A supplier's insurance certificate lapsed two months ago and nobody noticed. A refund is sitting one hundred dollars above somebody's approval limit. A contractor cannot go on site until three documents check out. An access request is waiting on a manager who is travelling. None of these is important enough to fund a project for. Together they are somebody's entire week, every week, in every company.
This is the long tail of business decisions, and it is the part of operations that automation has walked straight past for twenty years. Not because anyone decided to ignore it. Because the arithmetic never worked.
The line that runs through every operations team
Automating a process properly is expensive, and the numbers are public. In the analysis Forrester published on a large Camunda deployment, building one automated process takes an average of 3,428 hours of developer and analyst time, falling to around 2,165 hours once the practice matures. At the study's own fully burdened labour rate that is roughly two hundred thousand dollars of work per process, before anybody pays for software.
The same model tells you which processes qualify. It assumes the average process worth that investment runs 328,000 times a year. At that volume the build costs about twenty cents per case over three years. At a thousand cases a year, the identical build costs sixty-six dollars per case.
Companies have known this for a long time and set the threshold explicitly. Telefónica, in one of the most cited accounts of an early automation programme, would not consider a process unless it saved at least three full-time roles or ran more than a thousand times a week. Everything else was described as too small to justify development resource.
The habit is so settled that the industry's own benchmarking excludes the tail by design. When Panorama built its ERP benchmark dataset, one of the documented cleaning rules was to discard any project costing less than a hundred thousand dollars, on the basis that below that figure it is not really a project.
So a clean line runs through every operations team. Above it, software. Below it, a person, a spreadsheet and a shared inbox, indefinitely.
Worth saying plainly: the platforms above that line are good. Process orchestration, case management and enterprise decision engines do serious work and do it well. The point is not that they fail. It is that their economics require volume, and most of the decisions a business makes do not have it.
Nobody has ever counted what sits below the line
We went looking for the number, expecting to find it. It does not exist.
The concept is well established in the research. A group of academics formalised the long tail of business processes and demonstrated empirically that process portfolios really are long-tailed, with a handful of high-volume processes and a very long trail of low-volume ones. They opened by noting that the unrealised improvement potential in that tail "has not been quantified."
Then they could not quantify it either, and they were honest about why. Measuring processes at company scale requires those processes to already run inside systems that produce event logs. The long tail does not. It runs on email, attachments, spreadsheets and what somebody remembers to check. There is nothing to mine.
That is the detail worth sitting with. The reason nobody has measured the long tail is the same reason nobody has automated it. Both fail for want of a system that can see the work.
When somebody does bother to count, the numbers are large
Two places have been counted properly, and both are instructive.
Health care administration in the United States was costed rigorously in a peer-reviewed study published in the Annals of Internal Medicine: 812 billion dollars in 2017, or 34.2% of national health expenditure, against 17.0% in Canada. A third of all spending, on administering the spending.
Governments have a formal method for this too. The Standard Cost Model exists specifically to quantify administrative burden, and the European Commission applied it to just 72 pieces of legislation and arrived at 123.8 billion euros.
Both of those measure burden imposed from outside, by regulation or by a payment system. Neither touches the far larger pile a company quietly imposes on itself through its own internal processes. That slice has been named, theorised, shown to be long-tailed, and left uncounted.
The decision was never the slow part
There is a second reason the tail resisted automation, and it shows up everywhere once you look for it. The decision at the end of these processes is rarely the difficult bit. Assembling the evidence in front of it is.
The clearest illustration comes from health insurance. According to the CAQH Index, 98% of medical claims are now submitted electronically. The supporting documents that prove those claims are 68% fully manual, still moving by fax, post and email. The decision got automated. The evidence never did.
Trade finance tells the same story in a different vocabulary. The ICC Banking Commission estimates that between 65% and 80% of document sets presented under documentary credits are refused on first presentation. The listed causes are timing, conflicting data between documents, missing endorsements, unauthenticated alterations, wrong port names. Almost none of them is a disagreement about the underlying credit.
Accounts payable is the same shape again. On Ardent Partners' 2025 benchmark, 35.4% of invoices are processed straight through, the invoice exception rate sits at 18.4%, and AP teams spend 21.9% of their time answering supplier questions about what is missing or where something has got to. Best-in-class organisations, the ones who already bought software, reach 51% straight through. That still leaves half the invoices needing a person, which is where the cost of manual invoice approval actually accumulates and why three-way matching exceptions consume so much of a finance team's week.
Even payments, an industry nobody would describe as unautomated, breaks in the same place. Swift network data published by the Financial Stability Board shows 88.5% of wholesale cross-border payments cross the network within an hour, while only 61.7% are credited within an hour once they arrive at the receiving institution. The transfer is instant. The handling is not.
Four industries, four vocabularies, one shape: evidence scattered across several places, a rule that is not in dispute, and a person waiting.
