A conversation with a client turns into the document of the deal

1. In brief

Sam is an assistant inside a mobile app. He listens to the conversation between the tradesperson and the client, shows along the way what he has already understood, and on a tap gathers what was said into text: what the job is, when, for how much, what to buy, who to see and where to drive. From that the tradesperson makes a job card with one button, and comes back to it a day or a week later.

The second way in is equal to the first: you can have him listen to nothing at all and instead paste a chat thread from a messenger or a screenshot. Everything after that is the same.

The thesis in one line: a conversation with a client turns into the document of the deal. There is translation inside this, but translation is not the product. Earbuds will translate the conversation and forget it; when it is over the tradesperson has nothing left to send the client. With Sam, something is left.

Where we are. The app is built and running on iOS, it sits in TestFlight, and it understands five languages: English, Spanish, Portuguese, Russian, Ukrainian. We have zero outside users. We have not invited anyone into the alpha yet and we have brought in no guests. Everything said below about the product can be checked in the build; everything said about the market and about the community of tradespeople is a bet, and we write separately how we will test it.

2. What already works

There are no promises here. Everything listed is in the build and can be checked on a device.

Live listening. The tradesperson holds down a button and Sam listens to the conversation, keeping the board going: seven slots that fill up in front of you, the task, the price, when, the estimated duration, the client, the address, the materials. While he is listening Sam writes nothing into the chat and does not step into the conversation. An empty slot stays empty: if the address was never said, Sam does not invent it, he marks that the address needs to be confirmed.

The debrief after. When it stops, Sam answers the tradesperson in text, in his language, and where it fits he offers a ready line for the client: a separate block to read to the client, with playback on a tap and copy in one tap. The tradesperson sends it himself, from his own phone, with his own hands.

The second way in. A pasted chat thread or a screenshot is handled the same way as a conversation that was heard. For a tradesperson who got the request in a messenger, this is the main scenario.

The card. Saving happens when a person taps. Only then are the fields pulled out of the conversation and the card written to the server: task, price, date, address, materials, photos from the site. A week later the tradesperson opens the list of jobs and finds everything in place instead of scrolling back through a chat.

The amount is a field, not a calculation. Sam shows the amount the tradesperson named himself. He does not build the estimate for him, does not issue an invoice and does not collect payment.

Consent. While he listens, a ticker runs across the screen: "TRANSCRIBING, NOT RECORDING". Audio is not stored: speech is turned into text, the original recording is not kept. This is not a legal note at the bottom of a page, it is a piece of the interface that both the tradesperson and the client can see if they look at the screen.

What Sam does not do, on principle. He has no write permissions of any kind: text is all he can do. He does not save, does not schedule, does not send and does not pay. Every action in the system is taken by a person tapping. We treat this not as a limitation but as the condition under which the document can be trusted: only what the tradesperson let through ends up in it.

Soon. Invoicing, putting the meeting into the calendar with reminders, and a connection to messengers. The decision is made, the work is not done; wherever we name these items, they are marked "soon".

3. The problem

A solo tradesperson runs the whole cycle himself: found the client, talked, estimated, wrote it down, did the work, got paid. His only work tool is a phone. He knows his trade and he knows the material better than the client. He runs into a wall at one point: the conversation where the deal either closes or hangs.

The barrier here is not "his English is bad". Describing it that way is both wrong and insulting. The barrier is the difficulty of catching and confirming exactly what matters to this particular client when the language is not your own. It exists between dialects of one language too: you come out to a British client, it all sounds like English, and yet your read on the request drifts. The tradesperson leaves with the feeling that he forgot to ask about something, and he turns out to be right.

Then the second thing happens: what did get said gets lost. The details stay in someone's head and in a chat thread nobody goes back to. A week later the tradesperson does not remember whether they agreed on before noon or after, and whether the faucet has already been bought.

Sam stands between those two things: he helps catch what matters in the moment and leaves behind text you can come back to.

4. Who the user is and who pays

The shape of the audience. Someone who works for himself on short projects and goes through the cycle client → estimate → booking → work done → payment, with no storefront and no subscription routine. We call this criterion estimate-driven project services. It covers handymen, painters, tile setters, flooring and drywall, small plumbing and electrical, fences and decks, landscaping, cleaning, mobile mechanics, photographers who travel to the client. The list grows by the criterion, not by an industry directory.

The portrait matters more than the label. Good with his hands, proud of the craft, learns from short videos. High agency and low tolerance for friction: he will read the onboarding and figure it out if it pays back in time or money. The phone is the main tool, used in the moment: at the client's place, in the car, after the shift.

The product is a ladder, and the way in is the assistant. Level 1 is help in the moment: live listening and the debrief. Level 2 is a smart notebook: the card. Level 3 is light support for scheduling. Level 4 is CRM, far off. The front door of the first version is levels 1 and 2. This is a deliberate inversion of the funnel compared with field service management systems: they start with management, we start with the assistant, and management joins in once the tradesperson has grown into it.

