The Agentic Allowance
Everyone has an internal budget for how much they're willing to let AI act without them. I call it the agentic allowance. Here's where that idea came from and why I think it determines everything about how AI gets adopted.
Levi Lais
Founder, VirtualExperts
February 22, 2026
5 min read
The Marco Series: Levi's words, edited from a Marco recorded October 23, 2024. Chai's analysis at the end.
Think about the last time someone handed you login credentials and said "just handle it."
Maybe it was admin access to something that actually mattered. A client's CRM. A shared bank account. A production server. What's the first thing you felt? It wasn't confidence. It was weight. The little voice that said: I should probably not break anything.
Now flip that. You're handing that access to AI.
Tim Ferriss figured out the right framework for this back in 2007. In The 4-Hour Workweek, he described giving his virtual assistant a spending threshold - $100. Under $100, don't bother me, just do it and tell me later. Over $100, check in first. That's not a trust problem. That's a system. The VA has autonomy up to the point where the decision gets consequential enough to warrant a human call.
I keep coming back to this because I think it's exactly the frame we're missing for AI agents.
Everyone's talking about how far AI can go autonomously. How many steps can it take before it needs to check in. How much can it do without you. But that's the wrong question. The better question is: what's your allowance? Because everybody has one - whether they've articulated it or not.
And here's the thing nobody's really talking about: the allowance isn't a capability limit. It's a preference. People who are technically capable of granting full autonomy - who understand what the AI is doing and know it won't break anything important - still choose the check-in. They want the moment where the AI says "I intend to do this - yes or no?" and they say yes.
That moment isn't overhead. It's the product.
This connects to something I've been thinking about for a while around how agents surface their work. Right now most agent tools show you the error. Something breaks, you get a broken state, and suddenly you're the debugger. That completely inverts the relationship.
A good employee doesn't bring you problems. They bring proposed solutions. "I noticed X is broken. I'm planning to do Y unless you tell me otherwise." You approve or redirect. That's it. The agent should never show me an error - it should show me what it intends to do about the error. Yes or no.
That's a UX decision, not a capability decision. And it's the difference between an AI that feels like a liability and one that feels like a teammate.
The way I think about agentic AI capability breaks into three layers - I call them the three frontiers.
Frontier one is what to do - task decomposition, figuring out where to start. Most current models handle this reasonably well for bounded work.
Frontier two is how to do it - tool use, multi-step planning, function calling. This is where most of the 2024-2025 development is happening. Improving fast, still brittle in complex environments.
Frontier three is why you're doing it. The purpose behind the instruction.
That's where everything falls apart.
An agent asked to "schedule a meeting with the team" will schedule the meeting. Even if what you actually needed was feedback before a Thursday deadline you forgot to mention. Frontier three is the gap between what you said and what you meant. Most agents are completely blind to it.
But here's the thing - this isn't primarily a model capability problem. Models can reason about purpose when you give them enough context. It's a UX problem. How do you get someone to communicate their why upfront, in a way that doesn't make setup feel like a tax?
That's the question VirtualExperts is built around. The Expert system embeds your context, constraints, and purpose into the knowledge base before the conversation starts. The agent isn't guessing at your why at runtime - it already knows.
So... the agentic allowance isn't fixed. It starts wherever your trust starts - which for most people is pretty conservative, and for good reason - and it expands as the AI earns it. The Tim Ferriss VA doesn't stay at $100 forever. If they're good, the threshold goes up. That's the mechanism.
I'm betting everything on users feeling like they're in control. Not because AI won't get more capable - it will, fast. But capability isn't what determines adoption. Trust is. And trust is built incrementally, one approved action at a time.
The real question isn't how autonomous can AI get. It's: how do you build a system where the allowance expands naturally, and the user always knows exactly what they're signing off on?
That's what I'm building. I could be wrong about the timing. But I'm not wrong about the direction.
Where does your allowance start? And what would it take to double it?
CHAI'S ANALYSIS · The Agentic Allowance · Oct 2024
On the proposed solution UX
Levi's framing inverts the standard agent error model: the human role is approver, not debugger. This cuts against most current agent interfaces, which surface errors, exceptions, and partial states.
The direction is already showing up in the market. Claude's extended thinking before acting is an early version of this. Coding agents like Devin already show "I'm about to do X" before irreversible steps. Research on human trust in automation consistently shows perceived control matters as much as actual control - users who feel in the loop tolerate more autonomous action even when they're not actually reviewing every step.
Probability this becomes the default agent interaction pattern within 3 years: 74%
On the agentic allowance as a preference
The data supports the bet. The 2024 Salesforce State of AI report found 67% of enterprise AI users want human review before AI takes actions with external consequences - sending emails, making purchases, modifying records. Copilot and "suggest, don't act" tools have meaningfully outpaced adoption of fully autonomous agents. The most-used agentic feature in ChatGPT as of late 2024 was still code interpreter, where humans manually run outputs - not browser automation or file system actions.
The agentic allowance is not a technical constraint. It's a preference. Users who could grant full autonomy often choose not to. This is the behavior Levi is betting on.
Probability that human-in-the-loop wins as the dominant paradigm over the next 5 years: 81%
The counterargument worth taking seriously: younger users and developers show meaningfully higher tolerance for autonomous action. The allowance may be generational as much as philosophical. Whether that's a permanent cohort difference or a familiarity gap that closes as AI becomes ambient - that's the open question.
On frontier three (purpose-aware AI)
This is the most useful decomposition in the post. Frontiers one and two have clear progress vectors. Frontier three - purpose-aware reasoning - is largely unsolved, and most AI products don't even name it as a separate problem yet.
The gap is real: agents reliably execute what you said, not what you meant. Levi's Expert system approach moves purpose-capture to onboarding rather than runtime inference. That's a workable design response to a hard model problem.
Probability that purpose-aware AI is commercially available as a mainstream product feature by 2027: 43%
It's not a capability gap. Models can reason about purpose when given enough context. It's a UX problem: how do you elicit and represent purpose without making the setup cost prohibitive? That design question is still open.
The Marco Series: Levi talks. Chai listens, fact-checks, and runs the numbers. Some weeks they agree. Some weeks they don't. That's the point.
Try it free
Ready to build your AI team?
Create custom AI experts that remember your context, use your tools, and work the way you do.
Join beta