Platform · AI SDR
You decide what the SAGARIS AI SDR does alone
Pick one of four levels: manual, assisted, supervised or autonomous. These are not labels on a settings page. They are a type in our source code, and every action the agent takes is weighed against your choice before it runs. The whole table is further down this page, so you can see every answer in advance.
Included at $499 per seat per month
What each level may run alone
every action, every level
You set the ceiling and the code enforces it. That check runs before any level is read, so the limits you choose hold on every action, at every level.
How the gate runs
You can predict every answer before it runs.
Before the agent does anything, the action and your workspace's autonomy level go into one function. It answers run it, hold it for a named person, or stop it. No model takes part, so you get the same answer every time you ask. That is why the whole table below is printed here rather than described, and why anyone on your team can read it.
The gate, printed in full
action against level
| manual | assisted | supervised | autonomous | |
|---|---|---|---|---|
| draft_emaillow | rep signs | runs alone | runs alone | runs alone |
| enrich_contactlow | rep signs | runs alone | runs alone | runs alone |
| summarize_calllow | rep signs | runs alone | runs alone | runs alone |
| tag_leadlow | rep signs | runs alone | runs alone | runs alone |
| schedule_meetingmedium | rep signs | rep signs | runs alone | runs alone |
| send_emailmedium / irreversible | rep signs | rep signs | rep signs | runs alone |
| update_crmmedium | rep signs | rep signs | runs alone | runs alone |
| apply_discounthigh | rep signs | manager signs | manager signs | runs alone |
| place_callhigh / irreversible | rep signs | manager signs | manager signs | manager signs |
| send_contracthigh / irreversible | rep signs | manager signs | manager signs | manager signs |
| delete_recordcritical / irreversible | rep signs | manager signs | cannot run | cannot run |
| issue_refundcritical / irreversible | rep signs | manager signs | cannot run | cannot run |
Everything the gate is allowed to look at
Three properties on the action, plus the level your workspace is set to. That is the whole input. Anything the gate does not recognise is stopped rather than guessed at. That check runs before any level logic, so no setting can reach past the limits you chose.
- risklow, medium, high or critical
- reversiblecan the effect be cleanly undone
- externalSideEffectdoes it touch someone outside your company
Why this is different
Limits you set, enforced by the code itself.
Your AI SDR talks to your customers, so you want to know exactly what it will and will not do on its own. The table above is not a description of our policy. It is our policy. This page calls the same function the agent calls, and so does the autonomy screen in the product. All three always agree, because all three are the same code.
Sending wired as a setting
When the agent holds a send function, a preference decides whether it fires. The wire stays in place either way, and it only takes one default to bring it back.
SAGARIS
no send path exists for these four
- queue_email
- queue_sms
- launch_campaign
- book_meeting
The agent writes a proposal and stops. Sending lives somewhere else entirely, behind a gate that fails closed and a suppression list. There is no wire here for anyone to switch back on.
Built in, not switched on
For the four actions that reach a real person, the agent has no send function to call. It writes a proposal and stops. There is no setting to switch, so nobody turns it on by accident or in a hurry.
Your limits stay limits
Human sign-off sits on top of the gate, not above it. When the gate stops an action, no approval lifts it and no job title changes the answer. The line you drew holds on the busiest day of the quarter.
Held decisions you can replay, not just read
When an action waits for a person, its record keeps the inputs the gate used, not just the sentence explaining the verdict: the risk tier, whether the action can be undone and the autonomy level in force. Months later you can run those inputs back through the same function and see whether the policy, the model or the data changed.
Next door
What this connects to
- Sales engagementThe sequences the agent proposes into.
- Sales prospectingHow the agent decides who is worth the touch.
- AI dialerPlacing a call needs a named human at every autonomy level.
- AI cold calling softwareWhere an AI voice agent holds the call itself, and where a rep does, eight tools compared at source.
- SAGARIS for SDR teamsThe same cadence read from an SDR team's week, and the dial queue behind it.
The questions you ask before you trust it with your market.
- MOST ASKED
Yes, at the autonomous level. It sends the email, books the meeting, puts it on the calendar, updates the contact record and advances the deal stage, all on its own. At the supervised level it already books meetings and updates records for you. Two actions always wait for a person: sending an SMS and launching a campaign. Every action sits in a table in the source with its risk tier, so you know exactly which ones run unattended.
