AI labs now pay companies for old work records. Email threads, support tickets, CRM notes, project boards. They want real examples of how work gets done so they can train agents to do it.
Owners who hear about this ask one question first: what is mine worth? I built How Much Is My Company Data Worth? to answer that as honestly as the market allows. It is nine questions, a list of buyers that fit, and the payout ranges those buyers publish themselves.
Disclosure: I run the site. It introduces companies to AI data buyers and earns a referral fee from the buyer when a deal closes. The company pays nothing.
The stack
The site is Astro 7 in server mode on Cloudflare Workers. EmDash runs the blog. D1 holds two databases: one for CMS content, one for sign-in and estimates. R2 stores CMS media and KV caches CMS objects. Better Auth handles sign-in with an email code or a magic link. There are no passwords. Cloudflare’s Email Service sends the receipts.
Nothing exotic. Each piece has one job.
The estimator itself is plain TypeScript. No AI model sits in the loop. That was deliberate. Checking answers against published terms should give the same result every time, and I should be able to explain every line of it.
The nine questions
The flow shows one question per screen. A single-choice question advances as soon as you pick. A multi-select waits for OK. Without JavaScript, the form shows every step stacked and still submits.
- Is your company still operating?
- Where is the company based?
- How many full-time employees?
- How many years has it been in business?
- How many years of records do you still have?
- What industry are you in?
- Where does the company keep its records?
- How much of it is in English?
- Does anything limit sharing your data?
Every question is there because a buyer publishes a rule that depends on it.
Status matters because most buyers only work with operating companies. micro1 also reviews companies that are winding down or closed. Headcount is the hardest filter, so it comes in bands from 1 to 9 up to 1,000 or more. Years of records feeds the score. The systems question lists ten kinds of records: email, team chat, CRM, support tickets, projects, code, shared files, accounting, field service software and call recordings.
The last question filters no one out. Customer contracts and regulated data change what you can license, so they lower the score and add a plain caution to the result.
None of the nine asks for a name or an email. Contact details come later, and only if you want introductions.
How buyer fit works
Each buyer is a record in one TypeScript file. The record holds the program page, the published payout, which company statuses it accepts, the minimum headcount, the countries it lists, whether records must be in English, the systems it wants, the industries it names, and the date I last checked it.
Nine buyers are in the file today: micro1, Replay, Polyshares, Prism by Labelbox, Mode, Scale AI, Endless, Handshake AI and Turing.
Your answers are checked against every buyer. Each one comes back as a good fit, a possible fit, or no fit.
No fit is for hard rules. Wrong company status. A country the buyer doesn’t list. Records mostly in another language when the buyer needs English. A headcount band that sits entirely below the minimum. Too few years in business. Or a missing system the buyer requires, like Turing’s interest in engineering records such as code or project tickets.
Possible fit covers the gray areas. Headcount comes in bands, and a band can straddle a minimum. Mode publishes a six-employee minimum for law firms, so a law firm that picks “1 to 9” might qualify and might not. Some buyers publish no minimum but usually work with bigger companies. Some name target industries, and yours isn’t on the list. Some want English, and you said “mostly.” All of those land as possible.
Good fit means every published rule passed.
Each match carries plain reasons. “Meets the 30+ employee minimum.” “Names your industry as a target.” “Wants 4 of the systems you use.” Buyers that don’t fit fold away with the rule you missed, so you can see what would have to change.
Terms change without notice. Every buyer card shows the date its published terms were last checked. I re-check on the buyer’s own page before I touch a number.
Why it shows published ranges, not a price
This was the main decision.
Computing a number would be easy. Multiply headcount by years by a made-up rate and show a big, specific figure. People would like it. It would also be fiction.
No one knows the price before a buyer reviews a sample of the records. The buyers say so. Mode calls its headline range indicative. The books don’t help either. IAS 38 doesn’t let a company record internally generated customer lists as assets, so there is no book value to anchor to.
What is real is what buyers publish. So the headline on the result is always one fully matched buyer’s own range, with the source named under it. The page says it plainly: a published range, not an offer.
When several buyers fit, the most specific number wins. A tier sized to your company beats a headline range that covers everyone. Replay publishes tiers by headcount: $10K to $100K for 20 to 50 employees, $100K to $1M for 50 to 250, and $1M+ above that. micro1 publishes $100K to $500K for qualified companies and $500K to $2M for large datasets. Either of those is more useful to a 60-person company than Mode’s $100K to $5M, even though the bigger range looks better.
Some buyers fit and publish no number. Handshake AI is one. If those are the only fits, the result says how many buyers fit and that they price after a sample. If nothing fits, the result lists the minimums your answers missed. No dollar figure appears in either case.
The fit score is a separate thing. It runs out of 100:
| Part | Points |
|---|---|
| Company size | 20 |
| Years of records | 20 |
| Systems and how they connect | 25 |
| Industry demand | 15 |
| Freedom to license | 10 |
| Language and location | 10 |
The systems part has a small bonus I like. Buyers training agents want linked records: the conversation, the ticket it created, the file it produced. So a company with a way to talk (email, chat or calls), a system of record (CRM, tickets, projects, accounting or field service) and work product (files or code) earns extra points.
The score never turns into dollars. That rule is written into the repo’s instructions so no later change can quietly break it.
The lead flow
Submitting the nine answers saves them and the result to D1. No contact details. The result gets its own URL, so you can come back to it or send it to a partner. Each IP address can only run a limited number of estimates per hour.
If you want introductions, a second short flow starts:
- Pick the buyers. Only good and possible fits are listed.
- Company name and website.
- Your name.
- Your work email.
- Your role.
- Confirm.
The last step needs two boxes ticked. One says you are authorized to explore licensing the company’s data. The other agrees to introductions and to sharing your company name and contact details with the buyers you picked. The “how we get paid” disclosure sits on that same step, not buried in a footer.
Then you get a receipt by email and I get a lead alert. You can sign in later with an email code to see your requests.
The site never collects or sends company records. Each buyer runs its own review and its own contract. That line keeps the product small and keeps me out of the data handling.
What I’d keep if I built it again
- Rules live in data. Buyer terms sit in one file with a checked date. The copy reads from it, so the page can’t drift from the source.
- Fit and price stay separate. A score is a judgment. A range is a quote from someone else. Mixing them is how a tool starts making promises.
- Contact details come last. You see the full result before anyone asks who you are.
- A no comes with a reason. “You don’t fit” is useless. “Looks for 30+ full-time employees” is something an owner can act on.
If you run a company with years of email, tickets and files, the nine questions take a couple of minutes. The answer won’t be a promise. It will be what the buyers who fit you say in public, and who they are.