How we automated 60% of inbound RFQs at PE Energy

PE Energy is an industrial supply distributor selling parts across the US, Mexico, and Colombia. They are the kind of business AI hype usually ignores — no consumer app, no growth-hacking team, just a phone that rings and an inbox that fills.

Their problem was the inbox. Two hundred plus RFQs (requests for quotation) a week, each one a customer asking “do you have this part, how much, when can I ship”. Every reply needed a human to look up the part, find a vendor, copy a price, and write a polite message back. The team was drowning, customers were waiting, and the obvious questions were the most painful: half of the volume was the same dozen parts.

What we built

A pipeline we call Teddy. Three pieces:

1. An extractor that reads each RFQ ticket in Zendesk, pulls out part numbers, quantities, and shipping address, and looks them up against the vendor directory.
2. A search agent that, for the parts we do not stock, runs a Google search through Apify, finds the supplier’s product page, and pulls a price.
3. A drafter that writes a reply with the part numbers, suggested vendors, and source URLs — in the salesperson’s voice — and attaches it to the Zendesk ticket as an internal note.

A human still sends the reply. But the work that used to take ten to fifteen minutes per ticket now takes under two.

What we did not build

We did not build a chatbot. The customer never sees the AI. We considered it; we decided against it. In industrial supply, the customer wants to know that a person looked at their request. The fastest way to deliver that feeling is to make the person fast.

We also did not replace the CRM, the helpdesk, or the email system. Teddy is a thin layer on top of Zendesk that adds an action button to the agent’s sidebar. That decision saved us roughly six weeks of build time.

What surprised us

Two things.

Vendor matching was the hard part. Extracting part numbers from messy customer emails is easy — LLMs are great at it. Matching those part numbers to the right vendor in our directory was the real work. We ended up building a fuzzy lookup against organization notes in Zendesk, with a small Claude-powered tiebreaker for ambiguous matches.

The Google searches paid for themselves the first month. We were ready to swallow the Apify bill. We did not need to. The hours we saved on parts research alone covered the search costs by week three.

Where it is now

Teddy handles roughly 60% of inbound RFQs without a human touching the part-research step. The other 40% are the genuinely hard cases — obsolete equipment, custom assemblies, urgent jobs with vendor problems — which are exactly what we want the team focused on.

Reply times dropped from a same-day median to a 47-minute median. Customer NPS went up; agent NPS went up more.

If you run an inbox like this

Three things we would tell ourselves at the start:

1. Find the boring repeatable middle. Not the easy stuff, not the hard stuff, the volume. That is where AI pays.
2. Do not replace the system of record. Build the AI as a feature inside the tool the team already lives in.
3. Keep the human visible. Drafts, not auto-sends. At least at first.