These processes are not unique to your company
Here is the part that turns an observation into a product.
APQC, a non-profit, has spent more than thirty years cataloguing business processes in its Process Classification Framework, now on version 8.0. The cross-industry edition organises work into 13 categories and well over a thousand processes and activities, written explicitly to hold regardless of industry, size or geography, with 19 industry-specific editions alongside it.
Cataloguing is one thing. Measuring is the proof. Through Open Standards Benchmarking, APQC has measured 11,243 separate organisations against a single shared definition of what the finance function is, and 6,432 against one definition of procuring materials and services. You cannot benchmark eleven thousand companies against one process definition unless it genuinely is the same process in all eleven thousand.
The academic evidence points the same way. When researchers collected the master's admissions process from nine different universities, each documented independently by people who had never coordinated, domain experts were able to identify 202 one-to-one correspondences between activities across the model pairs. The same process really does turn up in the same shape, organisation after organisation.
Companies already behave as though this were true. In Deloitte's global shared services research, the top reason organisations centralise back-office work is not cost reduction. It is standardising the process, named by 88% of respondents in 2021 against 84% for cost, and still first in 2025 at 72% against 61%. And there is peer-reviewed evidence that the convergence is strongest exactly where you would expect: a classic study of enterprise system implementations found notably fewer mismatches in finance and accounting, where common standards already apply, than in specialised operational areas.
Your supplier insurance check is, give or take a field, everyone's supplier insurance check. Your claims intake looks a great deal like the next company's. That is not a criticism of anyone's operation. It is what makes the tail addressable at all.
Which is why the answer is not to build one
The instinct, once a team accepts that a small process is eating a day a week, is to get it built. That instinct is expensive, and there is now good evidence for how expensive.
Researchers at Oxford analysed 11,011 projects across 23 project types worth 4.64 trillion dollars and found that software is the only category where cost risk is mathematically unbounded. Not merely riskier. Unbounded, meaning the average overrun does not converge and the tail cannot be sized in advance. They named the cause as bespokeness, and the antidote as modularity. In their earlier dataset the single worst outcome was a small customisation budgeted at fifteen hundred dollars that finished at four hundred and twenty-five thousand.
A process that costs a company forty thousand dollars a year in scattered attention does not survive a build with that risk profile attached. It never has. That is why the tail is still there.
Modularity is the interesting half of that finding, because it points at the actual answer. If a process is genuinely the same across thousands of companies, it should be built once and adapted, not rebuilt from nothing every time somebody notices the problem.
What changes when the work is already done
This is why Floowed ships a library rather than an implementation programme. You choose a flow that already exists, adapt it to your systems and your rules, and it runs every case from then on.
What runs is the same five phases every time. It gathers what the case needs from your systems, from external sources like registries and screening providers, and from documents in whatever state they arrive, including handwritten, scanned and photographed. It interprets what those inputs actually mean rather than only what they say. It goes back to whoever still holds a missing piece, reads the reply, and works out whether it answers the question. It runs every gate your rules require. Then it acts on what was settled.
That fourth input, people, is the one that kept these processes human for so long. A system that cannot go and get an answer nobody has written down cannot finish a decision. It can only hand the case back.
What lands is a recommendation with the reasons and the source behind every check, not a verdict. The Decision Engine applies the rules you approved, the same way on every case, with the working shown. You decide how much of it decides, and you widen that as the results come back right.
The commercial consequence matters as much as the technical one. Because the flow is already built, you are not paying for a build, and you are not waiting a quarter for a consulting programme to finish before anything runs. Complex work still exists. A custom policy, an industry-specific set of checks, a real integration build: that is a project and it is priced as one. It should be the exception rather than the front door.
What this is not for
The bound matters, because breadth with no edges reads as no product at all.
This is not for decisions where two experienced people would reasonably disagree and both be right. It is not for the one-off judgment that happens twice a year. It is not for anything that has to resolve in milliseconds. And it is not for a process whose rule nobody can write down: if your team cannot articulate the policy, no system can run it, and the honest answer is to go and agree the policy first.
What it is for is narrower and much more common than it sounds. If you do it more than once, it belongs here.
The tail is where the work actually is
Big decisions get automated because they are business critical and somebody owns a budget for them. Underwriting. Claims adjudication. Customer onboarding. Those will keep getting attention, and they should.
Everything underneath has been left alone for a straightforward economic reason that no longer holds. It was never worth building each one from scratch. It was never possible to read what arrived or to chase what was missing. Both of those have changed, and the second one changed only recently.
Nobody has counted what that tail costs. Our guess is that when somebody finally does, the number will look a lot like the ones from health care administration and the regulatory burden studies: uncomfortably large, obvious in hindsight, and made up almost entirely of work nobody ever decided to do.