Photo by Tima Miroshnichenko on Pexels
Your warehouse lead needs $3,800 of replacement parts before a customer job on Thursday. She fills in a purchase request and emails it to her manager. Her manager is at a site visit, so the request waits. On Tuesday someone forwards it to the owner “just in case.” By Wednesday she has ordered the parts on her own card to keep the job on schedule, and the purchase order gets written afterward so AP has something to match the invoice to.
Nobody did anything wrong. The approval just had no clear owner, no clear limit and no clock on it. That is the problem a purchase order approval workflow is meant to solve, and it is worth solving before you add any software to it.
Where PO approvals get stuck
In most small and mid-sized businesses, purchase approvals live in inboxes and hallway conversations. The same few failure points show up again and again:
- No written thresholds. Everyone knows “big purchases need the owner,” but nobody knows whether big means $1,000 or $10,000, so people either over-escalate or skip approval entirely.
- One approver, no backup. When the person who signs is traveling or out sick, requests sit until they return.
- Requests arrive incomplete. No vendor, no quote, no job or cost code. The approver replies with questions, and the thread grows.
- The PO is written after the purchase. Once that becomes normal, the PO stops being a control and becomes paperwork.
- Nobody can see the queue. There is no answer to “what is waiting for approval right now, and for how long?”
For a sense of what “fast” can look like, APQC’s Open Standards Benchmarking reports a median of 2.0 days from purchase requisition to purchase order, across a sample of 1,250 organizations. If your requests routinely sit for a week, the delay is not the cost of doing business. It is a process with gaps in it.
Fix the process first: what a clear PO workflow needs
Automating approvals that have no rules just moves the confusion somewhere faster. Before any agent touches this, write down five things:
- Dollar thresholds by role. For example: team leads approve up to one amount, the operations manager up to a higher one, and anything above that goes to the owner. The numbers are yours to choose; what matters is that they exist on paper.
- A backup for every approver. Named in advance, with a rule for when it kicks in (say, no response in one business day).
- A minimum request. Vendor, item or service, amount, quote attached, and the job, department or budget line it belongs to. Incomplete requests go back to the requester, not to the approver.
- Which spend skips the PO. Utilities, rent, subscriptions under a set amount. Saying so explicitly keeps the PO meaningful for everything else.
- An emergency path. Real emergencies happen. Allow an after-the-fact PO, but require it within a set number of days and track how often it’s used.
That’s kAIzen: process first, then AI. Map how approvals really happen today, not how the handbook says they happen, and write the rules against that.
What a proactive agent does once the rules exist
With the rules written down, the work that is left is mostly remembering, checking and chasing. That is the part an agent can carry. It reads the signal already in your business (the request email, the attached quote, the budget in your accounting system, the vendor list) and acts on it:
- Checks the request is complete and sends it back to the requester with exactly what’s missing, before it reaches anyone’s approval queue.
- Routes it to the right approver based on your thresholds, and moves it to the named backup when the clock runs out.
- Compares it against what it already knows: the remaining budget on that job, the last price paid to this vendor, whether the vendor is new, whether the quote matches the request.
- Drafts the PO in the system you already use once approval is given, so the PO exists before the order is placed and AP has something to match the invoice to later.
What reaches the approver is not a dashboard to monitor. It is the exception, with the reason attached: “Request from the service team for $6,200 to a vendor you have never used, $1,400 over the remaining budget on job 214. The last comparable order went to your usual supplier at about 12% less.” Routine requests inside policy simply move.
Compare that with tools that wait for someone to log in, open a queue and remember to ask what is pending. That is a carburetor: it does less and demands more. A proactive system works like fuel injection. You state the intent (these are our limits, this is who approves), and it does the dosing in the background.
The PO is also the first leg of the match that protects you later. Our guide to accounts payable automation for small business shows where approvals fit in the whole AP flow. When the invoice arrives, it gets checked against the PO and the receipt, which we cover in three-way match automation.
The honest part: autonomy is earned
An agent that routes purchase requests should not start out deciding anything. The first weeks look like this:
- It suggests. It reads each request and proposes the route and the flags, and a person confirms or corrects. Every correction teaches it where your real rules differ from the written ones.
- It executes with review. Once its suggestions are consistently right, it routes on its own, and someone spot-checks a sample each week.
- It executes alone for the routine path, with every exception still going to a person.
Some things never leave a person’s view. The approval itself is a commitment of money, so the agent prepares and routes; a human approves. The same goes for new vendors, changes to vendor bank details, and anything with contractual weight. An ECU manages the mixture. It doesn’t choose the destination. We go deeper on this in our post on human-in-the-loop AI approvals.
If your approval limits touch internal controls your accountant or auditor relies on, confirm the thresholds and segregation of duties with them before you change anything.
Where to start this week
Pull the last month of purchase requests and note three things for each: how long it waited, who approved it, and whether the PO was written before or after the order. That one list usually shows where the workflow breaks, and whether it is a rule problem or a chasing problem.
If the chasing is eating hours every week, put a number on it with our ROI calculator. It will tell you whether an agent is worth building for this, or whether written thresholds and a backup approver are enough on their own.
Want to see this in your company?
Schedule a free discovery call with RubikSoft and learn how an AI Agent can transform your operations in weeks.
Schedule Your Free Consultation