The Customer the Filter Rejected — and How to Bring Them Back
A CRM pipeline is only useful when it contains customers. Most inboxes, however, contain something else: newsletters, delivery notices, invoices from suppliers, recruiters, agencies selling SEO. If every new sender becomes a card, the board turns into a second inbox and the sales team stops trusting it.
Firmspace CRM filters new senders before they reach the board. The decision has two layers. The first is a set of rules that cost nothing and answer instantly: robot addresses, service domains, unsubscribe headers, typical subjects of notifications and receipts, empty messages. Only what the rules cannot decide goes to the model, in batches of fifteen, with a ceiling of forty-five seconds per batch.
Lead qualification returns one of three verdicts. A lead is a person or company that wants something as a buyer. Junk is everything that is clearly not a buyer, including suppliers' invoices and people selling to the firm. Unsure is the honest middle: too little text, no clear intent, a missed call with no content. Unsure does not enter the pipeline, but it is not deleted either.
Every decision is written to a log: source, sender, subject, a short excerpt, verdict, category, a confidence score and one sentence with the reason. The log keeps the last four hundred decisions; the assistant reads the newest three hundred and shows up to fifty, filtered by verdict. This is the part that turns a lead filter from a black box into a colleague whose judgement can be checked.
When the filter is wrong, the fix is one command. The rejected item is restored from the log: a contact is created if needed, a card appears in the first column, and the note says it was returned manually. The entry is marked as a lead and linked to the new deal, so a second attempt is refused — restore a lead once and it stays in the pipeline.
The restore also emits the lead-created event. That matters more than it sounds: automation rules react to new leads, so a customer who was recovered still gets the welcome task, the notification and the assignment. The filter is not a gate that quietly swallows exceptions; it is a step with an undo.
One false positive is normal. The real problem is a filter that keeps making the same mistake. Firmspace CRM therefore lets the firm write its own lead rules in plain words, up to 1,500 characters, and those rules override the defaults when they conflict. Something like: we only take laptop repairs, and we do not work with intermediaries. No field mapping, no boolean logic — a sentence the team can read and correct.
The same filter runs where new contacts actually appear. For e-mail it looks at senders who are not yet contacts; a known contact simply gets the message in their activity feed. For chats it looks at the first messages of an unfamiliar conversation on webchat, Telegram, WhatsApp or another connected channel. A missed call without content is logged as unsure rather than turned into a lead. After three unsure verdicts in a row, the system stops guessing about that conversation.
Existing records can be reviewed in bulk: deals, contacts, companies, the last thirty days of mail and recent conversations. The check is read-only and returns counts plus the junk and unsure items with ids and reasons. Deletion is a separate step, always asks for confirmation and takes at most a hundred ids at a time. Nothing disappears because a model was in a bad mood.
There are budgets, because sending every message through a model would be neither cheap nor necessary. Calls for this task have a daily cap, four hundred by default per firm, adjustable on the server. If the model is unavailable or the cap is reached, a keyword fallback decides: it lets nothing questionable into the pipeline and marks it unsure instead. A global switch can turn the filter off entirely and return to the old behaviour.
The point is not that the AI is always right. The point is that its mistakes are visible, reversible and teachable. A pipeline you can audit is a pipeline a sales team will actually use.