated attempts, tool failures or escalations. Customer and operational measures can show the results. People can explain why a step exists and what happens when it is removed.

The value comes from connecting those views.

Imagine a customer-onboarding team with a persistent delay. The formal process locates the delay at an approval step. Agent traces show that the preparation task finishes promptly. Employees explain that the reviewer regularly has to find missing context before making a decision.

The apparent approval bottleneck could actually be an information problem upstream. Removing the approval might make the recorded process faster while leaving the underlying risk unresolved. Improving the context supplied to the reviewer might achieve a better result.

This is an illustration, not a reported customer case. Its point is the kind of explanation a leader needs: one that connects observed activity to the way the outcome is produced.

An operating picture built for a decision

There is little value in trying to reduce an entire company to one definitive score.

A service executive trying to improve response quality needs to understand different relationships from a leader trying to reduce avoidable administration. Both may examine the same function, but the decision determines which aspects deserve attention.

A useful operating picture might connect the distribution of effort with the flow of information, the purpose of handoffs, the location of judgement, the work delegated to agents, the consequences of errors and the constraints the team must respect.

It should also preserve differences that matter. A change may help normal cases while making exceptions harder. Experienced staff may use an agent effectively while newer colleagues need more support. One team’s improvement may depend on another team absorbing additional work.

Those differences can disappear inside a reassuring average.

The goal is sufficient depth to choose a useful intervention. A leader should be able to understand what appears to be happening, why that interpretation is plausible and what further evidence would alter it.

Observations need interpretation

Decision science describes a related problem as partial observability: the underlying conditions are not directly available, so decisions depend on what can be inferred from incomplete observations. Work on partially observable decision processes formalises how a decision-maker can maintain a belief about those conditions and revise it as evidence arrives. [2]

An organisation does not have to implement a formal probabilistic model of everything to benefit from that distinction.

Consider low use of an AI tool. It might reflect weak capability, poor outputs, incomplete data, missing permissions or a workflow in which using the tool is more difficult than completing the task directly.

“People need training” is one possible explanation. It is not contained in the usage number.

A useful system should make the interpretation explicit enough to challenge. It might identify repeated context retrieval and failed access as evidence for a workflow problem, while retaining output quality as an unresolved alternative.

That gives the leader somewhere concrete to begin. More precision in the usage count would not necessarily help.

People help explain the system

Employees are sources of understanding, not simply subjects of analysis.

A repeated manual step may be where someone notices an error that the formal system misses. An apparent delay may be the time needed for a difficult judgement. A workaround may be compensating for a known defect.

Their explanations will not always agree with each other or with the available records. That disagreement is useful. It identifies where the model needs work.

A self-improving company needs a way to resolve those differences through evidence and discussion. People should be able to see the interpretation, explain an exception and challenge a conclusion before it becomes the basis for changing their work.

This also changes the conversation around adoption. If a team returns to an older method, the useful question is what the new method failed to provide. The answer may concern trust, information, incentives or practical effort. Treating every deviation as unwillingness to change can obscure the thing that needs fixing.

Understanding without turning work into surveillance

Trust belongs inside this design.

An operating picture intended to improve a team’s work should have a clear purpose and limits. It should focus on patterns relevant to the outcome, rather than creating individual rankings or personal activity histories.

Privacy protection does not remove the need for detail. It changes how that detail is obtained and used. Team and workflow evidence can be combined with people’s explanations, while sensitive material and individual exposure are limited.

There is a real design challenge here: excessive aggregation can hide who is carrying a new burden, but excessive granularity can expose individuals. The answer should preserve meaningful differences where it is safe to do so and seek additional explanation where it is not.

The aim is to make the operation discussable. Leaders and teams should be able to use the evidence together to identify problems that were previously hard to see.

From knowing more to choosing better

Deep understanding has a practical test: does it change the next decision?

Perhaps the company discovers that better source information matters more than another training programme. Perhaps it learns that a review step is valuable but needs a clearer threshold. Perhaps a team is ready to delegate more, once an unnecessary handoff is removed.

These are decisions about the arrangement of work across humans, AI and software. They connect activity to purpose.

That is central to the self-improving company thesis behind Teho. The business needs a living understanding of how it operates—one it can use, question and update as the work changes.

A clearer picture matters because it gives leaders and teams a better starting point for action. The next step is to make a change that tests that understanding, then see what the operation teaches them.

References

[1] Feldman, M. S. & Pentland, B. T. (2003). Reconceptualizing Organizational Routines as a Source of Flexibility and Change. Administrative Science Quarterly, 48(1), 94–118. https://journals.sagepub.com/doi/10.2307/3556620

[2] Kaelbling, L. P., Littman, M. L. & Cassandra, A. R. (1998). Planning and Acting in Partially Observable Stochastic Domains. Artificial Intelligence, 101(1–2), 99–134. https://www.cassandra.org/arc/papers/aij98.pdf

[3] Korinek, A., Jones, C. I., Sacher, S., Cotter, T. & McCrory, P. (2026). Economic Scenarios for Transformative AI. The Anthropic Institute, Working Paper No. 2026-02, section 3.3, p. 27. https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf#page=27