SAGARIS is an AI SDR for teams that want the agent working inside the CRM, dialer and inbox it operates, under limits written in code, with its actions kept in an audit trail and texting a prospect, launching a campaign and placing a call always held for a person. The best AI SDR tool for you depends on how much you want it to do alone and where your contacts live. A single seller filling their own pipeline may be better served by a credit-metered prospecting workspace such as Regie.ai, which our Regie.ai comparison describes from Regie's own pricing page. A team that wants one system for the whole motion is who SAGARIS is for, with the AI SDR included in the $499 founding seat.
SAGARIS puts its AI on the phone for inbound calls: the SAGARIS AI receptionist answers your line in a live voice, tells the caller it is an AI, finds out why they are calling, and books a meeting, takes a message or transfers them to a person. Outbound works differently. SAGARIS does not let its AI cold-call prospects on its own: placing a call is an action that always waits for a named person, and outbound calls are made by your reps through the parallel dialer, with the AI writing the brief before the call and the notes after it. If you want software that holds outbound sales conversations by itself, that is a different category from SAGARIS, and worth asking every vendor to demonstrate on a real call.
You set the ceiling and the code enforces it. Deleting a record and issuing a refund never run on their own at any level. Placing a call and sending a contract always wait for a named person, even at the highest setting. If an action turns up carrying details the gate does not recognise, it stops. That is a rule in code, not a checkbox somebody can flip by accident.
Your hard limits hold, whoever is asking. An approval can release an action the gate was holding for review. It cannot push through an action the gate refused outright. No role, no seniority and no override flag changes that. When you draw a line here it stays drawn, and you can show a security reviewer exactly where it lives.
You decide the level and it stays where you set it. There is also a ladder in the code that can work out an earned level from a real track record. It reads how often people accept the work and how often it succeeds, over a big enough sample. A bad run drops it straight back to manual. It is off unless a workspace is opted in one at a time, and it tightens rather than loosens.
Each action the agent runs, proposes or is refused gets a log row. It holds the tool, its parameters, the evidence and citations behind it, the model, the gate's verdict and reason, and the outcome. When an action waits for a person, its decision record keeps more. It names who approved or rejected it and any reason they gave, and it holds a receipt of the outcome. It also stores up to three other options the planner considered, and the inputs the gate used: the risk tier, whether the action can be undone, and the autonomy level. Those inputs let the verdict on a held action be recomputed later rather than retold from memory. That is what you hand a reviewer who asks how a message got sent.
Judge it on your own pipeline rather than on a vendor benchmark. Every call, email and reply is logged against the contact as it happens, so reply rates, meetings booked and pipeline created show up in your own reporting from the first week. Bring a list of real prospects to a demo and we will run it in front of you rather than quote somebody else's numbers.
It is part of the same $499 per seat per month as the rest of the platform, and it is not metered per send, so the cost of running it does not rise with the volume of work it does. The agent is not a separately priced add-on sitting beside the CRM, the dialer, the inbox, SMS and enrichment; it is the thing that operates them, on one seat price, with its AI work included up to the plan's monthly allowance and nothing billed on usage. If you are weighing a credit-metered agent against this, our Regie.ai alternatives page sets Regie's published plans beside the seat, with the date they were read.
Run your own actions at the level you choose.
Bring the actions your team really takes. We set the autonomy level you would run, live on the call. Then we walk the table row by row until you can call the verdict before we do. You leave knowing exactly what your AI SDR will do on its own, and exactly where your line sits.
In this topic
AI governance, approval and audit trails
- Approval before the AI emailsWhere SAGARIS draws the line on what an agent may send without you.
- The audit trail of AI actionsWhat SAGARIS records about every agent action, and what it does not.
- Agents and autonomy, documentedAn agent proposes, a policy decides: each autonomy level and what runs at it.
- Audit log documentationWhat the audit trail records, and what it does not.
- Roles and access documentationWorkspace roles, permissions and read scope.
Essays on this subject
- Approval is a boundary, not a checkbox
- Consent that fails closed
- Fail-closed means the failure is silence, not damage
- Graduated autonomy, with the levels named in code
- Read-only by construction: making a dangerous tool unrepresentable
- An audit trail nobody can query is a log, not an audit trail
- Unauditable memory is a backdoor with a nicer name
- A deletion request an AI cannot honour is not a deletion
- Deleting a fact is not the same as knowing it stopped being true
- Provenance is the only thing that makes an AI answer arguable
- What did we believe on the day we lost the deal?
- Your RAG confidently quotes the policy you cancelled last March
- A superseded document beats a correct one, and similarity is why
- Grounding every draft in real pipeline signal
- The Company Brain: one memory across every channel.