Three weeks ago I ran an AI workshop for a dental clinic in Lisbon. Eleven people around the table in the staff room, coffee, pastéis de nata, and a whiteboard full of questions. Halfway through, I asked the team what took the most time in their week. The head receptionist didn't hesitate. "The inbox." Every morning she spends the first hour and a half sorting the clinic's shared email. Appointment requests, cancellations, insurance paperwork, supplier invoices, newsletters, and now and then a patient in pain. I asked if she could open it on the big screen. It was Monday, and there were 212 unread emails since Friday afternoon. Somewhere in the middle sat a message from Saturday morning. A patient with a swollen jaw was asking if anyone could see her that weekend. Nobody had read it. She had booked with another clinic on Sunday. The room went quiet for a moment. Then the owner turned to me and asked: "Can AI sort this for us?" My honest answer that day was a careful one. Yes, ChatGPT can read every email and put it in a folder. It also writes a paragraph about each one, takes a few seconds per email, and gets expensive at 200 emails a day. It sounds as sure when it's guessing as when it's right, and I didn't want a clinic trusting that with patients in pain. So I told them to start with clear inbox rules and a morning routine, and to come back to AI sorting later. Two weeks later, a new AI model came out that changed my answer. The AI that can't write a sentenceOn 15 September, Diogo Almeida released a model called Jev. He spent four years at OpenAI working on the training behind ChatGPT and GPT-4. His new model cannot write a single sentence. It can't draft an email, answer you in words or explain itself. When I read that, I laughed. The whole industry has spent three years teaching machines to talk more. Almeida told Forbes: "We've been optimizing for humans and we're super human at pleasing humans." The chatbots got very good at sounding helpful. That clinic's inbox needed something quieter. Writing and deciding are two different jobs. As of this month, they have two different tools. Paragraph brain or traffic copChatGPT and Claude are paragraph brains. You ask, they think out loud, and they answer one word at a time. Jev works like a reflex. When a traffic light turns red, you hit the brake. You don't write an essay about the colour red first. The easiest way to think about it is a traffic cop at the front door of your business. An email comes in, Jev looks at it and points. Patient in pain, to the top. Newsletter, to the reading folder. Invoice, to accounting. It decides in a fraction of a second, and it tells you how sure it is. Why the AI world sat upSpeed. Jev answers in 70 to 500 milliseconds. A blink takes about 300. Price. TypeSafe charges $42 per billion input tokens. One creator, Ryan Vogel, sorted 1,700 emails for 18 cents. A full week of that clinic's inbox, around 1,400 emails, would cost about fifteen cents. It admits doubt. This is the part I like most. Every answer comes with a confidence number. A chatbot sounds as sure when it's guessing as when it's right. Jev says "95% sure this is urgent" or "55% sure, please have a look". The sure ones get handled. The unsure ones go to a person. A word of caution. TypeSafe claims Jev is 193 times faster and 444 times cheaper than regular AI models. They tested it on workflows their own team built, and in their own notes they expect real results to land lower. I treat those numbers as a ceiling. That Saturday, with JevThis is how I'd set it up for the clinic. You don't write a prompt. You give Jev the email and a short list of questions with fixed answers.
The patient's email arrived on Saturday at 09:12. A fraction of a second later, Jev would have returned something like this:
The percentages are my own illustration. Your tests will show the real ones. Add one simple rule on top: anything marked as a patient in pain sends a text to the dentist on call. They would have seen it at 09:13 on Saturday, and the patient would have had an appointment before lunch. The newsletters would have gone to a reading folder, the invoices to accounting and the insurance forms to their own pile. On Monday morning the receptionist would have started with twenty emails that needed her, instead of 212. Setting it up in four stepsStep one, get access. Jev is in early access. You reach it through TypeSafe's own console, Vercel's AI Gateway (free until 25 September), Cloudflare or LangChain. I haven't seen a ready-made Zapier or Make module yet, so for now you or your developer call it with a web request step. Step two, write your questions. Decide the answer options in advance. "Is a client reporting something broken?" works better than "What is this email about?" Step three, draw a confidence line. Pick a threshold. Above it, the automation acts. Below it, the email lands in a small review folder. I'd start at 90% and lower it once the results earn your trust. Step four, let the writers write. Jev decides. If an email needs a reply, hand it on to ChatGPT, Claude or a colleague. The expensive model then works on the few emails that need words.
What the first builders foundDevelopers published dozens of tests in the first week. One of them shows the whole idea in two numbers. A builder took 1,018 AI research papers. A regular writing model summarised each one, and Jev sorted every summary into one of 24 topics.
Same pipeline, same papers. The writing cost fifty times more than the deciding. Most businesses pay writing prices for deciding work today, because until last week there was nothing else on the shelf. The clinic's inbox is one use. Three more stood out from launch week, and the figures are the builders' own. 1 / A traffic light for AI writing. If AI drafts your replies or posts, Jev checks each draft before it goes out. Does it answer the question? Are the claims backed up? Green sends, amber waits for a person, red goes back for a rewrite. This answers the worry I hear most from owners: "What if it says something stupid to a customer?" Ready for a small business now. 2 / Lead scoring before the call. Every contact form gets a weak, medium or strong score, and Jev can check whether the reply you're about to send fits that person. For an estate agency with 60 weekend enquiries, Monday starts with the serious buyers on top. Ready for a small business now. 3 / Web agents that click in milliseconds. The Browser Use team built an agent where Jev picks each next click on a website. It searched Google Flights in 7.1 seconds and cut its browser commands from 1,092 to 101. They tested one website, so it's early days, but this is where agents that handle supplier portals and booking systems for you are heading. Needs a developer, and patience. Your Saturday emailMost businesses I work with have a version of that Saturday email. In a hotel, it's the guest whose booking confirmation never arrived. In a transport company, it's the customer whose delivery went to the wrong address. It sits in a shared inbox under forty other messages until someone finds it on Monday. Your team isn't careless. They sort by hand, and on a busy day the important email looks like every other one. I'm also sceptical. Jev is eight days old. The numbers come from the company and from excited early users, and nobody has run it for a year inside a real business. So don't rebuild anything around it. Pick one inbox, run a two-week test next to your current process, and compare. Waiting has a price too. Each week your team spends hours sorting is a week a competitor's customers get an answer before lunch. Customers notice that long before owners do.
I still think about that patient with the swollen jaw. Her email sat in the clinic's inbox all weekend, under two hundred others, while she went looking for help somewhere else. That's the first job I'd give an AI traffic cop, in that clinic and in your business. Talk soon, P.S. Hit reply and tell me about your own Saturday email, the one that got missed. I read every answer, and the best ones shape what I test next.
Sources: TypeSafe AI, Wikipedia, Vercel, Cloudflare docs, DEV Community, AgentBreaking.
The AI Solopreneur • AI Strategy for European SMBs |
