AI in procurement

Human approval for AI agents in procurement

Updated 5 min readMarkdown
In short

An AI system in procurement may read, compute, prepare and draft on its own, because none of that has an effect outside the organisation. A person approves everything that does have one, which means a message to a supplier or an employee, a change at a vendor, the submit of a web form, the confirmation of an offer and an award. The approval point works when the person sees exactly what will go out, approves that one object, holds the permission for it, and the approval is recorded. What goes out by itself later, such as a reminder, was shown and approved before.

An AI system in procurement can be left alone with everything that has no effect outside the organisation: reading, computing, preparing, drafting. It needs a person's approval for everything that does have one: a message to a supplier or an employee, a change at a vendor, a submitted form, an award. The line is drawn by what an action does. A more capable model does not move it.

Where the line runs

A message to a supplier speaks for the organisation that sends it. A removed seat takes a tool away from a colleague. A submitted web form cannot be recalled. So two questions decide each action: who outside notices it, and can it be taken back?

ActionWho does itReason
Read invoices, seat lists, usage data and repliesThe system, aloneReading changes nothing. What a model read is labelled until a person has reviewed it
Compute a finding and its savingDeterministic code, aloneSame inputs, same result, with the calculation shown
Prepare the next step, draft a messageThe system, aloneA draft has no effect until someone sends it
Send a message to a supplier or an employeeThe system, after a person approved that messageIt speaks for the organisation
Change something at a vendorThe system, after a person approved the step, and only where the organisation allows such changesIt affects colleagues and contracts
Submit a form on a websiteAn agent, after a person confirmed the filled-in formA submit cannot be recalled
Count an offer that a model read from a replyA person confirms it firstA misread price would steer the comparison
Award to a supplierA personIt is a commitment

How to design the approval point

An approval is a check only when the approver can see what is being approved. Six rules make it one.

  1. Show the thing itself. The full text, every recipient, every value that will be entered. A line such as "12 messages ready" cannot be approved, only waved through.
  2. Approve one object. The approval belongs to this message or this step. If the text changes afterwards, it needs approving again.
  3. Tie it to a permission. Sending to suppliers, sending to employees and running a change at a vendor are separate permissions held by named people.
  4. Offer a second person. Where an organisation wants four eyes, whoever drafted something cannot approve it.
  5. Check again when it is sent. A do-not-contact list or the state of a sourcing event may have changed between the draft and the send.
  6. Record it. Who approved what and when, and what the system did by itself afterwards, marked as automatic.

What may go out by itself

Automation after an approval is legitimate when it adds nothing new. A reminder that repeats an approved question in fixed words, on a date shown together with the original, was approved with it. The same holds for passing an unanswered question on to the person's manager, provided the approver saw that step, its text and its date.

A follow-up to a supplier is a different case, because it is a new message in a negotiation. A workable rule is that every follow-up is a draft, except the first one, which repeats the approved message in fixed words and goes out alone only where the organisation has switched that on. The article on following up with suppliers covers timing and wording.

Nothing should happen because somebody did not answer. An employee who ignores two questions has not released a seat. A supplier who stays silent has not declined. After the last planned follow-up the matter returns to a person, who calls, tries another contact or closes it. A system that treats silence as an answer has taken a decision nobody approved.

Limits on what is prepared

Preparation needs limits as well, or the pile of drafts grows beyond what anyone reads. Three limits work in practice: prepare only for findings that rest on direct evidence, only above a threshold of worth, and only a fixed number per run. A person who is handed fifty drafts a day approves them without reading, and then the approval point exists on paper only.

What goes wrong

  • Blanket approvals. "Always allow messages to this supplier" turns an approval into a setting. The next message says something nobody read.
  • Approving the summary. The approver sees a count and a button, and the text is three clicks away.
  • Figures written by the model. A draft that states a price or a saving of its own has to be discarded before it reaches the approver. The article on why AI must not invent figures explains the guard.
  • Approvals in five places. A banner here, a mail there. One decision list ordered by deadline and worth keeps the approvals short and the important ones on top.
  • An agent without a stop. An agent that fills in a vendor's website and submits in one go leaves no moment for a second look. See browser agents for vendor portals.

How Cureman does it

In Cureman the connectors to vendors only read. The autopilot prepares the first step of a new finding: a draft campaign, a drafted message to a vendor or supplier, a draft sourcing event. It never approves, sends, changes anything at a vendor, switches a policy or books a module. Nothing goes to employees without a person approving the draft, and the reminder and the message to a manager are shown with the campaign and approved with it. Nothing goes to a supplier without a person approving that message. The one exception is a single reminder that repeats an approved request in fixed words. A change at a vendor is a step of an action plan that is prepared, shown in full and approved by a person, and it is off unless the organisation allows it. An agent stops before a submit by default until a person confirms. With the four-eyes rule switched on, nobody approves what they drafted themselves. Everything that waits for a person is in one list of decisions, and every change is written to the audit log. The rules are listed under principles, and how it works shows them in the daily routine.

Questions and answers

What may an AI agent in procurement do without approval?

Everything that stays inside the organisation and has no effect on anyone outside it. That covers reading invoices, seat lists and replies, computing findings, preparing the next step and drafting messages. The results are labelled as prepared or as read by the model until a person has looked at them.

Which actions always need a person's approval?

Every action with an effect outside the organisation. These are sending a message to a supplier or an employee, changing something at a vendor, submitting a form on a website, counting an offer that a model read from a reply, and awarding. The person approves the exact text and values, never a summary of them.

Can a reminder go out without a new approval?

Yes, if it was approved beforehand. A reminder that repeats an approved request in fixed words, on a date the approver saw, adds nothing the person did not agree to. A follow-up with new wording or new figures is a new message and needs its own approval.

What does the four-eyes rule mean for steps an AI prepared?

Under a four-eyes rule the person who drafted a message or a step cannot approve it, so a second person with the permission has to. A draft the system prepared has no human author, so any person who holds the permission can approve it. The rule is most useful for messages to suppliers, changes at a vendor and tasks an agent carries out in a website.

Does human approval slow procurement down?

It costs one decision per outgoing step, so the design has to make that decision quick. Put everything that waits in one list, order it by deadline and worth, and show the full text one step away. Work slows down when approvals are scattered over several tools or are demanded for steps that have no outside effect.