Who pays. A tradesperson who already earns and has hit a ceiling because of communication. Not a broke day laborer from the hardware store parking lot, and not a crew manager. For us the beginner is fuel for word of mouth and a customer for tomorrow, but not the one we take money from today.

The beachhead. Bay Area handymen. The handyman was chosen as the technically simplest way in: one user, the client stays off the platform, the two sided contractor ↔ crew story is deferred. We build distribution into the local trade rather than through an ad account: a closed environment with low trust is entered through its own people, and that determines both the channel and the pace.

5. Why this is not a translator

This is the first objection any reader has, and we answer it before it is asked.

Translation in the moment has already been made free and expected by the platforms. Arguing with that is pointless and unnecessary: for us it is good news. The cost of the feature has been taken off us, and the value has moved up, to a place nobody has taken.

The difference is mechanical, not marketing. Voice is ephemeral. It sounds and it is gone; to use it, both people have to have the same equipment and both have to make the same effort. Text stays. You can read it out of the corner of your eye in noise and in sunlight, you can come back to it a week later, you can correct it, you can attach a price and a status to it. That is why Sam listens to everything both sides say but answers in text, and what he hands over at the end is not a more accurate translation but a document.

Our best illustration is already filmed and sits on the site: the talk was in English, Sam answered the tradesperson in Russian and asked back exactly what the tradesperson would have asked himself, whether the P-trap actually fits, whether everything is already bought, what time on Saturday exactly, and the summary for the client he wrote in English. Not one word about translation is spoken in that scene, and the promise is kept.

The line that stands on the landing page: "A translation still isn't communication".

6. Why now

Until roughly 2024 the model could not take rough professional speech in the tradesperson's own language, with its jargon, its half sentences and its "that thing under the sink", and turn it into a sensible reply that accounts for the context of the trade. General purpose translators did it word for word and sounded worse, not better. That ability became workable and cheap over the past year and a half.

At the same time the platforms moved into the horizontal: translation as a button built into earbuds and into the phone. That changes expectations but does not close the task, because the horizontal is the same for everyone and knows neither the craft nor a particular trade. The horizontal is taken. The vertical, a specific trade and its outcomes, is taken by no one.

7. Scale

We write down only the numbers with a source behind them, and we leave out the ones we do not have.

• 50.2 million immigrants in the United States in 2024, 14.8% of the population and 18.2% (32.2 million) of the civilian labor force.
• 47% (23.5 million) of immigrant adults speak English less than very well; immigrants account for about 81% of all people with limited English in the country.
• Moving up one level of language proficiency adds roughly a third to earnings (Bleakley & Chin for the United States; a comparable plus 18 to 20 percent in Dustmann & Fabbri for the United Kingdom).
• In 2022, 17% of all self employed people in OECD countries were migrants, against 11% in 2006. That is the growing pool of micro businesses we are aiming at.

And the honest gap, which we name ourselves: there is no direct number in open sources for the language barrier at the moment of the deal for a self employed tradesperson. There are dense proxies next to it, and they sit a level above. We do not substitute one for the other, and we do not think the sum of the proxies proves our slice.

8. The landscape

We do not beat the ones listed below on features. Our claim is different: the layer of live communication between tradesperson and client, carried through to an artifact of the deal, is occupied by no one.

Field service management systems (Housecall Pro, Jobber, Workiz). These are tools for a business that has an office manager. They start from the paperwork and assume the conversation went fine. For a tradesperson who is not technical, their interface is a wall. At the same time they are themselves moving toward AI communication, which indirectly confirms our thesis.

AI estimating (Handoff and similar). A strong money moment, but this is back office after the conversation, and it is English by default.

Solo estimates and invoices (Joist, GetCost and the like). Proof that the market recognizes the segment. But they sell a form to fill in, not a partner in the conversation.

Voice copilots (Plaud, AudioPen, Otter). There is demand for "talk it out and get structure back", and that is good news. But a transcript is a dead end: no card, no client, no "answer for me", no money.

Translation in the moment (Apple, DeepL, Google). A shift in expectations, not a replacement. The mechanism is ephemeral and it hands over no document.

And the main competitor that is easy to miss: a well written prompt in ChatGPT. That is the status quo and the real rival for attention. A general purpose model does not sit inside a closed, distrustful trade, does not know the local craft, and does not turn a conversation into the paperwork a tradesperson's business runs on. Everything else it does.

In one line: when the tradesperson is standing in the room with the client, not one of these tools is there with him.

9. The moat: a path, not a fact

This is the section with the most at stake, and we will start it with a direct statement: as of today we have no moat. We have a path to one. We are not going to pretend otherwise.

What is already there

Personal memory. Sam remembers what the tradesperson did, how much he earned and how he handled a given kind of job; everything the tradesperson shared goes into his next job. That creates switching costs and it is the tradesperson's own intellectual property. But it is not a network effect, it is retention, and anyone who wants to can copy it. We do not pass it off as a moat.

The way into the trade. An immigrant trade is a closed environment with low trust, and you do not walk into it from the App Store. We build distribution through people inside it. Translation is the demo; access is the company. This is the most real part of our bet and at the same time the most fragile.

