Five steps to move from manual quote-chasing to agent-run sourcing — without waiting on your IT roadmap.

RFQ-to-PO automation is the use of AI agents to run the sourcing cycle — supplier identification, quote requests, bid comparison, and purchase order placement — with humans approving decisions instead of executing tasks. It sits a full layer above the RPA most manufacturers have already deployed. Where a bot copies quote data between screens, an agent reads part characteristics, decides which suppliers should see the RFQ, scores the responses, and recommends the award.

The question most sourcing leaders are asking is no longer whether this works, but how to sequence it. The short answer: start with structured part and supplier data, deploy agents first as decision support on supplier shortlisting and quote normalization, keep award approval human, and extend to automated PO placement only after the agent's recommendations have tracked human decisions for a full quarter. Done in that order, the cycle compresses without a rip-and-replace of the ERP — and without a multi-year line on the IT roadmap. According to Deloitte's 2026 Manufacturing Industry Outlook, 80% of manufacturing executives plan to invest in agentic AI by the end of the year. The ones who capture value will be the ones who sequence it, not the ones who spend the most.

Prerequisites: What Has to Be True Before Agents Touch Your Sourcing Cycle

Agentic RFQ automation works only when part data, supplier records, and award logic are explicit enough for an agent to reason over. If your buyers select suppliers from memory and compare quotes in personal spreadsheets, the first project is not an agent — it's making that tacit knowledge structured.

Three things need to exist, at least in rough form. First, a part master that carries the characteristics that actually drive sourcing decisions: commodity codes, materials, tolerances, certifications, and annual volumes. Second, a supplier master that records capabilities and performance — not just addresses and payment terms. Third, documented award logic: what weight price, lead time, quality history, and risk actually carry, and at what spend threshold a human must sign off.

What is not required is deep, write-level access to your ERP on day one. The highest-value early deployments run read-only — the agent sees item, supplier, and quote data, and pushes recommendations to a human. That matters for two reasons: it shortens the security review that stalls most external AI projects, and it means the work does not have to queue behind the dozens of integration projects already competing for internal IT capacity.

The Five Steps from Manual RFQs to Agent-Run Sourcing

The reliable path runs decision support first and autonomy last — each step earns the trust the next one spends. McKinsey estimates that autonomous category agents capture 15 to 30 percent efficiency improvements by automating non-value-added activities (McKinsey, 2025), and in the RFQ cycle nearly all of that sits in steps two and three below.

Step 1: Structure the data agents will reason over

Extract part characteristics and supplier capabilities into a form an agent can query. In practice this is two to four weeks of work against existing ERP tables, not a data-warehouse program. Watch for the classic failure: cleansing all 40,000 SKUs before starting. Pick the two or three commodity groups with the highest RFQ volume and structure only those.

Step 2: Automate supplier shortlisting by part characteristics

The agent's first job is answering "which of our qualified suppliers should quote this part?" — matching part attributes against supplier capabilities and performance history, and returning a ranked shortlist with reasoning. Buyers review and adjust. This is where drift from tribal knowledge to explicit logic happens, and where single-source awards made under time pressure start becoming competitive bids.

Step 3: Let agents issue RFQs and normalize the quotes

Once shortlists are trusted, the agent drafts and sends the RFQs, chases responses, and — the biggest single time sink — normalizes returned quotes into a comparable format across price breaks, lead times, tooling, and terms. This is hours of buyer time per RFQ cycle recovered. Failure mode to watch: agents sending suppliers RFQs with incomplete specs. Gate outbound messages behind a completeness check before anything leaves the building.

Step 4: Keep award decisions human — and make the agent argue its case

The agent scores quotes against your documented award logic and recommends a winner, with the trade-offs stated in plain language: what you give up on lead time to take the lower price, what the quality history implies. The buyer approves, overrides, or asks for a re-run. Log every override — that record is the evidence base for step five.

"RPA moves data between screens. Agents recommend which supplier deserves the order." — Nijansh Verma, Co-Founder at Heizen

This distinction is the whole design principle. If the deployment only accelerates keystrokes, you have bought expensive RPA.

Step 5: Automate PO placement last, under exception thresholds

After a quarter in which agent recommendations and human decisions have visibly converged, let the agent place POs autonomously below defined thresholds — spend limit, single-region supply, no quality flags — and escalate everything else. This is where cycle-time compression compounds: The Hackett Group's 2025 benchmarks show digital world-class procurement organizations run requisition-to-PO cycles 58% shorter than peers, and threshold-based autonomous placement is the mechanism that gets transactional POs out of buyer queues entirely.

Where RFQ-to-PO Automation Fails in Practice

The three most common failure modes are treating agents as faster RPA, waiting for the suite roadmap, and demanding deep system access on day one.

The RPA trap is the most expensive. Teams that frame the project as "automate the RFQ emails" get keystroke savings and no decision leverage — the scarce resource in sourcing is buyer judgment, not buyer typing.

The waiting trap is subtler. Gartner projects that 70% of supply chain software vendors will have agentic AI built into their products by the end of 2027, up from 1% in 2024 — which makes "wait for the suite" feel safe. But bolt-on features automate the vendor's generic workflow, not your award logic, and the two rarely match. Heizen is an AI-native software delivery company that builds supply chain systems for enterprise CPG and manufacturing companies, and in our work with enterprise manufacturers the pattern repeats: the sourcing teams that moved first with a scoped, read-only deployment had a year of override data and tuned award logic by the time the suite feature shipped — the ones that waited were starting from zero.

The access trap kills projects before they start. Any partner who needs administrator-level ERP access to begin is designing for their convenience, not your risk posture. Read-only scopes on a handful of tables are enough for steps one through four.

What Success Looks Like at 90 Days

Ninety days in, success is unglamorous: two commodity groups live, shortlists trusted, quotes arriving pre-normalized, and a growing log of agent recommendations against buyer decisions. No autonomous PO has been placed yet — and that is correct. The teams that get this right treat autonomy as something the system earns with evidence, one threshold at a time, rather than something the software claims on day one.