AI gets a highlighter
A new kind of AI finds the answer in your document and highlights it, instead of writing one. ARBI Docs already works this way. Here is how the two compare.
November 13, 2026
The short answer
When a lawyer asks AI about a document, the first question is always the same: where does it say that? An answer you cannot check is not much use.
A new AI built around that idea has just arrived. It does not write answers. It finds them in the document, highlights them, and says how sure it is. It is called Clairic, and it comes from Cloudflare, a large internet company whose services run behind many of the websites you use.
We welcome it. It is a useful new kind of tool for anyone building legal AI. It is also how ARBI Docs already works: every answer comes with the passage it came from, one click away, highlighted in the original document.
So we tested the two side by side on 53 documents where we knew every answer.*
- ARBI came out ahead: 109 of 127 answers right, against 105 for Clairic.
- Clairic is fast: about a quarter of a second for a short document.
- Clairic says how sure it is, and when it said it was very sure, it was almost always right.
* We tested the early version of Clairic that Cloudflare hosts. When Cloudflare releases the model itself, we will run it on our own systems and report back.
Writer or highlighter?
Picture two assistants reviewing a contract.
The writer reads it and types the answer: "The lease ends on 28 February 2030." A writer can also summarise and explain. But you have to trust what they typed, or go back and check.
The highlighter reads it and marks the date on the page. A highlighter cannot write a summary. But you can check their work at a glance.
Most AI tools are writers. Clairic is a highlighter. ARBI Docs is a writer that has to show its highlight: it can summarise and explain, but every answer must point to the passage behind it.
How ARBI Docs keeps every answer checkable
None of this needs a new kind of model. It is a combination of well-established techniques, applied strictly.
- Every document is read first. ARBI reads each document, including scans, and divides it into passages it can point to. The original file is never changed.
- Every answer names its source. Whether you ask a question in a conversation or fill in a review table across hundreds of documents, each answer cites the passages it relied on.
- No source, no answer. In a review table, if an answer cannot point to a real passage, ARBI shows "Not found" instead. In our own measurements, correct answers could always name their source, while most invented ones could not, so this one rule stops most of them before you see them.
- One click to check. Click any citation and the original document opens at that passage, highlighted.
- It can run on your own hardware. ARBI can run on our own serving software, so your documents can stay on servers you control.
What Clairic is, and what it isn't
Cloudflare announced Clairic on 19 October 2026 and was frank about its limits. Here is each claim made for it, and how far it holds up.
“It is a new kind of AI.”
“It cannot make things up.”
“It shows you where the answer is.”
“It tells you how sure it is.”
“It understands dates.”
“It is fast and cheap.”
“You can run it yourself.”
Head to head
We wrote 53 documents ourselves, so we knew every right answer: invoices, leases, contracts and receipts, scanned forms, documents in French, German and Japanese, and a 150-page hearing transcript. Some questions were traps, with no answer in the document at all. Both systems got the same 127 questions.
| What we asked | ARBI Docs | Clairic |
|---|---|---|
| Everyday facts from invoices and contracts | 46 of 46 | 44 of 46 |
| Dates, amounts and estimates | 16 of 26 | 18 of 26 |
| Spotting that the answer is not there | 4 of 5 | 5 of 5 |
| Lists and judgement calls | 10 of 10 | 6 of 10 |
| Long documents | 8 of 8 | 6 of 8 |
| Scanned documents | 25 of 32 | 26 of 32 |
| Total | 109 of 127 | 105 of 127 |
ARBI was better at long documents, at judgement calls such as picking out complaints, and at reading reference numbers from scans. On a second run, ARBI scored 29 of 32 on scans; one scan had not finished the first time.
Clairic was better at dates written without a year, because it can be told the date of the document, and at saying "not in the document". Once, ARBI filled in a missing purchase order number with the invoice number.
A welcome new tool
Clairic gets several things right that are worth paying attention to:
- Speed. A highlighter that answers in a quarter of a second makes new products possible, such as checking every incoming invoice or contract as it arrives.
- Honest confidence. A reliable "how sure am I" lets software accept the easy answers and send the doubtful ones to a person.
- Knowing the date. Telling the AI the date of a letter is a simple idea, and one worth borrowing.
We see Clairic as a useful new category for people building legal AI, alongside the citation-first approach ARBI Docs already uses. When Cloudflare releases the model, we will run it on our own systems and see how the two can work together.
Questions
▸Does ARBI Docs ever get it wrong?
Yes, occasionally, which is why every answer comes with its source. A citation lets you check any answer in one click, and in a review table an answer without a source is shown as "Not found" rather than as a fact.
▸Does my document leave my systems?
With ARBI, it does not have to: ARBI can run on hardware you control. Clairic today runs only on Cloudflare's service, which says it keeps no data.
▸What does Clairic cost?
$0.10 for every million tokens it reads (a token is roughly part of a word). That is well under a cent for a 100-page document.
Sources: Cloudflare's announcement, "Introducing Clairic, a new System One class of extraction models" (Cloudflare blog, 19 October 2026), and its Workers AI changelog entry of the same date · Gabriel Anhaia, "Cloudflare's Clairic writes no output tokens" (DEV Community) · the hearing transcript is the public record of the House Committee on Oversight and Reform hearing of 27 February 2019 (govinfo, CHRG-116hhrg35230). © 2026 Arbitration City Ltd.