AI agents Your agents read the CRM, write the summaries and create the tasks. Nothing is written without your approval. See how
Skip to content
AI agents

Agents inside the CRM.
You stay in control.

An agent reads the data you wire to it, at the pace you choose, and applies what you authorised it to do. Every run is logged, and nothing is written without your say-so if you require it.

app.naty.io — AI agents walkthrough
Live

From the deterministic workflow to the store, then the builder. The other walkthroughs are in the demo center.

The principle

An agent is
four things.

No magic and no black box. Each run makes one model call with a context you chose, and applies the result in the CRM.

01

A trigger

When the agent wakes up: a schedule, a CRM event, or you.

02

Some data

What it may read. Explicit sources, and nothing else.

03

Instructions

Your prompt, your model, your token budget.

04

Actions

What it may write, on the triggering record or on the ones it resolved.

Triggers

When the agent
wakes up.

Three ways to fire it, all of which compose with the rest of the CRM because they share the same event vocabulary as the automations.

01

Scheduled

A cron schedule with its timezone: daily, weekly, monthly. A 15-minute minimum interval, and configurable catch-up for missed runs.

Daily 08:00Monday 09:001st of the monthCatch-up
02

CRM event

The agent wakes up when something happens in the CRM, with configurable rate limiting so it never runs away.

contact.createddeal.stage_changeddeal.wondeal.losttask.completedform.submittedbooking.created
03

Manual

An on-demand run from the interface, to test a setting or catch up on something without waiting for the next schedule.

On demand
Inputs and outputs

What it reads,
what it writes.

The agent guesses nothing and digs nowhere. You wire sources to it, you tick actions, and its reach stops there.

Data sources

Eight context providers

Triggering record The contact, deal or company involved, with its activities.
Activity digest What happened over the last X days.
CRM statistics Volumes and trends over a period.
Deal pipeline The state of the book, stage by stage.
Stale deals Opportunities with no activity for N days.
Task digest Overdue, upcoming, by person.
Meeting digest Past or upcoming over N hours.
Contact digest New contacts, records with missing data.

Allowed actions

Six possible writes, to tick

Create a note On the record concerned, dated and attributed.
Create a task With an assignee and a due date.
Update a field Native or custom.
Add a tag To segment on afterwards.
Send a digest By email, to whoever you decide.
Trigger an automation The bridge to the deterministic engine.
Targeting happens on the triggering record, or on the records the agent resolved itself from its sources.
Guardrails

Nothing is written
without your say-so.

This is the part that matters. An agent editing the CRM on its own is a risk, so the product starts from the opposite assumption: by default, you approve.

01

Human approval

Switched on agent by agent. Nothing lands in the CRM without approval, in a dedicated queue with a configurable expiry.

02

Two distinct permissions

agents.approve is separate from agents.manage. Configuring an agent and approving what it writes are two different jobs.

03

Action cap

A maximum number of actions per run. An agent going off the rails cannot write a hundred records.

04

Skip when empty

No data to read, no model call, no bill. The agent goes back to sleep.

05

Full history

Context read, model response, actions applied, errors. Every run can be inspected.

06

Duplicate lock

An agent cannot run twice in parallel, even if two triggers overlap.

07

Model and budget per agent

The model and the token budget are set agent by agent, not globally.

Approval queue

3 follow-up tasks

Stale deal nudges · 4 min ago

Approve Reject
The builder

The agent explained
in plain language.

As you configure it, a panel sums up in one sentence what the agent will do, and what its context will cost in tokens.

The agent builder: name, trigger, sources, actions, and the panel that sums the agent up in plain language
Agent store

Ten agents,
installed in one click.

You start from an agent that works, then adjust it: schedule, sources, actions, approval. Nothing is fixed.

See the store demo
Reporting

Weekly activity digest

What moved this week, saved as a team note.

Monday 08:00

Reporting

Morning brief

Today's tasks, upcoming meetings, pipeline state, by email.

Weekdays 07:00

Pipeline Approval

Stale deal nudges

Spots opportunities with no activity for 14 days and creates the follow-up tasks.

Tuesday 09:00

Meetings Approval

Meeting follow-up

After a meeting, drafts the write-up and creates the next action.

Every hour

Prospecting

New lead triage

Rates every new contact hot / warm / cold and tags it.

contact.created

Prospecting

Form qualification

Reads each submission and creates the right task.

form.submitted

Reporting

Monthly pipeline review

The state of the book, sent by email.

1st of the month

Maintenance

Data hygiene

Spots records with no email, no phone, no job title or no owner.

Weekly

Pipeline

Won deal debrief

Captures what worked: why we won, what to take from it.

deal.won

Pipeline

Next step after a task

Automatically chains the following action.

task.completed

Choosing

An agent, or
an automation?

Both share the same event vocabulary. The question is whether you want a model in the loop, or a rule that never moves.

AI agent

When it has to read, summarise, decide

Writing a meeting recap, qualifying a lead, spotting what deserves a nudge: anything that means interpreting a context rather than applying a rule.

  • Written output
  • Context to interpret
  • Human approval available
  • Cost in tokens
Automation

When the rule is known in advance

Assigning an owner, adding a tag, creating a standard task: a visual trigger → conditions → actions builder, fully deterministic.

  • Predictable result
  • No model call
  • No variable cost
  • Run history
See the automations

An agent can in fact trigger an automation: it decides, the automation executes.

Which agent would save
you the most time?

Show us your organisation and your deals. You leave with two or three agents to install in the first week.