The most common objection we hear about agentic operations is not about capability. It is about control: if we turn this on, what stops it from doing something we did not want?

It is the right question. Our answer is that autonomy should be a dial you set per workflow, not a switch you flip for the whole business.

Four levels, one boundary

Agents on the Splenta platform run at the level you configure:

  1. Observe. The agent reads the record and reports what it would do. Nothing changes.
  2. Propose. The agent prepares the action, such as a draft journal or a recommended approval, and a person carries it out.
  3. Act with confirmation. The agent performs the action once a person confirms it.
  4. Act within policy. The agent performs routine actions on its own and escalates anything outside the policy you set.

What makes these levels meaningful is how the boundary is enforced. It is not a line in a prompt asking the model to behave. It is a scoped token. An agent at the propose level simply does not hold the scope to write, so no amount of clever reasoning gets it there.

Some actions stay human by default

Certain steps carry liability that should not move to an agent just because it would be convenient. Approving spend, releasing a clinical order, posting to the ledger, touching a payment.

On the procurement record, for example, the approve step is withheld from agents by default. Arin can read the pending purchase order queue, check each order against the same rules the workbench enforces, and surface the ones that need a decision with the detail attached. A person still owns the actual approve or reject on anything that crosses that line.

In finance, agents are built to triage accounts payable match exceptions without ever touching payment. The design assumption is simple: agents clear the routine majority and prepare the rest. They do not quietly absorb both.

Start narrow, widen with evidence

Teams that adopt this well tend to follow the same pattern:

  • Pick one workflow where manual work hurts most. Usually a reconciliation or a queue that never drains.
  • Run the agent in observe mode first. Compare what it would have done with what people actually did.
  • Read the escalations. The quality of what the agent hands back is the best signal of whether it is ready for more.
  • Widen the envelope one scope at a time. Every change is a configuration change, recorded like any other.

Because every action, and every denied action, lands on the same audit trail a person's would, the evidence for widening autonomy builds up as a by-product of running it.

Why this matters

A platform you can only adopt all at once is a platform nobody adopts. Treating autonomy as a setting means the first step is small, reversible and measurable, and each step after that is a decision you make with data rather than faith.

Read more about how Arin, PECO and Semorel apply this across continuous operations, in-app automation and agent access to your own product.