Skip to content

Governance · the layer under the agents

Anyone can answer the phone.
Ask what happens next.

Arbol turns every conversation into a record you can defend: who it was about, what was collected, what was promised, who owns the follow-up — and the consent that allowed the call in the first place.

One call, traced link by link

Illustrative trace · ids shortened

  1. CNV · 21:47

    Conversation

    Inbound voice call, 7 minutes, after hours

    Recorded, transcribed turn by turn, closed with a title, a summary and an outcome. Same shape whether it arrived by phone, WhatsApp or web chat.

  2. IDN · phone

    Person

    The caller is resolved, or created

    The number becomes an identity attached to a person. One match links. No match creates. Two possible matches never merge silently — they raise a task for a human.

  3. TOOL · 4 calls

    Tool trace

    What the agent looked up in your systems

    Every lookup and write is stored with its input, its output, its latency and its errors — through one execution path shared by voice, messaging and web.

  4. FACT · ×6

    Facts

    The conversation becomes structured data

    Not free text: values written against fields your team approved, each with the source it came from, when it was true and how confident the system is.

  5. TASK · owner

    Follow-up

    The open loop gets a name and a date

    A conversation that ends unfinished produces a task with an owner and a due date, or an explicit closure. Those are the only two endings allowed.

  6. EVT · caused by

    Causation

    Every link points at the one before it

    “Why did the system contact this person?” is a query, not an archaeology project. The chain answers it in one hop.

Six links, one chain. The last one is the point: every record names the record that caused it, so the answer to “why did this happen?” never depends on somebody remembering.

Four questions most AI phone stacks cannot answer

Not because the vendors are careless — because answering them requires a data model underneath the voice, and most of them only shipped the voice.

Why did the system call this person?

Because a specific datum changed, and a rule you approved was listening. The event that fired, the rule that matched and the task it created are all linked — you can read the reason without calling anyone.

What did the agent actually collect?

Fields, not vibes. Each answer is written against a definition your team created, with the sentence it came from, the moment it was true and a confidence value. Anything the agent is not authorized to record is dropped, not guessed.

Who owns this follow-up?

A person, with a date. Conversations that need a human close produce a task; tasks carry an owner and a due date; the queue is visible. An alert nobody owns is not an alert, it is a rumor.

Someone said stop. Where does that apply?

Everywhere, at once. A revocation heard on a call applies to voice, text and email against that person, and the outbound executor checks it before every dial — including for automated campaigns.

The doctrine

Five rules the system will not break for you

Every one of them costs us something — reach, speed, a bigger campaign. That is what makes them worth writing down.

  1. I

    Nothing dies in an inbox

    A conversation that ends open has exactly two legal endings: a task with an owner and a date, or an explicit closure written down by a person.

  2. II

    Absence is a trigger

    The strongest signals are the records that never arrived — the visit with no result, the follow-up that never happened. Rules can watch for a missing fact, not only for a new one.

  3. III

    There is a contact budget

    Each person has a cap per purpose per period. Twice a year, not six times. Proactive outreach without a ceiling is spam, and the ceiling is the governance.

  4. IV

    Sensitive facts never leave their conversation

    When the system has to reach a relative to update a phone number, that call is generic by design. Fields marked sensitive are excluded from the prompt. The relative learns that you are looking for someone, never why.

  5. V

    No silent merges

    Two people who look like the same person stay two people until a human decides. Ambiguity becomes a task, never a quiet overwrite of somebody's record.

What is written down, every time

Not a call log. Six kinds of record, produced by the same pipeline whether the conversation arrived by phone, by message or on your website.

The conversation
Channel, direction, duration, turn-by-turn transcript, title, summary, outcome and sentiment.
The people
The caller and the people the call was about — a guardian arranging visits for two family members is three people, not one row.
The tool trace
Every lookup and write against your systems, with inputs, outputs, latency and failures.
The facts
Values against approved fields, with provenance, the time they were true, and confidence.
The consent
Permission by person and purpose, plus every revocation and where it came from.
The usage
Seconds of voice, messages and model usage per conversation — so the bill is explainable.

It calls because a datum changed,
not because someone remembered

This is the difference between an AI that answers and an operation that runs. Five steps, and the fifth one feeds the first.

  1. 01

    A datum arrives

    A result from the lab feed, a document the patient sent, a fact from this morning's call.

  2. 02

    A rule wakes up

    Rules listen for new facts, for missing ones, or for a clock. The cohort is a live query, not a list somebody exported in March.

  3. 03

    Work appears with an owner

    The rule creates tasks — for the agent to call, or for a person to decide. Each one carries the event that caused it.

  4. 04

    The agent calls, inside the fence

    Consent checked, contact budget respected, calling window enforced in the patient's own time zone.

  5. 05

    The outcome comes back in

    Booked, no answer, needs a human. The result is written as new facts — which is what the next rule reads.

The loop runs inside the fence: consent per person and purpose, a cap on how often anyone is contacted, and calling windows in the patient's own time zone.

Straight answers

Prefer email? Write tosupport@getarbol.com — a person answers.

01Is this a separate product from the AI agents?

No. It is the layer underneath them. The agents hold the conversations; governance decides what the conversation becomes, what is allowed to happen next, and what gets written down.

02Do we have to move our clinical data into Arbol?

No. Nothing clinical is mirrored into a parallel chart. What enters the system is a fact attached to a person, with the record it came from — your source system stays the source of truth.

03Can we see why a specific patient was contacted?

Yes. Open the person, follow the chain: the datum, the rule, the task, the call and the outcome. Each link points at the one that caused it.

04Who defines the fields the agent is allowed to record?

You do. Fields are created by your team, with a type and a description. The agent can only write against fields that exist, and anything else is discarded rather than invented.

05What is live today?

Recording, transcription, analysis, structured facts, the tool trace and the scheduling engine are running in production. Parts of the proactive loop — clinical feed ingestion in particular — are being rolled out per customer. We will tell you exactly which on the call, in writing.

Bring one week of calls. We will show you the record they would have left.

Thirty minutes, your own operation, no slideware. If the record does not tell you something you did not already know, we will say so.