Failure demand: the work your operation creates itself

Failure demand is the request that arrives only because something was not done, or not done right, the first time. The progress-chasing call is the easiest example to recognize. In a study of 534 incidents at an English police force, failure demand accounted for 30% of all actions. Automating that queue without separating it makes the symptom cheaper and keeps the cause.
What is failure demand?
It is the part of the queue the operation manufactures for itself. In 2003 John Seddon defined it as demand caused by a failure to do something or do something right for the customer. The definition is quoted in the paper by Walley, Found and Williams, published in 2019 in the International Journal of Health Care Quality Assurance.
Its counterpart is value demand: the new request, the one the team exists to handle. A quote request is value demand. "Did you get my quote request?" is failure demand, and whoever answers it sits in the same queue and uses the same person as a new request would.
The concept applies to external and internal customers alike. The salesperson asking finance whether the invoice went out and the manager asking again for a report that came back wrong are internal customers generating the same kind of request.
How much of the work is failure demand?
It depends on who measures and what goes into the category. We found two documented measurements. One came to about a third. The other, with a much wider definition, came to 72%.
The first is a study by Morris and Walley published in 2021 in Public Money & Management. About 12 officers at Gloucestershire Constabulary re-examined 534 non-urgent incidents from the first two weeks of February 2017, listening to the original call for each one. They found repeat or failure demand in 168 of them (32%), with one to seven instances per incident. The 534 incidents needed 353 extra interventions, 30% of all actions. One of the authors works for the force studied, which helps with data access and deserves the caveat.
The second comes from the National Audit Office report on HMRC customer service, published in May 2024. HMRC itself assessed that 72% of the calls it received in 2023-24 were caused by failure demand, up from 65% in 2018-19. The category is wide: it covers HMRC's delays and errors, customer errors, and progress chasing, including before the response deadline has passed. Within it, at least 13% of all calls were chasing progress.
Seddon states that the share runs from 20% to 60% in financial services and reaches 80% to 90% in local authorities and police forces. He is the author of the concept and the figures come without a stated method, so we treat them as a claim. Morris and Walley themselves report the 80% figure as something that is "claimed", and they measured far less.
Why does your system not show this demand?
Because it records activity. Walley, Found and Williams put the problem in one sentence: staff record the work done. The demand goes unrecorded. An answered ticket counts as output whether it was a new request or the third chase on the same one.
Queue categories hide it too. "Question", "status" and "adjustment" describe what the agent did. None of them says why the request existed. A dashboard built on those categories shows a busy team and rising volume, and the natural reading is that you need more people or more automation.
HMRC illustrates the point. According to the NAO, the department knew its failure demand share and had no data on the time and money spent on it. It had not set a target to reduce it either.
What happens when you automate the whole queue?
You get faster at answering questions that did not need to be asked. An assistant that reports status in seconds lowers the cost of each answer, but the request keeps being created, because the person asking still cannot see the status.
The NAO report has a data point in that direction. HMRC expects failure demand to keep rising as a share of calls as customers use its digital services more. The same report records stakeholders' view that those services give customers insufficient reassurance on the progress of their applications and queries. Our reading: the digital channel absorbed the simple request and left on the phone what the process does not resolve.
The alternative is to fix the step that generates the request. In Gloucestershire, assessors concluded that 55% of incidents could have been dealt with in the first call. After the control room redesign in 2018, 43% of police response in the first full month was a remote or first-contact resolution, against 26% in the same month a year earlier. That figure describes the design change and first-contact resolution. We did not find a new measurement of the failure demand share in the study. The problem sat in the design of the step before it sat in the tool that answers for it.
It is the same reasoning as choosing what to eliminate first: the task that annoys people most is not always the one holding up the flow.
A practical example: the finance queue at a distributor
The scenario below is hypothetical and the numbers only illustrate the method. A distributor has two people in finance answering salespeople and customers by email and chat. The innovation manager receives a proposal to put an AI assistant on that queue.
Before approving it, she reads 100 recent requests and classifies each one. Suppose 38 are status questions ("has the payment gone out?", "was the invoice issued?"), 21 are corrections to an invoice or a record entered wrong, and 41 are new requests.
With that count, the project changes shape. The 38 questions call for the salesperson to see the status in the system where the order was entered. The 21 corrections call for validation at the point of entry. The assistant, if it still makes sense, is sized for 41 requests instead of 100. There is one more gain the count does not show: every status question interrupts someone, and the interruption costs more than the time it takes to answer.
How do you measure failure demand in your operation?
With a sample read by people who know the process. The Gloucestershire procedure works at a smaller scale:
- Pick one queue and pull the requests from a short, recent period. One or two weeks is enough to start.
- Read the original request. The category the system stored will not tell you this. Ask whether the request would exist if the previous step had worked.
- For each failure, note the step that generated it: missing information, invisible status, entry error, a deadline with no notice.
- Count the repeats per case. In Gloucestershire, each incident with failure demand had between one and seven.
- Give a few identical cases to every assessor and compare the answers. The study did this with five incidents and found no difference in how repeated activities were assessed.
One caution: use the result as a diagnosis. Walley, Found and Williams describe the measure as strictly a diagnostic and indicative tool, and Morris and Walley record that the UK attempt to turn it into a national indicator failed. Once it is enforced as a target, the classification itself becomes contested, as happens with any metric that becomes a target.
Frequently asked questions
Does this actually remove the bottleneck, or just move the problem?
It removes it when the fix is made at the step that generates the request. The service queue is where the failure shows up, and the previous step is where it starts. In Gloucestershire, in the first full month after the redesign, 43% of response was a remote or first-contact resolution, against 26% in the same month a year earlier. That is one month of comparison, and the number measures first-contact resolution, without measuring failure demand itself. If the cause sits in another company or agency, the classification at least separates what you control from what you need to negotiate.
How do you measure the result of reducing failure demand?
By counting failure requests per case, before and after the change, on the same queue and with the same classification rule. Total request volume is a poor indicator here, because it moves with sales and seasonality. HMRC shows what happens without that count: it knew the share and did not know the cost.
Will classifying the queue take too much of the team's time?
The sample is small and happens once per diagnosis. The English study used about 12 officers part-time over two weeks to re-examine 534 incidents with a 67-field form. An internal queue with a five-question form and 100 requests is a few days of work for two people, by our own estimate.
The next step
If your queue keeps growing and nobody can say how much of it is new requests, the Operational Architecture Diagnostic is a 30-minute conversation to locate the step that is generating the rest. Talk to an architect.
Related case studies

Did you recognize your operation in this article?
Book the Operational Architecture Diagnostic: 30 minutes to map where your operation's bottleneck is. No proposal in the first conversation.
Book a diagnostic