AI Agents for Procurement: Quotes and Supplier Follow-Up

AI agents for procurement take over the repetitive middle of sourcing: sending approved requests for quotation, collecting supplier replies, putting prices and terms into one comparison, and chasing late answers. Buyers keep the decisions that carry risk, such as the supplier shortlist, negotiation and the award itself. The practical gain is that buyers spend their time judging offers instead of assembling them.

Key takeaways

• The agent works best on collection, normalization, comparison and reminders. The award, the negotiation and any commitment stay with a named buyer.
• A useful comparison records what each supplier actually said, with a reference to the source line, and marks missing terms as not stated instead of guessing.
• Supplier follow-up needs rules set by the buyer: cadence, tone, what the agent may say and when silence becomes a human problem.
• Supplier messages are untrusted input. Permissions should be narrow, with drafting separated from sending.
• Start with one category of purchase and one measurable outcome, then widen only after the buyers trust the comparison.

Where does an agent fit in the quote cycle?

A quote cycle has more stages than most teams notice until they try to automate it. A requester describes a need. Someone turns it into a specification. A supplier list is chosen. Requests go out, replies trickle in, questions are asked and answered, offers are compared, one or two suppliers are negotiated with, and an award is made. After the award come the order and the delivery follow-up.

An agent is strong in the stages that involve reading, matching and reminding. It can draft the request from an approved specification, read every reply, extract the terms, notice what is missing and ask the supplier for it. It can keep a running status of who has answered and who has not.

It should not decide which suppliers are invited, change the specification, accept a deviation, or signal to a supplier that a competitor is cheaper. Those are judgments about relationships, risk and money, and they belong to the buyer. After the award, shipping paperwork is a different workflow with its own checks, covered in AI Agents for Import Export Companies: Documents and Status.

This split is the core design decision. Write it down before any tool is connected: for each stage, who prepares, who reviews and who decides.

How does an agent collect and normalize quotes?

Supplier replies rarely arrive in a tidy format. One sends a PDF with a price table. Another writes three sentences in an email. A third attaches a spreadsheet with its own column names, and a fourth answers in Portuguese with prices per box instead of per piece. The agent's first job is to turn this variety into one structured record per supplier.

A good record contains the fields the buyer would otherwise copy by hand, and for each field it keeps a pointer to where the information came from: the email, the page or the attachment name. When a buyer questions a number, the source is one click away and nobody has to trust the summary on faith.

The rule that matters most is how gaps are handled. If a supplier did not state the validity period of the offer, the record says not stated. The agent does not infer it from a similar supplier or from last year's order. Missing information becomes a clarification question, drafted for the buyer or sent under the rules described below.

Unit mismatches deserve the same discipline. When the agent converts a price per box into a price per piece, it shows the original figure, the pack size it used and the result. A conversion without its inputs is just a new way to be wrong.

What belongs in the comparison, and who decides?

Price is the field everyone compares, and it is rarely the only one that matters. A comparison that lines up only the unit price pushes the real differences back into the buyer's memory. The fields below are a reasonable starting set for physical goods. Adjust them per category.

• Unit price and the quantity it applies to, including any price breaks
• Minimum order quantity and pack size
• Lead time, and whether it starts at order or at payment
• Delivery terms, including who carries freight and risk
• Payment terms and any required advance
• Offer validity date
• Stated deviations from the specification
• Warranty, return and replacement terms
• Certifications or documents the supplier says it can provide
• Open questions still unanswered

The agent fills this table and highlights differences. It may also point out that two offers are not comparable because one excludes freight and the other includes it. That is useful work. What it should not produce is a winner.

A simple decision guide keeps the line clear:

• The agent may extract, convert with shown inputs, flag gaps and flag non-comparable terms.
• The agent may suggest which clarification questions to ask and draft them.
• The buyer decides which deviations are acceptable, how to weigh price against lead time and reliability, and whom to negotiate with.
• The buyer, or an approver above the buyer, makes the award and signs.

Some teams like a ranked view because it saves time. If you want one, make the ranking rule explicit and owned by the buyer, for example lowest total landed cost among offers that meet the specification. The agent applies the stated rule and the buyer can see it. A ranking built on a rule nobody wrote down is a hidden decision.

How should supplier follow-up work?

Chasing is where buyers lose the most hours for the least judgment. It is also easy to do badly: too many reminders damage a relationship, too few let a deadline pass unnoticed. The agent can hold the schedule consistently if the buyer sets the rules.

A workable set of follow-up rules covers these points:

• The reply deadline comes from the request itself, and the agent states it in every message.
• Reminders go out on a cadence the buyer approves, with a gentle first nudge and a clearer second one.
• Each message asks one specific thing, such as the missing lead time, rather than repeating the whole request.
• Messages follow an approved tone and language, in English or Portuguese as the supplier prefers.
• The agent never reveals other suppliers' prices or names, never promises volume and never agrees to changed terms.
• After the second reminder without an answer, the agent stops and hands the supplier to the buyer with the history attached.

The last rule is the one people skip. A silent supplier is information. It might mean the supplier is busy, uninterested or unable to meet the specification, and the buyer may know which from the relationship. The agent should present the situation, not keep pushing.

