An AI agent for supplier research

Twenty sites, twenty price lists, twenty ways of hiding the lead time. Supplier research is the job people most want to hand off — and the one where a chat-only agent is nearly useless. Here is what doing it properly takes, from a real recorded run.

Ask a chat assistant to find suppliers for a part and you get a list of company names it remembers, some of which exist. What you need is different: the current price at the quantity you actually buy, the minimum order, the lead time, the shipping terms, and a contact you can write to — from the supplier's own site, today, not from training data. That is not a question; it is an afternoon of clicking. Which is exactly the kind of work an agent with a computer is for.

There is a real recording of a MyClawn agent doing this job on its own desktop — searching, building the comparison in a spreadsheet, and sending the enquiry from its own address. You can watch it and take over from the last frame. The rest of this page is what that run is made of.

What supplier research actually consists of

1. Finding candidates on the real web

The agent opens a real browser and searches the way you would: directories, marketplaces, manufacturers' own sites, the results behind "request a quote" forms. It reads each page as rendered — including the tables and PDFs where the useful numbers usually are — rather than fetching HTML and guessing. This is the step that a text-only agent cannot do at all, and the step that takes a person the longest.

2. Extracting the numbers that are never in the same place

Price per unit at your quantity band. Minimum order. Lead time. Incoterms. Whether the price is ex-works or delivered. Every supplier presents these differently, several hide them behind a login or a quote request, and some only state them in a downloadable sheet. The agent pulls what is visible, notes what is missing, and — for the ones behind a form — fills the enquiry and sends it (see below) rather than leaving a gap.

3. Building the comparison you can actually use

The output is not a paragraph; it is a spreadsheet on the agent's machine, one row per supplier, with the columns you asked for and a source link on every number, so any figure can be checked in one click. Because the machine persists, the sheet is still there next week when the second round of quotes comes in, and the agent updates it instead of starting over.

4. Sending the enquiry — from its own address

Where a price needs asking for, the agent writes to the supplier from its own email address, not yours, with your requirements in it. That matters for two reasons: the thread is the agent's to follow up, and nothing goes out under your name that you did not see. What it sent is kept on the machine and shown to you in the dashboard, so "I emailed them" is something you check, not something you take on trust. (One honest limit: replies to the agent's address are not delivered into its mailbox yet — see giving an agent its own email address.)

Where it saves the most

• Breadth. Checking twenty sites instead of the five you had patience for. The best supplier is often number fourteen.
• Repeat rounds. Re-checking the same suppliers next quarter, with the sheet and the logins already on the machine.
• The boring middle. Unit conversions, currency, shipping added to price, all done consistently across rows.

Where you still need to look

• Quality and legitimacy. An agent can find a supplier and read its page; it cannot tell you the parts will be good or the company will still exist in six months. Treat the sheet as a shortlist, not a decision.
• Numbers that need a phone call. Some quotes only exist in a conversation. The agent gets you to the point of knowing who to call.
• The commitment. Placing the order is yours. Set the agent up to stop before anything binding — the desktop is streamed live and you can take over, so the line is easy to hold.

How to brief it

Give it the part, the quantity, the region, the columns you want, and the deadline. Say what counts as a deal-breaker (minimum order above X, lead time above Y) so it can drop candidates instead of listing them. Drop any spec sheet or drawing onto its desk — it reads the whole thing. Then watch the first few sites it opens; you will know within minutes whether it understood the job.

Where this is the wrong tool

If you already know the three suppliers you will use, there is nothing to research. If the parts are so specialised that every quote is a negotiation, the agent does the finding and you do the talking. Supplier research earns its machine when the list is long, the numbers are scattered across sites, and the job comes back every quarter. See what the agent's machine and mailbox are made of in Cloud clones.