If you work in legal marketing or business development, the phrase "Artificial Intelligence" has probably worn out its welcome. Just when you figured out what a Large Language Model (LLM) is, and just when your firm started rolling out Copilot licenses in Outlook and Word so people could use AI without citing fake court cases, a new batch of jargon appeared in your LinkedIn feed: "System One models," "decision models," and "extraction models."
When you have a pitch deck due at 4:00 PM, three partners asking where their conference leads are, and an executive committee asking whether those $30-a-month Copilot licenses are paying off, you don't have time for a computer science lecture. Does this help your firm win business and serve clients, or can you ignore it?
It helps, and the idea behind it is simpler than the jargon. These new models are built for the routine jobs where chatbots are slow, expensive, and unpredictable.
On October 19, Cloudflare released a preview of Clairic, which it calls the first "System One" extraction model. It joins a newer category called decision models, which includes TypeSafe AI's Jev and Cloudflare's own Clef. Together they change how much routine work you hand to a chatbot like Copilot and how often you have to double-check what comes back.
To see why, start with a high school history assignment.
Imagine you are back in 10th-grade history class. Your teacher hands you a three-page article about a famous train robbery and a worksheet with three blanks:
- Who robbed the train?
- What date did it happen?
- How much money was stolen?
Now imagine you hire a tutor to complete the worksheet.
If the tutor is a standard AI chatbot (an LLM), you have hired an over-thinker who insists on writing an essay, one word at a time, to answer a fill-in-the-blank worksheet.
It reads the pages, clears its throat, and starts typing: "Based on the historical text provided, the individual responsible for the robbery was..."
That takes several seconds. Do it across thousands of documents and the computing bill adds up. Worse, because LLMs are built to be writers and conversationalists, they sometimes try too hard to be "helpful." If the article never says how much money was stolen, a chatbot might guess or make up a number (what tech people call a "hallucination") so it doesn't leave your worksheet blank.
For the last couple of years, businesses, including law firms and legal tech vendors, have used these slow, expensive "essay writers" for basic administrative work. We have been asking Copilot or ChatGPT to pull invoice totals off bills, read intake forms, and type business cards into Excel.
The new models take a different approach, borrowed from how people think.
Fast Brain vs. Slow Brain ("System 1" vs. "System 2")
Daniel Kahneman's Thinking, Fast and Slow describes two ways we think:
- System 2 (Your Slow Brain): If someone asks you to multiply 17 x 24 in your head, you have to stop and concentrate as you work through it step by step. The same goes for writing a client pitch in Word or drafting a sensitive email in Outlook. Chatbots like Copilot are "System 2" AIs.
- System 1 (Your Fast Brain): If someone throws a baseball at your head, you duck. When you pull up to a red octagon, you register "STOP" instantly.
Until recently, there were few good "System 1" AI tools for everyday business workflows. Now there are two kinds of "Fast Brain" models, each with one narrow job.
Decision Models (The Multiple-Choice Pickers)
TypeSafe started this category with a model called Jev. Cloudflare followed with Clef and Clef-flash, its own open-source decision models, released October 1.
Think of a decision model as a very fast multiple-choice test taker. It cannot write an email in Outlook or draft a memo in Word.
You hand it a piece of information and a strict list of options (A, B, C, or D), and in a fraction of a second it picks one and gives you the probability that its pick is right. It can't invent an "Option E," because it only answers from the list you gave it.
Extraction Models (The Fact Finders)
That brings us to Clairic. It works like a highlighter and a pocket calculator.
Why couldn't Jev or Clef do that job? You can't turn every question into multiple choice. If you ask, "What is the total dollar amount on this hospital bill?" or "What date did this accident happen?", you can't list every possible number or date as an option in advance.
Cloudflare built Clairic for questions like these. You give it an unstructured email, a PDF, or up to four smartphone photos, plus a blank form to fill out. It typically answers in under a second and handles four kinds of blanks in the same pass:
- Extracts text: It pulls the exact name, company, or order number from the document, so a person can check it against the source.
- Computes math: If an email says, "I paid $5,000 in May and $3,500 in June," and you ask for the total, Clairic returns $8,500 in the same pass. If an email mentions two dates, it can calculate how many days passed between them.
- Estimates values: If an exact dollar amount isn't written down, it uses clues in the text to give you a realistic range.
- Forecasts numbers: Give it a history of past numbers and it projects the next values, each with a likely range.
Because Clairic doesn't generate text, it can't invent a sentence the way a chatbot can. It can still make mistakes, so every answer comes with a confidence score.
If you ask it for a Police Report Number and the client never mentioned one, Clairic returns null (blank) instead of guessing. And because it reads the whole page in one pass, Cloudflare says asking for 30 fields costs about the same as asking for three.
How the "AI Relay Team" Works in a Law Firm
These models make Copilot and the other chatbots more useful.
