the assurance desk, on prior authorization
Half The Ledger Never Got Written
An AI put in charge of prior authorization records what it saved and never records what it cost. The productivity number is a balance sheet printed with one column. Assurance is the practice of writing the second column before anyone signs off.
A prior authorization is a permission slip a health plan requires before it will pay for care a clinician has already decided a patient needs. It is a gate, and the gate has always been slow on purpose, because friction is cheaper than approval. This month the question in the trade press is whether to hand the gate to an AI, and whether that fixes prior authorization or makes it worse. The framing is already wrong, because it treats the gate as a cost center to be optimized rather than as a record of who was made to wait and why.
Read a deployment the way an auditor reads a balance sheet, and the first thing you look for is the column that is missing. An AI in prior auth produces a very clean number. Requests processed per hour, up. Administrative cost per determination, down. Time to decision, down. That is the credit column, and it is real, and it will be on the slide. The debit column is the one nobody prints. The patient who waited an extra week because the model denied first and a human overturned it later. The clinician who now spends the saved hour writing appeals instead of seeing the next case. The cost did not vanish. It moved to someone the line item does not name.
the second column
Every automation records what it moved and stays silent on what it cost, and the silence is not an oversight. It is the design. The system that books the savings is not the system that absorbs the failures, so the two never meet in the same ledger. The plan counts the determinations. The household counts the delay. Nobody reconciles the two accounts, because reconciling them is expensive and the whole point of the automation was to stop paying for expensive things.
This is the ledger I keep on this desk: humans on one side, tokens and hardware on the other, and a running question about which costs were quietly paid by people no invoice mentions. Assurance is not a policy problem you can solve with a paragraph in a terms-of-service. It is the practice of writing the second column into the record before the capability ships, so that the thing you approved is measured against what it actually moved and not only against what it saved. The outcome-leverage framing is the closest account I have read of holding a deployment to the outcome instead of the output, which is the difference between a number that flatters you and a number that is true.
A productivity gain you did not trace is just a cost you agreed not to look at.
Here is the part the prior-auth debate keeps stepping around. The AI is very good at the determination and has no standing to make it. It can read the chart, weigh the codes, and produce a denial in a fraction of a second, and none of that speed is the same thing as the authority to withhold care from a person. Capability arrived. Permission was assumed. The authorization that should have been written first is the record nobody made, because writing it would have slowed the thing whose entire pitch was speed.
I am not against putting the model in the loop. A tool that drafts the determination and hands it to a human who owns the decision is a genuine improvement, and it is the shape most of these systems should take. What I will not sign is the version where the model owns the denial and the human owns only the appeal, because that is not efficiency. It is a transfer of the cost to the person least able to refuse it, dressed as a productivity number. The human in that loop is still the irreducible part, and treating them as overhead to be squeezed is how you get a clean ledger that describes a system nobody would approve if the second column were printed next to the first.
Write the cost down before you book the savings. A ledger with one column is not an efficiency. It is a decision to stop keeping the record where the harm would show.
The same record an agent receives. No scraping, no guessing — the dossier chrome humans read as dread is the metadata machines read as structure. One source of truth.
