Technical white paper · September 2026
From Answers to Records
Every form your AI output takes today fails in the same place. The chat answer is gone when the window closes. The generated document looks like a record and is a snapshot severed from its source. The emailed digest loses its permission context in the first forward. The dashboard is persistent, governed and live, and it carries no claim at all. None of these fail for reasons formatting can fix. This paper defines what has to replace them, using a definition of “record” that has existed for decades, and then shows what the replacement refused to say in one real scheduled run.
Ali Yıldırım — Founder and CEO, D-CAT Group. Twenty-five years in enterprise business intelligence. Computer Engineering, Boğaziçi University; MBA, Istanbul Bilgi University.
The case
The argument, in four sentences
- Observability tells you what the agent did. Auditability lets you reconstruct it. Neither tells you whether you can stand behind the claim it produced, which is the thing a board asks about, years later.
- Five properties decide that: identity, address, authority, type and lifecycle. Most stacks hold some of them on the container, applied at read. The question is whether any of it travels with the number you acted on.
- Records management answered this question decades before agents existed. The paper takes that definition and applies it to AI output without softening it.
- Then it stops arguing and opens the database: one scheduled run, every gate outcome, every refusal with its recorded reason.
The part a demo leaves out
What a demo will not show you
Any vendor can show you a good briefing. The interesting record is the one the system refused to produce. One scheduled run, nobody asked a question: it began at 05:00, finished 285 seconds later, and cost 78 cents of model spend, frozen at the price in effect that day. Thirteen candidate records reached the database. Eleven published. Two were held, each with a reason still readable in the Guard Ledger. And once, mid-run, the engine sent the agent’s entire draft answer back. This is what it wrote:
[engine] Structural rule (numeric provenance): every figure in a final answer is exactly one of three things. ADDRESSED (declared in the claims block with the ref you read it from), COMPUTED (declared with op and operands that are addresses), or STRUCTURE the engine or the user put there. Nothing else is proven. A value that merely EXISTS somewhere in the evidence is not proven until you name its address. Your answer was NOT delivered.
Fourteen figures in that draft had no address. Across the whole operating record, nine of the eleven held objects were the same family of sentence: fastest-growing, weakest, highest. The class of claim the gate catches most often is exactly the class an executive is most likely to act on.
What is inside
Section by section
- Why a chat answer, a generated document, an emailed digest and a dashboard all fail in the same place, and why none of it is a formatting problem
- The five properties that decide whether a delivered number can be relied on later, read against all four of those forms
- The definition of a record that records management settled decades ago, and what happens when you apply it to AI output without softening it
- Observability, auditability and defensibility: three properties sold under one word, and which one your vendor is actually selling
- How a number earns its place: envelope, row, cell and a check digit, and what happens to a figure that cannot name its address
- The eight checks a candidate passes before it becomes a record, in the order they run, none of them involving a model
- One real scheduled run, opened in full: what published, what was held, and the draft answer the engine sent back
- What the operating record shows, what it cannot show, and four limits stated before a CIO has to ask
- What is shipping, what is next, and four states that exist in the schema but have never produced a live record
- Appendix A: twelve questions for every vendor in this category, written to be used against us as well
Before you spend the time
Read it if. Skip it if.
Read it if
- You are evaluating more than one product in this category and the demos have started to look alike.
- You have to sign something that says an AI-produced number was defensible.
- Your audit, risk or legal function has already asked where a number came from and the answer took more than one click.
- You run the data platform the agents would sit on top of.
Skip it if
- You want a market overview. This paper describes one architecture and argues for it.
- You want customer results. There are none in it. Every number in it is the system’s own operating record, and it says so on every page that carries one.
Stated before you ask
What this paper does not claim
- It does not claim zero error. It describes a mechanism and the behaviour of that mechanism when it refuses.
- It does not claim a certification. ISO 15489-1 is a concepts-and-principles standard, not a certifiable one; the certifiable records management standard is ISO 30301. We hold neither. We designed against the definition and we say so in the paper.
- It does not report a customer result. Every figure comes from the product’s own database: 16 discovery runs, 3 agents, three and a half weeks.
- It does not measure reading. Every published brief was opened at least once, and that is all the record supports. There is no open rate, because there is no denominator yet.
Any vendor can show you a good briefing. Ask to see what the system refused to say, and why.