Skip to content
Saleswoman

AI sales agent workflow, from research to a checked draft

Four steps, in this order, every time. Three of them are what every tool in the category does. The fourth is the one that gets skipped, and it is the reason this page exists.

Run it on a prospect

  1. Research, and keep the line each fact came from

    The agent assembles a brief before it writes a word. It reads the CRM record, the notes your team left, the account pages you connected and the history of any previous contact. Every field it collects is stored with the place it came from, and that pairing is what the other three steps are built on.

    A field it could not fill is not an error and it is not quietly guessed. It is simply not available as evidence, which is a different thing, and the drafting step is told so.

    It does not scrape a company it was never given access to, it does not buy contact data, and it is not a lead database. You bring the list. It reads what you connected.

    The brief it wrote

    • CRM note, 14 March: the renewal on their routing tool lands in July. CITED

      Source: HubSpot note on the account record.

    • Careers page lists two dispatch roles open in Columbus. CITED

      Source: the careers page you connected, read today.

    • Headcount growth over the last year. UNVERIFIED

      Nothing found in the sources connected. It stays empty rather than estimated.

  2. Draft it short, and mark every sentence that asserts something

    You pick the angle and the tone. The agent writes to it, keeps the email inside the length a busy operator will actually finish, and puts the unsubscribe line and the address block in before anyone has to remember them.

    Then it reads its own output sentence by sentence and sorts each one: does this assert something checkable about the world, or is it a greeting, a question, an offer of a meeting? Assertions get underlined. An assertion that traces back to the brief is underlined in one colour, one that does not is underlined in the other.

    The unsupported sentence is often the best sentence in the email, because it is the one that shows you understand the situation. Deleting it silently would leave you a safe, bland, unopened email. Marking it hands you the decision instead.

    The draft, as it would send

    empty miles on the ohio lanes

    Your careers page has two dispatch roles open in Columbus. Usually that means the lanes are growing faster than the desk can cover them.

    Most fleets your size are losing a double digit share of backhaul. We help dispatch teams price return legs before the truck is empty.

    Worth fifteen minutes to see whether the numbers hold on your lanes?

    Unsubscribe link present. Postal address block present.

  3. Check the draft, not a report a week later

    Six checks run on the text that is about to leave, and each one can hold the send on its own. There is no composite wellness number and no traffic light without arithmetic behind it: the spam score names the triggers it found, and the reading grade is next to the range it is being judged against.

    The claim ledger is the one that matters. It lists every sentence that asserts a fact, marks it cited or unverified, and for anything unverified it proposes a rewrite that removes the guess, usually by asking instead of stating.

    This is the same check the spam score checker page walks through line by line, and the same one the demo runs on whatever you paste.

    REVIEW

    Spam risk 12/100 Grade 7.1 84 words 1 of 40 today

    Held for review: one claim in the draft has no source behind it.

    Claim ledger

    2 of 3 cited

    • Your careers page has two dispatch roles open in Columbus. CITED

      Source: careers page, read at draft time.

    • Most fleets your size are losing a double digit share of backhaul. UNVERIFIED

      Suggested rewrite: ask how much backhaul goes empty, rather than telling them.

    • The renewal on their routing tool lands in July. CITED

      Source: CRM note dated 14 March.

    Spam triggers found: none. One link. Reading grade 7.1, inside the 6 to 9 range this check targets.

  4. Hold, then send inside the cap you typed

    A clean draft joins the queue for that mailbox and goes out inside the daily ceiling you set. A draft carrying an unresolved claim does not. It waits for you to supply the source, take the rewrite, or cut the sentence.

    The cap is a number you type, per mailbox, and the agent has no override for it. No burst mode, no catch-up after a quiet day, no quota it is trying to reach. When the ceiling is hit the queue stops and waits for tomorrow.

    Replies stop their own thread. A prospect who answers never receives the next step, and a stop applies to everything in flight, because sends are queued against the cap rather than fired in a burst.

    The queue, this morning

    • Queued to send 6 drafts All six cleared every check.
    • Held for you 1 draft One claim has no source behind it.
    • Sent today 6 of 40 The ceiling is yours, and the agent stops at it.
    • Stopped on reply 2 threads They answered, so the follow up was cancelled.

Four things this workflow deliberately does not do

Every one of these is a real product in someone else's category. Saying we do them would be a fabricated fact, and the whole argument on this site is that we do not make those.

Record or transcribe calls

No call recording, no transcription, no meeting notes from audio. That is a different category and we are not in it.

Test inbox placement

The checks run on the draft. They do not seed test inboxes, they do not look you up on a blocklist, and they cannot tell you which folder a message landed in.

Warm up a mailbox

There is no warmup network and no reciprocal sending pool. The cap is how volume is managed here.

Sell you contact data

No lead database, no email finder, no enrichment resale. You bring the list and the sources.

What it does do is set out capability by capability on sales automation software, priced flat rather than per seat on AI SDR software pricing.

Watch the four steps run on your own prospect.

Paste two lines about an account, pick an angle, and read the ledger the agent produces. No account, no card, no signup before the result.

Run it on a prospect

Your data stays yours. Prospect data is never used to train a model. Nothing is charged today.