Record Everything
If it happened, it is in the record.
Nobody is asked to remember to log anything, and nothing later depends on how well they remembered. Calls arrive from the AI Softphone, meetings and emails and notes arrive from the work your team already does, and all of it lands in one timeline per contact, company and deal. Transcripts make what was said searchable. Coaching, deal review and accountability all read from here rather than from anyone’s memory. What was said does not fade, does not get remembered charitably, and does not go home at five. It is there, word for word, whenever the question comes. That is the difference between remembering what a customer agreed to and being able to show them.
In the product
One stream, not six tabs.
Capture
How the record gets made
Nothing here asks a rep to remember to log something. The record is a by-product of the work, which is the only kind of record that stays complete.
Retrieval
And how you get it back out
A record nobody can query is storage, not memory. These two pages are about reading it back.
Limits
What is kept, and what is not
Recording a conversation is a legal act before it is a product feature. Both of these pages are about restraint.
The mechanism
What “one timeline” is actually made of
There is no timeline table. A timeline is assembled at read time from the tables the work already wrote to, which is why it stays complete without anyone maintaining it — and why what it can show you differs by what you opened.
Nine tables, three questions, one shape
Open a contact and six generic touch tables are read for that person — activities, calls, sent emails, meetings, notes and attribution touchpoints — alongside their campaign enrolments. Open a company and it reads its own rows plus the same touches on every contact attached to it. Open a deal and it reads a different four: the deal’s own activities including its stage transitions, the documents attached to it, the stage notes written on it, and the research briefings requested for it.
All three come back in one shape, so the screen does not have to know which question it asked. And all of it is derived rather than stored: nothing writes a timeline row, so there is no second copy to fall out of step with the record it describes.
A feature appears here by writing a touch, not by being added
The tables are generic on purpose. Nothing in the assembly knows what a journey is, or a sequence, or a campaign — it reads touches, and anything that wrote one shows up. That is why a manually logged call and an automated email land in the same stream in the order they happened, and why a capability built next year appears here without this code being opened. The cost of that choice is stated rather than hidden: two sources that ever describe the same event would show it twice, and the assumption that they do not is written down next to the code that depends on it.
What each view can and cannot roll up
A company timeline rolls up its contacts. It does not roll up its deals — deal documents, stage notes and briefings are read on the deal, and reach a company view only where an activity happens to carry one of its contacts. Saying so is more useful than implying an account view that gathers everything, because the difference is exactly where somebody would go looking and not find something.
The suppression of machine chatter is a contact-timeline rule as well. It is the view a rep lives in and the one the noise was drowning; a company view is not filtered the same way.
Questions
The things people actually ask.
Does MADDOX make the calls?
No. Calls are placed and received by the AI Softphone, which is a separate application. MADDOX receives them through a scoped integration API and turns them into records. Nothing in the CRM dials a phone, and no page here will tell you otherwise.
What happens to the audio?
It is deleted once transcription has succeeded, and everything derived from it — the transcript, the utterances, the behaviors, the activity — survives. A transcription that failed is never purged, because deleting the audio would turn a retryable hiccup into a permanent loss. The third case is the one to know about: audio the assessor finds unusable is deleted as well, and nothing was ever derived from it.
Is every activity type shown to reps?
No. Fourteen system activity types are suppressed from the human-facing timeline. They are real history and stay queryable, but they must not be what a rep sees when they open a contact.
Can the same call be recorded twice?
No. Ingestion is idempotent on the tenant and an idempotency key. A replayed webhook returns the original record marked as a replay, and produces no side effects — no second activity, no second last-contacted stamp, no second transcription spend.
Who can read a transcript?
Opening a call reads the call: the same tenant, the same permission that lets you open call records at all. Searching across calls by meaning is scoped more tightly than that — it is limited to the reps you are already permitted to see, which for an individual rep is themselves. The two are deliberately not the same rule, and the stricter one governs the surface that can reach across everybody at once.
Ask it what someone actually said.
The honest test of a capture layer is whether you can find a sentence you half-remember from a call three weeks ago. Bring one.