The autonomy dial: approve every move, or automate the routine

The false choice everyone is arguing about

Most conversations about AI at work collapse into two camps. On one side: AI that does nothing without you. It drafts, it suggests, it waits, and you click approve on everything, forever. Safe, but it recreates the bottleneck you were trying to remove. On the other side: fully autonomous agents that run your business while you sleep. Exciting in a demo, terrifying the first time one emails a customer something you would never have said or moves money you did not authorize.

Both extremes are real, and both are wrong for how a business actually runs. The all-manual version means you have hired an assistant who cannot do anything you have not personally watched them do. The fully-autonomous version means you have handed the keys to a system whose judgment you have not yet earned the right to trust. Neither matches how you would onboard a human employee, and neither should be the default for software either.

The honest answer is a dial, not a switch. You do not decide once, globally, how much freedom your AI workforce gets. You decide per task, per risk level, and you move the dial as evidence accumulates. Some things you will always approve by hand. Some things you will let run on their own within limits. Most things live in between, and they migrate over time as you learn what the system gets right.

The default is approve-every-move, and that is a feature

At Kirality, the starting position for every new tenant is simple: the AI drafts, you approve. Agents work inside your own tools, your CRM, your inbox, your docs, your project tracker, and they prepare the action but do not commit it. A follow-up email sits as a draft. A CRM update waits in a review queue. A status change on a project is proposed, not applied. Nothing leaves your control until you say so.

This is not a limitation to apologize for. It is the fastest, safest way to learn whether a given task is one you can eventually hand off. Every approval is a data point. When you find yourself clicking approve on the same kind of draft over and over without changing a word, you have discovered a candidate for automation. When you routinely rewrite or reject a certain kind of action, you have learned exactly where the AI is not ready, and you have caught it before it did any damage.

Approve-every-move also does something subtler: it keeps you fluent in your own operation. You see the drafts, so you see the patterns, the edge cases, the customers who need a human touch. Automation that hides all of this from you does not make you a better operator. The default is designed so that when you do decide to dial something up, you are making an informed decision, not a hopeful one.

Standing mandates: handing off the routine within limits you set

Once a task has proven itself in the approval queue, you can grant a standing mandate. A mandate says: for this specific kind of work, within these specific limits, you do not need to ask me first. The AI can act on its own, and you review after the fact instead of before. This is where the time savings actually show up, because the highest-volume, lowest-judgment work stops routing through your attention entirely.

The critical word is limits. A mandate is never a blank check. You define the boundaries: which task, on which records or segments, up to what threshold, during what window, and what happens at the edge of the boundary. Categorize and tag inbound leads automatically, but flag anything that looks like an enterprise deal for a human. Keep project statuses in sync, but escalate anything blocked for more than a set period. The mandate encodes your judgment as rules the system follows, so hands-off does not mean unattended.

Mandates are also reversible and observable. You can narrow one, pause it, or revoke it entirely the moment something feels off, and you can watch what it has been doing without waiting for a report. The goal is not to set it and forget it. The goal is to move routine work off your desk while keeping a clear line of sight and a fast path back to manual control. Dialing up is a decision you can always undo.

The lines that never move: money, comms, and legal

Some categories of action are simply not eligible for hands-off automation, no matter how much the AI has proven itself or how much you would like the convenience. Anything that moves money, anything that speaks to the outside world in your name, and anything with legal weight always keeps a human in the loop. These are not defaults you can override in settings. They are hard boundaries.

The reasoning is about asymmetry of consequences. A miscategorized lead is a small, easily-corrected mistake. A payment sent to the wrong party, a contract signed on your behalf, or a message that damages a customer relationship is not something an audit trail can undo after the fact. The value of automating those actions is real but modest; the cost of getting one wrong is severe and sometimes irreversible. When the downside is catastrophic and hard to reverse, a human approves it. That trade is not close.

This also protects the parts of automation that do make sense. Because the genuinely dangerous actions are fenced off permanently, you can be far more relaxed about dialing up the routine ones. You are not one bad rule away from a wire transfer or a legally binding commitment. The dial has a hard stop, and knowing exactly where that stop is makes the rest of the range safe to use.

Every setting rides on a tamper-evident audit trail

Whether you approve a move by hand or a standing mandate runs it for you, the action is recorded on an append-only, tamper-evident audit trail. Every draft, every approval, every mandated action, and every boundary the system hit is logged with who or what did it and when. This is what makes the dial trustworthy: dialing something up is not a leap of faith, because you can always reconstruct exactly what happened and why.

Tamper-evident is a deliberate and honest claim. It means the record is append-only and any attempt to alter history is detectable, not that the log is a magically complete account of the universe. That distinction matters. The audit trail gives you a defensible, reviewable history you can hand to a partner, an auditor, or your future self when you are trying to understand why an agent did what it did three weeks ago.

Two more things sit underneath all of this. Per-tenant isolation means your data and your agents live in your own space, not commingled with anyone else's. And bring-your-own-key means the underlying model runs on your Anthropic, OpenAI, or Bedrock credentials, so you keep control of the provider relationship and the cost. The autonomy dial is only as trustworthy as the foundation it sits on, and the foundation is designed to be inspected, not taken on faith.

How to decide what to dial up, and when

A practical way to think about the dial is a short set of questions per task. How reversible is a mistake, and how expensive is it to fix? How much does this task vary, or is it nearly the same every time? How well has the AI performed on it in your approval queue so far? High reversibility, low variance, and a strong track record point toward a mandate. Any one of those being weak is a reason to keep approving by hand a while longer.

Start narrow and widen slowly. When you grant your first mandate for a task, scope it tightly, a single segment, a low threshold, a short window, and watch the after-the-fact reviews. If the results hold, widen the boundary. If they wobble, tighten it or pull it back to manual. There is no penalty for moving the dial back down, and treating every mandate as revisable rather than permanent is exactly the right posture. The system is built to let you adjust, not to lock you in.

Over time, a healthy operation ends up with a spread. A handful of high-stakes tasks stay firmly manual by design. A growing set of routine, well-understood work runs under mandates within limits you trust. And a middle band lives in the approval queue, either on its way to a mandate as it earns confidence or holding there because the judgment involved is worth your attention. That distribution is not a failure to fully automate. It is what a business run well by a human and an AI workforce actually looks like.