When an agent should ask for approval

An illustrative order workflow; not a report of a deployed client system.

Suppose our usual replenishment order is between 100 and 200 units. If an agent proposes 500, I want it to bring that decision to a person before placing the order.

There may be a good reason for the larger quantity. Demand may have changed, or an upcoming event may require more stock. I would want the agent to investigate: check incoming stock, examine supplier lead times and compare the consequences of ordering different quantities. It could do considerable work to prepare the decision while still needing approval to commit the business.

As I think about connecting agents to our order management system, that distinction helps me decide what access to give them. I would start with a narrow responsibility and the tools and information needed to fulfil it. A replenishment assignment would specify which business unit and stock it covers. Permission to investigate within that scope would not include changing supplier details or taking stock allocated to another part of the business.

The usual quantity range would help define a review policy, but an order for 150 units could still be unnecessary. Suppose a supplier's EDI update confirms that stock will arrive earlier than expected while the agent is preparing its recommendation. If that stock now covers the need, the proposal should change. The information available when the agent started would no longer be enough to justify placing the order.

I would therefore want the system accepting the order to check current stock, outstanding orders, total commitments and the authority behind the request before execution. Those business checks should apply whether the request comes from a person, an EDI flow or an agent. Splitting one large purchase into smaller requests should not bypass the approval requirement. As more applications act on the same operation, the order system needs to keep those decisions consistent.

These checks belong in ordinary software, with permissions enforced behind the tools. A predictable replenishment calculation may also fit an existing workflow. I would use an agent where interpreting changing circumstances or investigating an exception helps someone reach a better decision.

At the beginning, I would keep people close to the work. Testing and supervised use should establish whether the agent takes the expected actions and whether the controls hold when it encounters an exception. I would include cases with missing information, changing stock and proposals outside its authority. Repeating straightforward orders successfully would give us only part of the evidence needed to reduce supervision.

I would also examine the decisions it brings back to people. Does it catch the unnecessary 150-unit order? Does it explain a justified request for 500? Can it identify what it could not verify? An agent that asks for approval on every ordinary order would leave us with a different problem: a review queue that still consumes the team's day.

Once a frequent task had proved dependable within its agreed limits, I would want routine work to proceed without individual approval. For something happening a hundred times a day, I would track unnecessary interruptions and the time people spend investigating exceptions, alongside the quality of the orders themselves. Delegation should give the team time back while preserving the decisions that need their judgement.

I would keep records, automatic checks, alerts and periodic review, including samples of routine work. Changes to the model, tools or business conditions could justify closer supervision again. Some consequential actions would continue to require approval every time, especially when difficult to undo. For infrequent decisions, keeping a person involved may offer a sensible balance between effort and control.

The quality of each request for approval would matter too. If I have to reconstruct the stock position and search for the source of a demand change myself, much of the investigation remains with me. I would expect the agent to bring the relevant records, explain the alternatives and make any uncertainty visible. That would let me spend my attention on deciding what to do.

If an agent brings me a 500-unit order, I want to understand what changed, what is already committed, and what happens if we stay at 200. Preparing that decision is useful work, even when a person makes the final call. Over time, I would judge the agent by the routine work it handles reliably and how much easier it makes the difficult decisions that still belong to us.

← Back to thinking