Why operational processes stall between tasks, and how agentic AI keeps them moving

A request that should take a day often takes three weeks. A buyer asks for supplier certificates, a customer sends an enquiry with half the specification, or an auditor asks where a figure came from. In each case, the work itself is modest: someone checks a document, finds an answer and sends a reply.

The delay rarely comes from the task. It comes from the time between tasks, when the work waits for a reply, a missing document, someone to notice it has stopped, or the one colleague who knows where things stand.

Most operational delay sits in these gaps. This is where agentic AI can contribute most, and where, so far, most AI has contributed least.

Most of the time a process takes is spent waiting, not working

In their study of how companies compete on speed, George Stalk and Thomas Hout traced orders from the moment they arrived to the moment they were delivered. Only 0.05% to 5% of that time was spent working on the order itself. Their figures came from manufacturing and order handling, but the pattern is familiar in any operation.

The point is not that people sit idle. They are usually busy with other work. It is the work item that waits: in a queue behind other requests; in a clarification loop because information arrived incomplete or wrong; in a handover between teams; in an inbox awaiting approval; in rework after an error; or in a batch that is only dealt with at the end of the week.

Knowledge work shows the same pattern. Asana's Anatomy of Work Index 2021, a survey of more than 13,000 knowledge workers, found that people spent 60% of their time on what it called work about work: searching for information, chasing status and coordinating with others. Only 26% went on the skilled work they were hired to do.

The implication is that making each task faster achieves little while the work still waits between tasks. The larger gains lie in the gaps.

Uncertainty, rather than task difficulty, keeps work waiting for people

In the 1970s, the organisational theorist Jay Galbraith offered an explanation that still holds. The more uncertain a task, the more information has to be processed while it is being done. When information is missing or unclear, someone has to find it, check it or decide what to do. Organisations cope by adding rules, layers of approval, spare time or more coordination.

In operations, uncertainty is routine: a certificate has expired, a drawing and an enquiry use different units, a supplier answers a different question from the one asked, or a reply never arrives. Each is minor in itself, yet it typically becomes a loop with a person who must notice it, chase it, wait, and check again.

This is the main reason work stalls. The tasks are rarely difficult, but every piece of uncertainty waits for someone to pick it up.

Unmanaged dependencies between tasks are the hidden workload

Thomas Malone and Kevin Crowston defined coordination as managing the dependencies between activities. One step needs the output of another. Two tasks compete for the same person. A larger piece of work depends on many smaller ones being finished first.

Most stalled processes trace back to a dependency that nobody is actively managing. A quote waits for a price check, which waits for a supplier, who in turn waits for a clarification that was never sent. Each person did their part, yet the process stopped.

Current approaches speed up tasks but leave the gaps between them to people

Most AI in operations today takes one of three forms. Each has value, but none takes responsibility for the work between tasks.

AI that analyses data. Dashboards, forecasts and insight. Valuable, but it assumes the data is already clean, complete and in place. Getting it there, which is most of the operational effort, stays with people.

Platforms that ask organisations to reshape their processes around them. In 1990, Michael Hammer argued that organisations should not automate their old processes but redesign them from scratch: don't automate, obliterate. The idea has merit, and some processes deserve it. But rebuilding is slow, costly and risky, and most organisations cannot stop running to do it. A platform also still waits for people to feed it the right inputs at the right time.

Tools that process documents. They read, extract and summarise, and they do it well. Then they hand the result back. The chasing, the follow-up and the memory of where things stood remain with people.

Carrying a process means owning the work between tasks

Carrying a process means owning it between the tasks, not only performing them. In practice, that takes five things.

  • Knowing the state of the work. What is done, what is outstanding, who owes what and by when, at every moment.

  • Remembering. The context of a thread across days and weeks: what was asked, what was promised, why something was decided, and where it was left.

  • Handling uncertainty within a mandate. A missing document prompts a request to the right person. A mismatch is checked against the source. An unclear answer gets one precise follow-up question. A silence gets a reminder.

  • Keeping momentum. No step waits for someone to notice it has stopped.

  • Bringing people in to decide. Briefly, and with everything prepared, because the system has already done the work up to that point.

Consider a supplier evidence request. A buyer asks for current certificates from 12 suppliers, and the system identifies the certificates required and requests them. Two suppliers reply with expired documents and are asked for current versions. One sends a valid certificate in an unexpected format, which is read regardless. Three do not reply and are followed up. By the third week, the evidence is complete, and the sustainability lead reviews two judgement calls before signing off. No one on the team spent those three weeks chasing.

A system can only carry a process it is embedded in

A system can only carry a process it is part of. That means working inside the operation as it runs: the same inboxes, the same systems of record, the same people and the same rhythms. It also means configuration that goes as deep as the work needs: the organisation's rules, its exceptions, its formats and the order in which its people decide.

This is the opposite of asking an organisation to move into a new platform. Nothing has to be rebuilt first. And because the system works with the organisation's own systems, the workflow, its data and its records remain the organisation's own.

People gain time and come to each decision prepared

The obvious change is time. Hours spent chasing, checking status, and rebuilding context come back. The less obvious change is attention. When the system carries the routine progress of the work, people spend their time on the decisions, relationships and exceptions that need them, and they come to each one prepared.

Questions to ask of any system that claims to run a process

  1. Where does our work wait today, and who notices when it stops?

  2. What does the system do when an input is missing, late or wrong?

  3. Does it remember the context of a thread from last week?

  4. Does it work inside our operations, or ask us to move into it?

  5. When it brings a person in, how much of the work is already done?

How Mitochondria's agentic AI systems carry the process

The approach. Each unit is configured to an organisation's rules, formats and systems, and embedded deep within its operations. It owns its part of the process from start to finish: tracking what is outstanding, remembering context, resolving what it can and following up on what it cannot.

What it enables. Work moves without waiting for someone to notice it has stopped. Uncertainty is handled as it appears, not when someone finds time. People are brought in for the decisions that are theirs, with the work already done around them.

The impact. Shorter cycles, fewer stalled processes and teams who spend their time on judgement rather than chasing. The workflow, its data and its records remain the organisation's own.

Read more about Mitochondria's approach.

Sources

  • Stalk, G. and Hout, T. M. (1990). Competing Against Time: How Time-Based Competition Is Reshaping Global Markets.

  • Asana, Anatomy of Work Index 2021 (press release)

  • Galbraith, J. R. (1974). Organisation design: an information processing view. Interfaces, 4(3).

  • Malone, T. W. and Crowston, K. (1994). The interdisciplinary study of coordination. ACM Computing Surveys, 26(1).

  • Hammer, M. (1990). Reengineering work: don't automate, obliterate. Harvard Business Review, 68(4).

—

Mitochondria builds agentic AI for evidence-led core operations. Research-led and model-agnostic. ISO/IEC 27001:2022 certified, with controls aligned to the GDPR, UK GDPR, India's DPDP Act and Australia's Privacy Act. Amsterdam and Pune. Request a working session →

Next
Next

How to measure AI value beyond cost savings