PrymeIntelligence

Finance teams·Playbook

How to let an AI agent raise purchase requisitions

Let an AI agent raise purchase requisitions for routine replenishment, with a named human requester, budget checks, budget holder approval and a full log.

An agent can raise requisitions at depth 3 for routine replenishment, such as consumables reordered against usage, because a requisition is an internal request that commits nobody until it is approved. Every requisition still names a human requester who answers for the need, goes through the normal budget holder approval and records the data that triggered it.

Why is a requisition a safe place to start?

A requisition asks for permission to buy. It sits before the purchase order, which is the commitment to a supplier, and before the invoice and payment. Nothing leaves the business when a requisition is raised, and it can be cancelled without anyone outside knowing. That makes it a low-risk action for an agent, provided the controls after it stay intact.

The requester owns the need. The budget holder owns the decision to spend. Procurement owns how the need is met: which supplier, which contract, which price. An agent raising requisitions takes over the requester's typing, not the requester's accountability.

Where the agent instead works on requisitions people have raised, checking and routing them, see how to process purchase requisitions with an AI agent.

Which permissions do you grant per action, and at what depth?

  1. Agent, depth 1: reads stock levels, usage, open orders and the catalogue. It needs to know what is already on order before asking for more.
  2. Agent, depth 2: drafts a requisition for a requester. For anything outside the routine list, the agent prepares it and the named requester submits.
  3. Agent, depth 3 once certified: submits routine requisitions. Items on an agreed replenishment list, from the catalogue, within a quantity range and a value limit per requisition and per month.
  4. Agent, held: quantities outside the usual range, items not on the list, or requisitions that would exceed the remaining budget line. The named requester confirms or edits.
  5. Agent, refused: approving requisitions. Approval stays with the budget holder through the normal workflow.
  6. Agent, refused: creating purchase orders from requisitions it raised. If you also let an agent create purchase orders in your ERP, use a separate capability and keep the budget holder's approval between them.
  7. Person, named requester: answers for each requisition raised in their name.
Action The agent does Depth Stays with a person
Read demand data Stock, usage, open orders 1 Nothing
Non-routine request Drafts it 2 Requester submits
Routine replenishment Submits within limits 3 Requester owns the list
Unusual quantity Holds with the data held Requester confirms
Approve Nothing refused Budget holder
Convert to order Nothing in this role refused Buyer or separate capability

Why must a person be the requester of record?

Budget accountability works because a named person asked for the spend. If the requester field shows the agent, nobody answers when the budget holder asks why forty boxes of gloves were needed. Set each routine list against a named requester, usually the site or department lead, and show "raised by agent on behalf of" in the record. That person reviews the list of items and ranges periodically and can withdraw an item at any time.

What is recorded for each requisition?

The filed record holds the trigger (stock below reorder point, a usage forecast, a scheduled replenishment), the data the agent read, the quantity rule applied, the named requester, the limits in force and the verdict. When an approver queries a requisition, the answer is one lookup away.

The ERP record also matters. The agent should write the trigger into the requisition's justification field, in plain words: "stock 20, reorder point 50, average weekly use 30, two weeks' cover".

Which controls matter, and what evidence does each leave?

Open-order check. The agent does not raise a requisition for an item already on order in sufficient quantity. Evidence: open orders read at the time, on the verdict.

Quantity range. Each item has a minimum and maximum per requisition. Evidence: the range version used.

Monthly value cap. Total value raised by the agent per cost centre per month. Evidence: running total on each verdict.

No split requisitions. One need is one requisition, not several under an approval threshold. Evidence: linked requisitions flagged together.

Revocation. The requester or procurement can remove items from the routine list, or the capability entirely, with immediate effect. Evidence: list changes in the log.

What should you measure?

  • Requisitions rejected by budget holders, and why.
  • Stock-outs on items the agent manages, compared with before.
  • Excess stock on those items, the opposite failure.
  • Held requisitions by reason.
  • Duplicates caught by the open-order check.

Where does it usually go wrong?

Usage data is wrong. Write-offs and returns are booked as usage, the agent sees high consumption and orders more. Have the requester sanity-check the usage source when the list is set up.

The list grows by default. Items are added "while we're at it" and the routine class quietly becomes most purchasing. Review the list quarterly against spend.

Approval becomes a formality. Budget holders approve agent requisitions without reading because they are always fine. Send approvers a monthly summary of what they approved, by item, so patterns are visible.

Requisition and order merge. In some setups a catalogue requisition turns into an order automatically once approved. That is acceptable only if approval is real. Whether an agent can raise a purchase order directly is a separate decision, and so is whether it can approve purchase orders.

Where Pryme fits

In Pryme, drafting and submitting requisitions are separate capabilities, and submission at depth 3 is earned on your own past replenishment data. Quantity ranges and caps are your limits in the minimum chain, and each verdict names the bound that decided. The finance bounds behind this are in AI agents for finance.

Questions people ask

Can the agent raise requisitions for services?

Rarely at depth 3. Services need a description of scope that someone has to decide on. The agent can draft them for a requester.

What if the budget is exhausted?

The agent holds the requisition and tells the requester. It does not raise it against a different budget line.

Can staff ask the agent to raise a requisition for them in chat?

Yes, as a draft at depth 2. The staff member reviews and submits it in their own name.

Decide what your agents may do. Prove it. Enforce it.

Pryme Intelligence certifies each agent on your own cases, checks every action against five bounds at the moment it happens, and files every verdict.