A GTM system that finds companies at the moment they cross from product-led growth to a sales-led motion, and acts on it before anyone else does.
Most outbound starts from a list. Cortana starts from a moment.
The target is a Series A to C company whose self-serve product is pulling it upmarket. That moment leaves tells in public, weeks before it shows up in a database. I built Cortana to run my own Founding-AE search across 34 of these companies.
Each arm adds a capability, not a new stack. OS knows every account. Engage reaches the right person. Signal reacts when something changes. Intel plans the account.
Keeps 34 target accounts continuously researched, scored and queryable. Every Monday one scheduled call runs the whole pipeline and writes the result to Salesforce. Claude reads and writes the same data through the same tools, in plain English.
A real MCP server over a real CRM, not the stock integration. Calibrated scoring instead of a hand-tuned formula. An unattended pipeline with guardrails.
Built. I found the schedule had gone silently inactive after Sept 22, turned it back on Oct 7, and a catch-up run re-scored 34 of 34. The first fully unattended run is Mon Oct 12.
Turns a scored account into a conversation, for the top accounts only. A tool assembles the facts, the model writes, and a person reviews and sends. The first three intros went out Oct 7, and every link back to this site is tagged so a click or a visit can be traced to the exact email.
Personalization grounded in evidence, not templates. Every claim in a message traces back to a source.
Three real intros sent and tracked so far. The bar is five or more, with replies logged.
OS runs on a calendar. Signal runs on events. It scores each change with the same engine OS uses, then routes it: an Engage draft, an Intel refresh, or a note. The first source is live: every visitor to this site is tagged with the owner of their network, and a visit that follows a tracked email is credited to that exact message.
One scoring engine serving a second consumer. Signal-based GTM, where timing is the edge, not list volume.
A real event triggers a real action end to end, with no one watching.
For an engaged account, Intel reasons over everything the system knows: Salesforce history, archived JDs, cited evidence, contacts and notes. A tool assembles the evidence pack and the model writes the plan from it, citing sources and labelling inferences. Plans are versioned and logged to Salesforce.
An agent that reasons over a system of record it helped build, not a chat window over a PDF.
v1 is built and the first plan is written. The bar is a plan used in a real conversation, and one that updates itself after an event.
Researching one account by hand took me 60 minutes, and the notes went stale within a week. Cortana now assembles the same evidence brief in half a second and re-scores all 34 accounts every Monday. That is machine time. My review of each brief is still being measured.
Job-board slugs collide. 4 of 28 auto-discovered boards belonged to other companies, and the data looked perfectly plausible. Fixing it moved three rankings by 41 to 62 points.
Re-scoring unchanged input moved 19 of 34 accounts by 1 to 6 points. So only a move of 7 or more counts as real movement.
The scheduler step passed standalone tests and still failed inside n8n twice. It only counted as working once it ran in n8n's own engine.
The weekly schedule sat switched off for two weeks and nothing complained. Now a Monday task checks that the run happened, drafts the digest, and says so if it didn't.
A mail connector rewrote every tracked link and deleted the tracking pixel on save. I only caught it by inspecting the saved draft. Tracked emails are now sent from a rendered page, and links to my own site carry a tag instead of a redirect.
My own browsing was being counted as a prospect's visit, because the tag sticks to the browser. Now only a landing that carries the email's tag after the real send time counts, and scanner clicks within two minutes are ignored.
The first Clay harvest cost ~185 credits. Weekly hiring signal now comes from free public ATS APIs, saving ~61 credits a week.
The full case file: architecture, scoring calibration, and the first outbound it sends, with the reply data. Five quick questions, about 30 seconds.