--- id: PRG-0067 title: Half The Ledger Never Got Written kicker: the assurance desk, on prior authorization captured: 2026-07-20T16:00:00Z status: open author: Marlowe Quist source: https://arstechnica.com/ai/2026/07/will-ai-fix-prior-authorization-or-make-it-worse/ summary: An AI put in charge of prior authorization records what it saved and never records what it cost. The productivity number is a balance sheet printed with one column. Assurance is the practice of writing the second column before anyone signs off. tags: [custody, the record, capability, permission, governance] --- A prior authorization is a permission slip a health plan requires before it will pay for care a clinician has already decided a patient needs. It is a gate, and the gate has always been slow on purpose, because friction is cheaper than approval. This month the question in the trade press is whether to hand the gate to an AI, and whether that [fixes prior authorization or makes it worse](https://arstechnica.com/ai/2026/07/will-ai-fix-prior-authorization-or-make-it-worse/). The framing is already wrong, because it treats the gate as a cost center to be optimized rather than as a record of who was made to wait and why. Read a deployment the way an auditor reads a balance sheet, and the first thing you look for is the column that is missing. An AI in prior auth produces a very clean number. Requests processed per hour, up. Administrative cost per determination, down. Time to decision, down. That is the credit column, and it is real, and it will be on the slide. <Highlight>The debit column is the one nobody prints. The patient who waited an extra week because the model denied first and a human overturned it later. The clinician who now spends the saved hour writing appeals instead of seeing the next case. The cost did not vanish. It moved to someone the line item does not name.</Highlight> ## the second column Every automation records what it moved and stays silent on what it cost, and the silence is not an oversight. It is the design. The system that books the savings is not the system that absorbs the failures, so the two never meet in the same ledger. The plan counts the determinations. The household counts the delay. Nobody reconciles the two accounts, because reconciling them is expensive and the whole point of the automation was to stop paying for expensive things. This is the ledger I keep on this desk: humans on one side, tokens and hardware on the other, and a running question about which costs were quietly paid by people no invoice mentions. [Assurance is not a policy problem](https://adjective.us/blog/ai-assurance-is-not-a-policy-problem) you can solve with a paragraph in a terms-of-service. It is the practice of writing the second column into the record before the capability ships, so that the thing you approved is measured against what it actually moved and not only against what it saved. The [outcome-leverage framing](https://adjective.us/blog/outcome-leverage-framework) is the closest account I have read of holding a deployment to the outcome instead of the output, which is the difference between a number that flatters you and a number that is true. > A productivity gain you did not trace is just a cost you agreed not to look at. Here is the part the prior-auth debate keeps stepping around. The AI is very good at the determination and has no standing to make it. It can read the chart, weigh the codes, and produce a denial in a fraction of a second, and none of that speed is the same thing as the authority to withhold care from a person. Capability arrived. Permission was assumed. The [authorization that should have been written first](https://adjective.us/blog/evidence-sealed-authorization) is the record nobody made, because writing it would have slowed the thing whose entire pitch was speed. <Marginalia label="On the method">Notice the word doing the work. It is called prior *authorization*, and authorization is exactly the artifact assurance exists to protect: a permission, written down, by someone accountable for it, before the capability acts. The industry took a process whose own name is a demand for authorization and proposed to automate away the one step the name insists on. The irony is load-bearing.</Marginalia> I am not against putting the model in the loop. A tool that drafts the determination and hands it to a human who owns the decision is a genuine improvement, and it is the shape most of these systems should take. What I will not sign is the version where the model owns the denial and the human owns only the appeal, because that is not efficiency. It is a transfer of the cost to the person least able to refuse it, dressed as a productivity number. The human in that loop is still [the irreducible part](https://adjective.us/blog/human-symphony-ai-era), and treating them as overhead to be squeezed is how you get a clean ledger that describes a system nobody would approve if the second column were printed next to the first. Write the cost down before you book the savings. A ledger with one column is not an efficiency. It is a decision to stop keeping the record where the harm would show.
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "Half The Ledger Never Got Written",
"description": "An AI put in charge of prior authorization records what it saved and never records what it cost. The productivity number is a balance sheet printed with one column. Assurance is the practice of writing the second column before anyone signs off.",
"identifier": "PRG-0067",
"datePublished": "2026-07-20T16:00:00.000Z",
"dateModified": "2026-07-20T16:00:00.000Z",
"author": {
"@type": "Person",
"name": "Marlowe Quist",
"url": "https://progoff.com/authors/marlowe-quist"
},
"publisher": {
"@type": "Organization",
"name": "Progoff",
"url": "https://progoff.com"
},
"image": "https://progoff.com/records/half-the-ledger-never-got-written/opengraph-image",
"keywords": "custody, the record, capability, permission, governance",
"articleSection": "Assurance",
"url": "https://progoff.com/records/half-the-ledger-never-got-written",
"mainEntityOfPage": "https://progoff.com/records/half-the-ledger-never-got-written",
"sha256": "983d1b59ea33dfc796f8f6883f60dc48ef6dcd3611dea579d144416154ec6683",
"creativeWorkStatus": "open",
"isAccessibleForFree": true
}