What we are betting on

What follows is the bet. It is designed but not built, and it may not hold up.

Tradespeople in one metro share suppliers, local codes, climate and typical housing stock. The knowledge that is valuable here is not the average market price, it is what actually holds up in this climate and under these codes, and how to price work of this type. A general purpose model knows the world but does not know what worked for two hundred tradespeople in a particular metro. The bet is that accumulated anonymized experience by type of work turns out to be useful to the next tradesperson, and that this, rather than a set of features, is what keeps people.

How it works at the level of principle. The data is split into two layers. Private, meaning clients, photographs, trade secrets and the tradesperson's own specific pricing: never shared, exportable as a file, deletable. Shared, meaning anonymized practice by type of work, with no personal portraits, anonymized to the level of practice rather than the individual job: in a small world where everyone knows the same houses, a specific job is easy to de-anonymize, so the address and the unique details are stripped out. We treat consent as a public contract with the trade, not as a checkbox: we explain what, how and why, we ask on the way in, and after that contribution is on by default, with a visible way to see what you contributed and to put it on pause. The frame is "the trade teaches itself", not "we collect data".

Separately: an expert here is not a persona and not the model's own assessment of itself. It is a proofread body of knowledge plus a weighting rule plus a mark of origin, either verified in practice or taken from the label. Community sources are connected later and through a gate, not automatically.

What we will never do

We do not aggregate specific prices. Not in any form, not as a metro average, not as a range. The reason is simple and worth saying out loud: one story of "they leaked my prices" inside a trade group chat and the channel is dead. Everyone sets a price according to his own competence; an average price is both harmful and useless. Our moat and our main risk are the same object, and we would rather have a smaller moat and a living channel.

How we will test it

The signal on this part is behavioral, not declarative: the tradesperson agrees to contribute to the shared layer without resistance once the mechanics have been explained to him. The condition for scale is local density: until a metro has on the order of two hundred active tradespeople, the shared layer cannot be useful, and we will not claim otherwise. The sign that we are wrong: tradespeople agree to contribute but do not come back for anyone else's experience.

One sentence worth keeping in mind while reading this section: today Sam knows exactly what a particular tradesperson has shared with him. Everything else here is a bet.

10. How we make money

A subscription of about $50 a month, first month free.

The principle that matters to us more than the number: we ask for money after the product has already made someone a deal, not before. The market we are going into is poisoned by scams and by prepayment for a promise; trust here is fragile and spread runs on word of mouth. Charging up front burns exactly the asset we are doing all of this for. So the free part is bounded not by the calendar but by delivered value: the payment conversation starts once the tradesperson has been through the full cycle, the conversation, the card, the price, the message to the client, and repeated it in another week.

We frame the price around the risk of loss rather than an abstract "you will earn thousands": one estimate that does not close is hundreds of dollars gone, and the subscription pays for itself with one extra closed deal a month.

Cost is dominated by model tokens, not by storage. By our current internal estimate it runs about $3 to $4 per paying user per month. The bottleneck in the economics is not here: it is in the conversion from free to paying and in the cost of distribution.

Willingness to pay is a hypothesis, not a fact. We test it not with a survey but by taking money after closed deals and seeing who reaches for a card.

11. Risks we name ourselves

Zero users. We do not hide it behind "everyone uses ChatGPT anyway", because that is demand for AI in general, not for us.

The trust of the trade is fragile. Said above: the moat and the risk are one object. One leak of someone else's prices closes the channel entirely.

Legal. Transcribing someone else's speech requires consent, and in California that is two-party consent. We answer that with more than a footnote: the line "we transcribe, we don't record" is built into the screen, and audio is not stored.

"The client hears a robot instead of the tradesperson". Awkwardness in the moment can kill the deal, which means it hits exactly where we promise to help. The client's reaction has to be measured in the field, not assumed.

Retention after the tradesperson has learned the phrases. The key feature teaches the user and makes itself unnecessary. Our answer is memory, the workflow and the shared layer; and we say it plainly instead of hoping nobody notices.

Narrow distribution at the start. Getting into a closed trade rests on personal connections, not on a channel you can buy more of. That is our advantage and our ceiling on speed at the same time.

12. How we will test that we are right

The test is written so that it can be failed.

What we do: founder dogfooding, then two or three real tradespeople from the beachhead on real jobs.

A yes signal, both conditions at once:

1. the tradesperson came back in the second week without a reminder;
2. he told another tradesperson about the product.

"I like the idea" does not count. A third signal, separately, is about the moat: he agreed to contribute to the shared layer without resistance.

Conditions for a clean read: cohorts are split by language, Russian, Ukrainian and Spanish counted separately; a tester's time is counted in active days, not calendar days; invitations are filtered by the criterion, short projects with an estimate, the beachhead, real practice.

13. Status map

A summary for a quick check: what in this document is verifiable and what is a bet. In the next version some of the rows should move from the right column to the left; if they have not moved, that is information too.

This document will be updated as rows in the table change status. Questions and suggestions go through the early access form on the main page.