Zinta

The agent worked for us, which is not the same as being a product. Turning it into one meant answering questions an internal tool never has to: how a stranger’s data stays separate from ours, who pays for the model, what happens when a hunt finds nothing worth sending, and how someone takes their pipeline with them if they leave. The category made it harder — AI SDR tools churn at 50 to 70% a year because buyers stopped trusting output they could not verify, so anything we shipped had to be checkable rather than merely plausible.

• Kept the behaviour as a specification rather than a script. Targeting, the signal hierarchy, mail craft and compliance live in reviewable markdown, so changing how the agent works is an edit rather than a deploy.
• Made evidence a hard rule instead of a preference. Every company carries a source URL fetched during that run, and the page at it must contain the fact stated. No verified page, no prospect — and when nothing clears the bar the customer is told so in writing rather than handed a padded list.
• Never send. The agent drafts and stops; the customer sends from their own mailbox. Their domain reputation stays theirs, they remain the sender of record for GDPR and CAN-SPAM, and a draft we got wrong cannot reach a stranger without a human reading it first.
• Shipped delivery over MCP, so the hunt runs inside the Claude, Cursor or Codex subscription the customer already pays for. That is why the entry plan is $19 rather than $79 — our marginal cost on it is a few database writes.
• Built the multi-tenant boundary as one rule with a test behind it: every customer route reads the account from the session token, and no route anywhere accepts one from a path or a body. The test is mutation-tested — each property was deliberately broken to confirm the suite catches it.
• Treated data rights as buttons rather than an inbox. Export everything as JSON, export the pipeline as CSV, delete the account and every project, all self-serve — a right you have to email someone about is a right with friction in it.

• A measured refusal rate of 17% across six real target sites — five produced a company with a dated, sourced reason and one honestly reported that nothing cleared the bar. That number is the product working, not failing.
• Mails that hold to the spec measurably: recent batches sit at 93 to 116 words, zero em-dashes, an opt-out line on every one.
• Contact discovery that reports its own gaps. A verified address is found for roughly half a batch; the rest arrive with a contact page or LinkedIn profile rather than a guessed address, because a fabricated contact is worse than none.
• A human approval step that has never been bypassed — by design rather than by policy, since the product has no ability to send.
• Near-100% gross margin on the entry tier, because the customer’s own model does the work, and honest unit economics on the managed tiers where model spend is the bill and AWS is two cents a customer a month.
• Proof of what we tell clients: we do not only build AI agents, we run one in production every day — and we found out where it does not work by measuring rather than guessing.