If your firm runs on Microsoft 365, Copilot is already on your lawyers' and marketers' desktops. On October 15, Microsoft made more than 20 legal skills and partner plugins generally available inside Copilot.
The problem: ask Copilot alone to read 200 blurry photos of business cards, or to calculate unbilled hours across disorganized intake emails, and it gets slow and clunky. It also sometimes gets numbers wrong.
A better setup splits the work like a three-person relay team:
- Runner 1, the Extraction Model (Clairic): Reads the disorganized documents or photos and pulls out clean facts and numbers in under a second.
- Runner 2, the Decision Model (Jev or Clef): Looks at those facts and picks the right category or next step from your firm's dropdown menus in under a second.
- Runner 3, the LLM (Copilot, ChatGPT, Claude, or Gemini): Takes the baton once the facts and decisions are locked in. Then it does what it does best: writing the email, pitch, or summary inside Microsoft 365.
Here are two examples, both with Copilot running the last leg.
Example 1: Legal Marketing & Business Development (The Post-Conference Pileup)
The Scenario: Five of your firm's partners just got back from a three-day industry conference. On Monday morning, October 19, they dump 200 photos of business cards into a shared folder, along with a jumbled list of people who scanned their badges at the firm's booth. Half the cards have handwriting scribbled on the back from cocktail receptions.
One blurry smartphone photo shows a business card for "Sarah Jenkins, General Counsel at FinTech Corp." On the back, a partner scribbled in pen: "Met Oct 18. Expanding to Europe in Q1. Worried about GDPR fines. Currently paying BigLaw LLP $85k/mo on retainer and unhappy with slow replies. Follow up before Nov 1!"
Normally, a marketing or BD coordinator spends two days on this work. That means squinting at handwriting while typing 200 cards into the CRM or an Excel tracker, then figuring out who is a hot lead and nagging partners to write follow-up emails in Outlook before the trail goes cold.
Here is how the relay team handles the same 200 cards:
Step 1: Clairic (The Extraction Model) Reads the Photos
Instead of a person typing, or an LLM slowly reading 200 images one at a time, Clairic reads the front and back of Sarah's card in under a second and fills in your CRM or Excel columns:
- Extracted Text: Name = "Sarah Jenkins", Title = "General Counsel", Company = "FinTech Corp", Current Counsel = "BigLaw LLP", Key Pain Point = "GDPR fines / European expansion".
- Computed Math: Clairic sees "$85k/mo" on the card and calculates Annualized Legal Spend = $1,020,000. It sees "Follow up before Nov 1," compares it to today's date (Oct 19) and fills in Days Left to Follow Up = 13. ⚬ Estimated Value: Based on the company clues and current spend, Clairic estimates First-Year Pitch Value = $250,000 to $600,000.
- Confidence Score: What if one word of the partner's handwriting is a mess? Clairic attaches a confidence score to every blank. If it is 99% sure the card says "$85k," it moves on. If it is only 60% sure because of a coffee stain, it flags that one box for your BD coordinator to check.
Step 2: Jev or Clef (The Decision Model) Grades and Routes the Leads
With all 200 cards turned into clean facts and numbers, a decision model (like Clef or Jev) answers two multiple-choice questions your BD team set up in advance:
- Question 1: What is the BD action for this contact? (Your firm's preset options: A. Immediate 1-on-1 Partner Pitch | B. Invite to VIP Client Dinner | C. Add to General Firm Newsletter | D. Ignore / Vendor Trying to Sell Us Something). Because Sarah is a General Counsel spending more than $1M a year with a pressing Q1 need, the model picks Option A (97% probability).
- Question 2: Which firm pitch deck or one-pager in SharePoint matches this lead? (Available choices: M&A Overview, US Employment Law, EU Data Privacy & GDPR, IP Litigation). It picks EU Data Privacy & GDPR (99% probability).
Step 3: Microsoft Copilot (The LLM Writer) Drafts the VIP Pitch in Outlook and Word
Of those 200 cards, probably 150 belong to vendors, law students, or people who stopped by the booth for a free pen. You don't want to clutter a partner's inbox by having Copilot write 200 custom pitches.
So the workflow only calls Copilot for the 15 "Partner Pitch" leads the decision model flagged. Because Copilot lives inside your Microsoft 365 environment, it takes the verified facts from Clairic, pulls the EU Data Privacy & GDPR pitch document from SharePoint that the decision model selected, customizes the cover page in Word with FinTech Corp's details, and drops a ready-to-send email into the partner's Outlook Drafts folder:
"Hi Sarah, it was great speaking with you on Sunday. Best of luck as you gear up for the European expansion in Q1. I know keeping GDPR compliance from slowing down a launch is a major headache. I attached a quick two-page overview of how our team handles EU rollouts. Do you have 15 minutes next Tuesday on Teams to compare notes?"