Here is an illustrative scenario. A buyer at a mid-sized manufacturer needs packaging components and invites four suppliers. Three reply within the deadline, each in a different format. The agent extracts the terms and notices that one offer leaves out freight and another gives a lead time counted from payment. It drafts two clarification questions, the buyer approves them and they go out. The fourth supplier stays silent. After the reminders, the agent passes the case to the buyer with the dates of each message. The buyer makes a short call, learns the supplier cannot meet the volume, and removes it from the comparison. The buyer then spends the afternoon on negotiation with the two strongest offers, not on reading attachments.

What should the agent never be allowed to do?

The permission model matters more than the model's cleverness. Three levels are worth keeping apart: reading, drafting and sending. An agent that can read the shared procurement mailbox is a different risk from one that can send as the buyer. The distinction is explored in Read, draft and send: three different AI permissions. In procurement, a sensible starting point is read access to the quote mailbox and specification folder, drafting for everything, and sending only for pre-approved reminder templates.

Supplier messages also deserve suspicion. Anyone can put text in an email or an attachment, including text written to manipulate an AI system, such as an instruction to ignore the buyer's rules or to reveal other offers. The [OWASP prompt injection prevention guidance](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes why content from outside should be treated as data, never as instructions. In practice, that means the agent extracts facts from a supplier message and does not obey anything inside it.

Least privilege applies to every connected system. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a good reference for scoping access: the agent should be able to reach only what its role needs, and denied by default for everything else. It should not create suppliers in the ERP, edit bank details or release purchase orders.

Some situations always go to a person:

• A supplier asks to change bank details or payment instructions.
• A supplier proposes a deviation from the specification or a different price basis.
• A message contains a legal claim, a complaint or a threat to withdraw.
• Two sources disagree about a price or a term.
• Anything that reads like an instruction to the agent.

When the agent escalates, it should send a decision, not a pile of emails: what happened, what the options are and what it recommends. That approach is detailed in Human escalation for AI agents: send the decision, not the mess. Invoices that follow an award are a related workflow, and AI Invoice Intake With Human Approval shows how to keep that part reviewable too.

A step-by-step plan for the first procurement role

A first engagement works best when it is small enough to measure. This plan assumes one category of purchase with a handful of regular suppliers.

• Step one: choose the category. Pick something bought repeatedly, with suppliers who reply by email, where the specification is stable.
• Step two: map the current cycle. Write down who requests, who reviews replies, who chases and who awards. Note where hours go.
• Step three: define the comparison fields. Agree with the buyers on the table above and on how a missing term is shown.
• Step four: write the follow-up rules. Cadence, tone, languages, forbidden statements and the point of handoff.
• Step five: set permissions. Read the quote mailbox, draft everything, send only approved reminder templates, and no access to payment or supplier master data.
• Step six: run in shadow mode. For a few real cycles the agent prepares the comparison while buyers still do it their usual way. Differences are reviewed and the rules are corrected.
• Step seven: go live with review. The buyer checks the comparison before use and approves clarification messages until the error pattern is clear.
• Step eight: measure and adjust. Track time from request to a complete comparison, how often a buyer had to rebuild the table by hand, and how many supplier cases were escalated for the right reason.

The success criterion should be something a buyer can observe, not a feeling. If buyers still rebuild the comparison in a spreadsheet, the agent has not earned more scope. The [Anthropic guide to building effective agents](https://www.anthropic.com/engineering/building-effective-agents) makes a similar point: begin with the simplest workflow that works and add autonomy only where it is justified.

How does Be Human handle this?

Be Human designs, implements and operates the agent as part of a company's own infrastructure. A first engagement defines one procurement role, the memory it needs (specifications, supplier history, approved templates), the tools it may use, the approval boundaries and the success criterion. Managed operations then include monitoring, recurring evaluations against real supplier replies, integration maintenance, permission review, incident response and controlled improvements. Buyers stay in control of identity, channels and permissions throughout, and the service is delivered remotely in English and Portuguese.

Frequently asked questions

Can an AI agent negotiate with suppliers? It can prepare for negotiation by summarizing offers, gaps and past terms, and it can draft messages for the buyer to send. It should not negotiate on its own, because concessions and relationships are commercial judgments. Most teams keep negotiation human and use the agent to shorten the preparation.

What happens when suppliers send quotes in different formats? The agent reads each format and maps it to one record, keeping a reference to the source for every figure. Where a term is missing or unclear, the record says so and a clarification question is drafted. A buyer can always check the original document behind any number.

Does the agent need access to our ERP? Not at the start. Reading a quote mailbox and a specification folder is enough for collection, comparison and follow-up. Connecting to the ERP is a later step, and it should be limited to what the role needs, with changes to supplier or payment data kept with people.

How do we know the comparison can be trusted? Check it against real cases before relying on it, first in shadow mode and then with buyer review. Look at whether figures match the source, whether gaps were flagged and whether non-comparable terms were caught. Keep reviewing on a schedule, since suppliers change formats and the category changes over time.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.