The Twenty Percent Is the Business
Co-Founder & CEO of CLRT
In a release published on 19 September, PureHealth announced an AI system, built by its subsidiary PureCS, that will process documents for Daman, the health insurer its release calls The National Insurance Company. The scale is real: more than 60 million documents a year, serving more than three million members. The system is expected to convert more than 80 percent of those documents into structured digital data, manual rework is set to fall by more than half, and processing, the release says, goes from days to minutes. Most readers will stop at the 80. The more useful number is the one the announcement never prints. On its own round figures, as many as 12 million documents a year may not come out the other side as clean data. Someone still handles them, and that someone is where this business now lives.
Start with the arithmetic, because the release invites a reading it does not support. Eighty percent of sixty million leaves up to a fifth, on the round figures about twelve million documents a year, roughly a million a month, that the system is not expected to structure on its own. That is our arithmetic on the release's round numbers, not a published figure. What matters more than its size is its composition. An extraction system clears the documents that look like the documents it has seen: the standard form, the clean template, the legible scan. What it hands back is everything else. The release itself says the system flags missing or incomplete information at the point of submission. Those flagged documents, the handwritten referral, the endorsement that contradicts the policy, the claim with a field nobody can read, are not a random fifth. They are, almost by construction, the fifth where something is unusual, and in an insurer something unusual is where the money and the disputes are.
Notice the grammar of the launch as well. The 80 percent is expected. The halving of rework is set to happen. These are forecasts made on day one, which is normal and honest, but every one of them measures the documents that flow. The phrase days to minutes carries no figure at all. The document that does not flow keeps its days. The release also claims enhanced accuracy without a number, and it does not mention any human review of what the system produces. It does say what the freed time is for: fraud detection, clinical quality assurance and member support. That is the right instinct, and it contains a quiet admission. The valuable work at an insurer was never keying in clean forms. It was always judging the documents that do not fit, and the automation has just made that work the whole of the human job.
Here is the second-order effect most automation business cases miss. Remove the easy fifth of anything and the remaining human queue gets harder; remove the easy four fifths and the queue becomes nothing but exceptions. A team sized and measured on the old mix of documents now faces a stream in which every item needs judgment, and its throughput falls exactly when the dashboard is celebrating. The risk also has a second home that no queue reveals. A field misread inside the automated 80 percent does not land in the residue. It flows on, structured and confident, into the systems downstream, and the release presents it as the foundation for future claims and prior authorisation processing. So the exposure lives in two places at once: the visible residue that people handle, and the invisible errors inside the share that nobody is looking at because it was declared done.
The regulator has already said where it will look. The Central Bank of the UAE issued a guidance note on AI and machine learning in February, and its scope explicitly names insurance providers. It defines a high-impact decision as one that materially affects a customer's access to services, and gives an insurance claim among its examples. It allows AI to run without direct human involvement only for low-risk, non-material processes. It expects customers to be able to request human review or an explanation of an AI-generated decision. It expects testing at least once a year and whenever a model is upgraded or materially changed, and it expects models built or hosted by third parties to sit in the same inventory and meet the same standards as those built in-house. None of this is about how fast the 80 percent moves. Every expectation is about the boundary, and about who answers for what crosses it.
The lesson travels well beyond one insurer. An automation rate is the wrong headline for any regulated business, because it describes the part of the work that was already becoming a commodity. The defensible questions are all about the remainder. What kind of document lands there, and does the mix shift as the system learns? Who owns it, how fast must it move, and what evidence does each decision leave behind? What is the error rate inside the share declared done, and who would notice if it doubled? A business that can answer those has designed an operation. A business that can only quote its rate has bought a machine and handed its hardest work, unannounced, to whoever happens to be holding the queue. The rate is what the vendor sells. The remainder is what the insurer runs.
The eighty percent will look the same at every insurer that buys it. The twenty percent is the business.
A deeper dive
The mechanism underneath the headline is a threshold, and that is what makes the rate slippery. A document system does not simply know which documents it has read correctly; it scores its own confidence and routes everything below a chosen line to a person. Move the line down and the automated share rises, the residue shrinks, and more misreads pass silently into the structured data. Move it up and the residue grows, the errors fall, and the rate on the slide looks worse. An automation percentage is therefore not a property of the technology. It is a setting, one point on a trade-off between the work handed to people and the errors handed to the business, and quoting the rate without the error that comes with it says almost nothing. The setting also drifts. Document formats change, new providers submit in new layouts, policy wording is revised, and a threshold that was sound at launch quietly stops meaning what it meant. Choosing where the line sits, defending that choice to a regulator, and watching it move is judgment work that no vendor can do on the insurer's behalf, because only the insurer carries the cost of each kind of mistake.
The trap that follows is slower and more expensive. The residue is the most valuable data the insurer now produces: every correction a reviewer makes is a labelled example of exactly where the machine fails. Captured deliberately, it is how the system improves at its edge and how the insurer proves to an examiner that the boundary is governed rather than assumed. Left as an operations backlog, cleared as fast as possible and never fed back, it is waste, and the system stays weakest exactly where the claims and the complaints are. Meanwhile the people working that queue lose the easy cases that once trained them, so their judgment thins just as it becomes the whole of their job. The release already lists complex claims among the requests the system streamlines, so the boundary touches the decisions the guidance calls high impact today, and the same foundation is meant to carry future claims and prior authorisation processing. That is where an insurer discovers whether it designed the twenty percent or merely inherited it.
Work with CLRT
This is the work CLRT does before anything is switched on: we map what an automation will hand back, size the residue in real cases rather than percentages, and design the routing, review and evidence that let a regulated business defend its boundary to customers and examiners alike. The machine that clears the easy majority is increasingly a commodity. Knowing where to draw the line, and what to do with everything on the far side of it, is not. Start with the CLRT Ascent diagnostic at ascent.clrtstudio.com, or talk to CLRT directly about the residue in your own operation.

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.