By 9:05 AM on Monday, the CRM is updated, the 185 lower-priority contacts are sorted into newsletter lists, and the partner opens Outlook, reviews 15 pre-written emails with the right attachments, and hits "Send."
Example 2: The Practice Side (New Client Intake & Triage)
The same relay works on the practice side. Client intake is a good test, because speed and accuracy decide whether you sign a strong matter or lose it to the firm down the street.
The Scenario: At 11:00 PM on a Tuesday, a potential client fills out the "Contact Us" form on your firm's website. They attach two crooked photos of hospital bills sitting on their kitchen counter.
"Hi, I got rear-ended on I-95 by a delivery truck from Apex Logistics on Oct 2, 2026. My neck is messed up. The first ER visit was $4,200 and the follow-up MRI was $1,850. My insurance won't cover all of it and I've missed two weeks of work. Can someone call me?"
Here is what the relay team does while your intake staff is asleep:
Step 1: Clairic (The Extraction Model) Pulls the Case Facts
In under a second, Clairic reads the website message and the two photos of the hospital bills. Here is what it returns:
- Extracted Text: Adverse Party = "Apex Logistics" (pulled straight from the message, so your conflicts software can run a check right away).
- Extracted Date: Incident Date = 2026-10-02.
- Computed Math: Clairic adds the ER bill and the MRI bill: Total Medical Bills So Far = $6,050. It also subtracts Oct 2 from today's date (Oct 20) to fill in Days Since Incident = 18, which is important for preserving the truck's black-box data and tracking statutory deadlines.
- Estimated Value: Based on the commercial fleet involvement, the ER visit, the MRI, and two weeks of lost wages, Clairic estimates Potential Matter Value = $45,000 to $120,000.
- The "IDK": Your intake form has a box for Police Report Number. A chatbot might grab an account number off the top of the hospital bill and put it in that box. Clairic sees no police report mentioned and marks Police Report Number = null (blank).
Step 2: Jev or Clef (The Decision Model) Triages the File
Your decision model reads Clairic's worksheet and answers two routing questions:
- Question 1: Which practice group handles this? (Options: Family Law, Corporate Disputes, Personal Injury - Standard Auto, Personal Injury - Commercial Trucking). It picks Personal Injury - Commercial Trucking (99% probability).
- Question 2: How urgent is this intake? (Options: High Priority / Alert On-Call Team, Standard Morning Queue, Reject / Refer Out). Because a commercial logistics company is involved and the medical bills already top $6,000, it picks High Priority (94% probability).
Step 3: Microsoft Copilot (The LLM Writer) Preps Outlook, Teams, and Word
Because the decision model flagged the file as High Priority, the workflow calls Copilot inside the firm's Microsoft 365 workspace for three writing jobs:
- In Outlook: Copilot drafts a warm reply telling the injured person that the Commercial Trucking team has their file and asks: "When we speak tomorrow morning, if you happen to have the Police Report Number handy, that will help us move even faster." Copilot knows to ask because Clairic left that blank as null.
- In Microsoft Teams: Copilot posts a three-bullet summary in the Commercial Trucking practice group's channel and tags the partner on call, so they see a high-value matter waiting before they pour their morning coffee.
- In Microsoft Word: Using the firm's approved SharePoint templates, Copilot pre-fills a draft Engagement Letter with the client's name, the incident date (2026-10-02), and the adverse party (Apex Logistics), so it is ready when the partner finishes the morning call.
Why the Split Matters for CMOs and BD Leaders
If you are in a budget meeting, or sitting across from a legal tech vendor pitching their "latest AI software," you don't need to know how to code. Splitting these jobs affects your firm's money, risk, and time.
You Don't Have to Be an Engineer to Be Smart About AI
When a new tool like Clairic, Jev, or Clef hits the news right after another round of Copilot updates, it can feel like the goalposts moved again.
This time the change works in your favor. AI is maturing in a way that makes it more predictable for law firms.
At first, everyone tried to use one giant, unpredictable tool (the LLM chatbot) for administrative tasks and writing. In a profession where a wrong date, a miscalculated invoice, or a made-up number can damage your reputation, that made people rightly nervous.
Now the industry is building specialized tools that stay in their lane:
Extraction models (like Clairic) pull the facts and leave a blank when the information isn't there.
Decision models (like Jev and Clef) pick the right box on the multiple-choice test.
LLMs (like Copilot, ChatGPT, Claude, and Gemini) write the prose inside the Word and Outlook tools your team already uses every day.
If you work in legal marketing or BD, you will probably never configure Clairic yourself. These models are more likely to show up inside the CRM and intake software your firm already buys.
What you need is the relay race in your head. The next time a vendor says their tool "uses AI to read your RFPs, sort your conference leads, or triage client intake," ask one question:
"Are you using a slow, expensive chatbot to guess our numbers, or an extraction model to get the facts right before handing them to Copilot?"
If the answer is "the chatbot does everything," ask what it puts in a field when the information isn't there.