Smarter procurement: AI agents that request quotes, compare suppliers and recommend

The problemYour purchasing team spends more time chasing quotes than negotiating.

Tuesday, 4:40 p.m. Andrea, purchasing manager at an industrial maintenance company, has three email tabs open, a spreadsheet with twelve columns and a message from the plant floor: "Did the bearings order go out yet?" She requested quotes from five suppliers on Thursday. Two replied. One sent a PDF with no delivery date; the other quoted a different model. The line goes down on Friday if the parts don't arrive, so Andrea does what she always does: she buys from the supplier she trusts, knowing she is probably overpaying.

This scene plays out every day in thousands of companies. Procurement with AI agents isn't about replacing Andrea. It's about taking everything off her plate that doesn't require judgment: requesting, chasing, organizing and comparing. That frees her to decide and negotiate.

Why procurement turns into a bottleneck

The problem is rarely a shortage of suppliers or budget. It's that purchasing is made of small steps nobody wants to do, and nobody notices until they break.

• Quotes arrive in every format imaginable. A PDF, an email with the price buried in the text, a photo of a price list. Comparing them means retyping them.
• Terms matter as much as price. Lead time, payment terms, warranty, freight included or not. A cheap quote that requires prepayment can end up expensive.
• Supplier history lives in someone's head. The fact that a vendor "always ships late in December" isn't written down in any system. It's what we call tribal knowledge in another article.
• Approvals stall. The order waits for a signature from someone stuck in a meeting, and nobody knows which step it's on.

When time gets tight, the team stops comparing and buys out of habit. That's the hidden cost.

How procurement with AI agents works

An agentic procurement system isn't a chatbot that "helps you buy." It's a set of agents, each with a specific job, working from the same business memory and following the rules you define. The structure around the model is called a harness: the scaffolding that gives the agent tools, context, limits and checkpoints.

In procurement, the typical workflow has four parts:

1. Request agent. Captures the need (from the plant floor, from the ERP when inventory drops below a minimum, or from an email) and turns it into a clear request: what, how much, by when and to what spec.
2. Quote agent. Identifies the approved suppliers for that category, sends them the request, follows up with anyone who doesn't respond and extracts the data from every reply, whatever the format.
3. Comparison agent. Puts every offer in the same table and scores them against your criteria: total cost, lead time, payment terms, delivery track record and quality. It doesn't invent weights; it uses the ones your team set.
4. Recommendation agent. Writes up a recommendation with the suggested option, the alternatives and the reasoning, and prepares the purchase order ready for signature.

Above all of this sits the most important rule: no order goes out without approval from the right person.

Spend-based approval: the gate that builds trust

The most common fear among operations leaders is "What if the agent buys something it shouldn't?" The answer isn't to trust the model more. It's to design the approval points well. In the harness, every purchase order passes through a gate: a checkpoint that only opens with a signature from a defined role.

A typical setup, as an illustrative example:

You set the amounts. What matters is that the rule is written into the system, not stored in someone's memory, and that every approval is logged: who signed, when and with what information. We go deeper on this design in human approval by role.

A practical framework to get started

You don't need to automate all purchasing at once. The sensible move is to start with one category and do it well.

1. Pick a repetitive category. Spare parts, consumables, packaging. Frequent purchases, multiple suppliers and clear specs.
2. Write down your criteria. What matters most: price, lead time or track record? Which terms are deal-breakers? If your team can't agree, that's your first finding.
3. Gather the history. Past orders, late deliveries, complaints. Even if it's messy, the system can structure it and store it in each supplier's profile.
4. Set approval thresholds. By amount and, where relevant, by spend type or new supplier.
5. Measure before and after. Time from request to purchase order, number of quotes compared per purchase and on-time deliveries.

Illustrative example: before and after

Back to Andrea. Imagine her company rolls out an agentic procurement system for the bearings and spare parts category.

Before. The request comes in by text message. Andrea emails five suppliers, waits, follows up by phone, retypes two quotes into a spreadsheet and, under time pressure, buys from the usual vendor. Nobody records that a third supplier replied late with a better price.

After. The plant logs the need. The quote agent sends the request to approved suppliers within minutes and nudges the ones who haven't replied the next morning. When the responses arrive, the comparison agent builds the table with total cost, lead time, terms and historical performance. Andrea receives a recommendation: Supplier B, because even though it's slightly more expensive, it delivers in three days and Supplier A has been late on its recent deliveries. Andrea reviews it, changes a payment term, signs, and the order goes out.

Andrea didn't stop buying. She stopped retyping. Now she spends her time on what actually needs a person, like renegotiating with Supplier A or sourcing a new vendor for a problem category.

Common mistakes when automating procurement

• Automating without written criteria. If nobody has defined what "the best purchase" means, the agent will optimize for whatever is easiest to measure, which is usually just price.
• Not connecting to finance. Purchasing affects cash flow and accounts payable. A good design coordinates with your finance, collections and reconciliation processes.
• Leaving memory out. If the system doesn't remember how each supplier has performed, it compares quotes, not suppliers.
• Skipping observability. You should be able to see which step each request is on, what each supplier replied and why an option was recommended. Without traceability there's no trust and no audit trail.
• Not capping the system's own spend. Beyond purchasing spend, the harness controls what each agent run costs, so operating costs stay predictable.

What to do on Monday

Ask your purchasing manager for the last twenty orders in one category, with three data points for each: how many quotes were compared, how long it took from request to purchase order, and whether it arrived on time. That list will quickly tell you whether your problem is process, information or approvals.

At Addabra we build custom agentic systems for functions like procurement: we diagnose the process, design the agents and approval gates, connect them to your tools and support them in production. Your team keeps the judgment and the signature; the agents handle the requesting, chasing and comparing.

procurementsuppliersAI agentshuman approval