AI sales agents: validation first, autopilot later

Handing your outbound to an AI agent feels a bit like hiring a new rep. They are fast, eager and available around the clock, but on day one you don’t know how they write, how they read your ICP, or how they handle a prospect who is clearly the wrong fit. You wouldn’t let a new hire email your best accounts unsupervised in their first week. The same logic applies to agents, and that is exactly why Norixia ships with two autonomy modes: Validation and Autopilot.

This post covers why we recommend starting every agent in Validation, what to actually look at in the approvals queue, the signals that an agent has earned Autopilot, and why some routine steps should keep asking for your approval no matter what.

Two modes, one simple rule

Every agent in Norixia has an autonomy setting:

Mode	What happens	Best for
Validation	Every email waits for your approval before it goes out.	New agents, new offers, new audiences, new mailboxes.
Autopilot	The agent sends on its own, within its quotas and sending hours.	Agents with a proven track record on a stable playbook.

The rule we follow ourselves is simple: trust is earned in Validation, then spent in Autopilot. You can always switch an agent back. Nothing about the choice is permanent.

Why start in Validation

Your knowledge base is a first draft

During setup, Norixia reads your website and prefills your Company, Offers, Targets, Selling arguments, Objections and Tone & signature. It gets you to a working team in minutes, but it is still a first draft of how you describe your business. The fastest way to find the gaps is to read what your agents write with it. If every email leans on the same selling argument, or misses the objection your buyers always raise, you will see it in the first ten drafts.

Your ICP sounds clearer in your head than on paper

“Mid-size SaaS companies in France” can mean a lot of things. Validation lets you see which leads an agent actually sources and how it interprets your targets before a single email lands in a real inbox. If the leads are off, you fix the ICP in the knowledge base or the routine’s Source leads step, not after a week of misdirected sends.

Reputation is slow to build and quick to lose

A new domain or mailbox needs care. Norixia already applies guardrails on every send: warmup that starts at 10 emails per day and adds 5 per day, a default cap of 50 emails per mailbox and 30 per agent, random spacing between sends, and a Mon–Fri 09:00–18:00 send window. Those protect your volume. Validation protects your content, which is the other half of deliverability: relevant, personal emails get replies, and replies are a strong signal to mailbox providers.

What to review in the approvals queue

The approvals queue shows each draft with its lead context, so you can judge the email against the person it is going to. You can edit a draft, click Approve and send, or reject it with a reason. Use J and K to move through the queue quickly, and bulk approve once you are confident in a batch.

Here is a short checklist we use for every draft.

1. Is this the right person?

Does the lead match your ICP and the persona the agent was briefed on?
Is the company size, industry or role a fit, or is the agent stretching?
Is there any reason not to contact them right now (existing customer, partner, competitor)? If so, add them to your suppression list so no agent tries again.

2. Is the email actually personal?

Does the first line reference something specific and true about the lead or their company?
Would the email still make sense if you swapped in a different company name? If yes, it is too generic.

3. Is the offer accurate?

Are prices, features and claims consistent with your Offers section?
Does the agent promise anything you don’t do?
Is the call to action small and clear, like a short call or a quick question, rather than a big ask?

4. Does it sound like you?

Does the tone match your Tone & signature section?
Is the length right? Most cold emails work best short.

5. Why did you reject it?

When you reject a draft, write a real reason. “Too long, lead with the integration angle” is far more useful than “no”. Then look for patterns. If you keep giving the same feedback, the fix belongs in the knowledge base or the agent’s brief, not in individual edits.

Tip: Keep a running note of your edits for the first week. Every repeated edit is a line you should add to the agent’s brief, the Objections section or the Tone & signature section.

Signals that an agent is ready for Autopilot

There is no magic number, and it depends on your volume and risk tolerance. But these signals, taken together, are a good sign that an agent can work on its own.

You approve most drafts without edits. When you are mostly clicking Approve and send, your review is no longer adding much.
Your edits are small and cosmetic. A word here or there, not a rewritten first line or a corrected claim.
Rejections have stopped repeating. The same mistake no longer shows up week after week.
Lead quality is consistent. The leads in the queue match your ICP without you second-guessing them.
Replies look healthy. Check the agent’s KPIs on the Overview: open rate, reply rate and positive replies. Look at how replies are classified (Interested, Not now, Not interested, Unsubscribe, Out of office). A rise in unsubscribes or “Not interested” is worth investigating before you scale.
The mailbox is in good shape. The mailbox health panel shows bounce rate, recent sends and replies, and today’s capacity. A low bounce rate and a completed warmup mean the infrastructure is ready too.

If most of these are true, switch the agent to Autopilot. It will keep sending within its daily limit, its sending hours and every EmailGuard check: suppressions, unsubscribed or already-replied leads, quotas and warmup, send window and spacing, and cross-agent dedup.

Autopilot doesn’t mean unsupervised

Moving to Autopilot changes who clicks Send, not who is in charge. A few habits keep you close to the work:

Read the Overview regularly. KPIs by role, the pipeline, latest runs and usage tell you quickly if something drifts.
Watch email threads. Threads marked “Needs you” are where the agent wants a human. You can take over any thread at any time, and use or regenerate the AI suggested reply.
Go back to Validation when things change. New offer, new market, new persona, new mailbox: each one is a reason to review again for a while.

Keep “Always ask for approval” on sensitive steps

Autonomy isn’t all or nothing. In the routine builder, any routine step can force Always ask for approval, even when the agent runs on Autopilot. That lets you automate the routine work and keep a human on the moments that matter.

Good candidates for this setting:

Replies to interested leads. A routine triggered “when a lead replies”, filtered on Interested, is often the step closest to revenue. Many teams prefer to read those responses before they go out.
Emails to high-value accounts. If a branch of your routine targets strategic accounts, keep a human on it.
New sequences. When you publish a new routine or a new email step, require approval until you have seen enough output to trust it.

Here is how a typical setup might look:

Trigger: Every week on Mon–Fri at 09:00
→ Source leads (Apollo, ICP in natural language)
→ Send an email                      (Autopilot)
→ Wait 3 business days
→ Condition: has replied?
   No  → Follow-up                    (Autopilot)
   Yes → Condition: reply classification = Interested?
          Yes → Send an email         (Always ask for approval)
          No  → End

The first touch and the follow-up run on their own. The reply to an interested prospect waits for you. You stay in the loop exactly where judgment matters most, and nowhere else.

Try it with your own team

Norixia makes this progression easy: start every agent in Validation, review drafts with full lead context, and turn on Autopilot when you are ready, with guardrails on every send. Learn more about how agents work and how to build routines with approval steps where they count. The approvals queue and EmailGuard checks are described on the outbound page, and the setup guide walks you through choosing autonomy for your first team.

You approve. They execute, day and night. Create your free account and hire your first AI agent today. No credit card needed.