Shadow, approval, autonomous: rolling out an AI agent safely
Every Mokanata agent moves through the same three levels before it sends on its own. Here is what each level is for, what to measure in it, and when to move on.
Mokanata team4 min read
The fastest way to lose trust in an AI agent is to let it talk to customers on day one. The second fastest is to never let it talk to them at all, so that it stays an expensive demo. Mokanata takes the path in between: every agent moves through the same three levels, shadow, approval and autonomous, and you decide when it has earned the next one.
This guide walks through what each level does, what to look at while an agent is in it, and the signals that tell you it's ready to move on.
Why a ladder, not a switch
A switch asks you to make one big bet: either the agent is good enough, or it isn't. A ladder turns that bet into a series of small ones, each backed by evidence you collected at the level below.
It also gives your team time to adjust. Reps who have read a few hundred drafts in shadow mode know what the agent tends to get right and where it hesitates. When it starts sending on its own, they already know which threads to keep an eye on.
Level 1: shadow
In shadow mode the agent reads every conversation it's assigned to and drafts a reply, but it never sends anything. Drafts show up in the agent's runs as "drafted in shadow", next to the reply a person actually sent.
This is where you find out whether the agent understands your business at all. Use the time to check:
- Coverage. Does it have a source for what customers actually ask? The answer check in the agent's knowledge settings rates a question as Covered, Partial or No source. Every "No source" is a gap to fill before anyone relies on the agent.
- Tone. Would you be comfortable if this draft went out under your company's name?
- Judgement. When it isn't sure, does it say so? A good draft with a low confidence score is more useful than a confident wrong one.
Note
The create-agent wizard recommends a shadow period of a set number of days, and new workspaces create their first agent in shadow mode. You can go live straight away, but you give up the evidence this level collects.
Level 2: approval
In approval mode the agent drafts and a person signs off before anything is sent. Drafts land in the approval queue with their confidence score, the sources they cite and the agent's reasoning. A reviewer can approve, edit, snooze, discard, or reassign the draft to a teammate with a note.
This is the level where most teams spend the longest, and that's fine. What you're measuring changes:
What to watch
| Signal | Where to find it | What good looks like |
|---|---|---|
| Edit rate | Approval queue, agent runs | Most drafts are approved as written |
| Confidence vs. outcome | Low confidence tab | Low scores line up with drafts you'd edit |
| Escalations | Escalated tab | Escalations are for real reasons, not noise |
| Time to draft | Agent performance trends | Drafts are ready before a person would be |
The second row matters most. If the agent's confidence is well calibrated, you can trust it to tell you which drafts need a human, and that's the foundation for the next level.
Approving in bulk
Once the routine drafts are clearly fine, review them in batches. "Approve N with confidence 0.90 or higher" lists every draft it would approve and asks you to confirm. Anything a guardrail holds stays in the queue regardless.
Level 3: autonomous
An autonomous agent sends on its own unless a rule holds the draft. That last clause is the whole point: autonomy in Mokanata is never unconditional.
Two kinds of rule decide what still comes to a person:
- Guardrails are hard limits, such as a price below the approved band, a delivery promise, a contract change or a refund. A draft that breaks one is held, routed to a person, blocked or revised, depending on how you set the rule up.
- Auto-approve rules release the routine drafts by agent, confidence floor, channel and deal value, optionally after a review window.
Even at this level, the agent's own limits still apply: a discount band, a cap on new conversations a day and actions an hour, and quiet hours in the contact's local time.
When to move up a level
There's no universal number, but these questions work as a checklist:
- Have you read enough drafts to have an opinion about this agent's judgement, not just its writing?
- Are the drafts you edit mostly the ones it flagged as low confidence?
- Do you have a guardrail for every outcome you'd never want sent unreviewed?
- Does the team know how to stop an auto-send and take over a thread?
If any answer is no, stay where you are a little longer. Moving down is always available too: pausing an agent or changing its autonomy level takes effect immediately.
The short version
Start in shadow and fill the knowledge gaps. Move to approval and calibrate confidence. Go autonomous only with guardrails in place, and keep reviewing by exception. The agent does more of the work at each step, and you always know why it's allowed to.