3 operating levels your AI needs
Our AI Sales System prototype made a massive mistake early on. I actually still use it as a warning today.
Our AI Sales system prototype made a massive mistake early on. I actually still use it as a warning today.
It started offering customers products that our client didn’t sell anymore.
It sounded totally natural while doing it, and it was even scheduling meetings for these fake products.
We only caught it because we were still manually reviewing every single conversation.
That was the exact moment the right architecture became obvious to me. We had given the system more freedom than it had actually earned.
Today, it handles hundreds of thousands of conversations, fully on autopilot.
But it didn’t start out autonomous. It had to graduate through three distinct levels: Assisted, Delegated, and finally, Autonomous.
Most operators get this wrong. They treat AI like a volume dial they can just slowly turn up. It is not a dial. It is three entirely different operating modes. Each one has a different role for the human, and each one has a completely different surface area for failure.
The biggest mistake I see is not people using AI too cautiously. The mistake is promoting the AI without proof.
Your role changes at every level
The cleanest way to classify any AI process is to ask one question: What is the human’s role?
1. Assisted Level
Here, the AI generates the output, but the human owns the decision and the final result. Every single output is reviewed before anything ships, sends, updates, or acts.
Best for: Drafts, scoring, classification, summaries, and suggestions.
The Risk: Bottlenecking. If the AI is consistently right but the human still reviews every single action, the human becomes the constraint on your scale.
2. Delegated Level
At this stage, the AI owns the step. The human sits in the exception path, not the default path. Anything falling below a strict, defined quality threshold gets kicked to a human for review. Everything above the threshold just runs.
Best for: The vast majority of real-world business systems (where they should honestly live for a long time).
The Risk: Invisible degradation. The process looks like it’s running smoothly, while unmapped failures silently accumulate in the background.
3. Autonomous Level
The AI makes decisions entirely inside a clearly defined scope. The human is still present, but strictly as the architect and the final escalation path.
Best for: Tasks that have earned absolute trust through the first two levels. It’s not “AI doing everything.” It is AI doing one bounded thing perfectly.
Moving up
Before I ever move a system up, I run through six specific checks.
I look for three things before moving from Assisted to Delegated:
A one sentence quality threshold. “Good enough” does not count. A trained reviewer must be able to read it and check the output consistently.
Proven variance. That quality has to hold up across 20 or more real examples. I am not talking about sterile demos or perfect inputs. I mean real production variance.
Mapped failures. I need to know exactly what a failure looks like, where it routes, and who is responsible for fixing it.
Then, before moving from Delegated to Autonomous, I look for three more things:
A low, stable exception rate. My personal baseline is fewer than 5% of runs going to human review across a 30 day period.
Automated edge cases. If an exception pops up three times, it is not rare anymore. It has to become a rule, a route, or a guardrail.
Human trust. This is not a technical metric. It is a social one. Your team needs to see the system work in their context enough times that they stop checking it by reflex. If you skip this, your technically correct system will just fail socially.
It’s alright to hit a ceiling
You also have to accept that not every AI step is meant to climb all the way up that ladder.
Some steps have a permanent ceiling at the Assisted level. Think about founder sales calls where the human relationship is the actual product. Tasks with legal or ethical accountability belong here too. Same goes for client facing judgment calls when the client expects human ownership. The AI can prep, flag, summarize, and challenge you. But the human decides.
Other steps have a permanent ceiling at the Delegated level. Strategic judgment belongs here. You can use AI to model scenarios for positioning, major resource allocation, or creative direction. But I still want a human making the final call. Rare or novel processes also stop here. Autonomous systems need stable patterns. If the work changes shape every single time, autonomy is just performance theatre.
That was the deeper lesson implementing thousands of automations taught us.
That hallucinating prototype belonged at the Assisted stage. Manual review was not friction. It was the correct architecture. We only moved it to Delegated after we tested it on real inputs, with real variance, and mapped all the failures.
The AI owned the outreach, and humans reviewed the exceptions. Only after those exception paths were proven did we make it Autonomous.
Even today, every new custom implementation starts with live data testing before we ever call it fully Autonomous.
There are two reasons for that:
First is knowledge base completeness. A new client has a different market, a different ICP, different products, and different objections. Until the system runs on that live data, it is back at Delegated.
Second is client team trust. The system itself might be technically ready. But the client’s team is not. Trust does not transfer from one deployment to another.
Assisted means you are the loop. Delegated means you handle exceptions. Autonomous means you designed the system and remain the escalation path.
That’s a rule of thumb for you to audit every single AI system in your business.
🔧 Tools & Resources
Helicone: Highly useful when a Delegated step needs live quality monitoring instead of a dusty written QA rule nobody actually checks. Just remember that observability doesn’t fix a fundamentally bad process design.
Mem0: A great tool when an Autonomous system needs persistent memory across sessions and client contexts. The catch? Memory can preserve bad assumptions if your foundational knowledge base is weak.
Tango: Perfect for the transition before Assisted moves to Delegated. It captures real workflows into SOPs that both humans and agents can follow. The constraint is that captured steps will still require human judgment and cleanup.
Autonomy is not the absence of humans. It is the result of placing humans in the right part of the system.
At the start, they review everything. Then they handle exceptions. Later, they design the boundaries and stay available when the system reaches them.
That sequence is slower than pretending the system is autonomous on day one.
But it is also the exact reason it keeps working when real clients, real data, and real edge cases arrive.
Build a calmer business,
– Yuri
Yuri Vonchitzki
LinkedIn · YouTube
Want to reach 21,000+ business founders and operators implementing AI? Contact us for sponsorship details.
The graduation framing is the useful part. In OR-adjacent AI, I would want the same promotion test before any agent moves from assisted to delegated: what failure looks like, which human owns the exception, and what signal proves the workflow became less brittle rather than just less supervised.