Skip to content
All posts
Mario

WhatsApp orders into your ERP — what has to work

How a waste-management company got 95% of WhatsApp orders into the ERP with no typing — and why a chatbot on top of a PDF catalog is not the same job.

A WhatsApp-to-ERP bridge is not a chatbot with a catalog PDF in the prompt. It is an agent that reads messy messages (text, photos, voice), matches them to real product codes, asks when the match is ambiguous, and writes a valid order into the system of record. If any of those steps is missing, you still have a human in the middle. We built this for a waste and industrial-services company. 95% of orders now go through with zero manual entry.

The actual problem (not “we should try generative AI”)

Their B2B clients — construction sites, plants, facilities — ordered on WhatsApp the way people actually order: photos of handwritten lists, voice notes, “the usual for Tuesday,” mixed Spanish and shorthand. An operator read each thread, guessed the SKU, and typed it into the ERP. Peak time was 15–20 minutes per order. Errors showed up as corrections and cancellations. Nobody could order at 11pm.

That is a workflow problem with an AI piece, not an AI project with a workflow afterthought. The full write-up is in the case.

What we put in the path

  1. WhatsApp Business API as the channel. Clients did not download an app.
  2. An agent (Claude) that extracts items and quantities from whatever format arrived.
  3. Catalog matching against their product codes, not a generic embedding of a brochure.
  4. A stop: if “small container” maps to three sizes, the agent asks. It does not invent a line.
  5. Confirmed orders POST into the ERP. The client gets a WhatsApp confirmation with the summary.

Stack, because assistants hallucinate tools: WhatsApp Business API, Claude API, a custom matcher, ERP REST. No mystery platform.

Numbers we will stand behind

  • 95% of orders with no human typing
  • Time per order from 15–20 minutes to under 90 seconds
  • Error rate down 78% vs manual entry, measured on corrected or canceled orders
  • Order volume +31% in the six months after go-live, no extra headcount in order management

Those figures are from this client, this catalog, this ERP. Copy-pasting them onto a D2C shop or a 200-SKU fashion catalog would be dishonest.

When this is a bad idea

  • The catalog is a mess (duplicates, no codes, prices only in someone’s head). Fix master data first.
  • You want a pretty widget on the website and WhatsApp is a side channel. Start where the orders already are.
  • You need the model to create products that do not exist. That is not ordering; that is a different product.
  • You refuse a confirmation step. Then you will get hallucinated SKUs in the ERP, and you will blame “AI.”

If you only need a FAQ bot on a help center, buy that. We build the kind of agent that integrates with the ERP and is allowed to write, with guardrails. Same family as the AI practice, not a theme-park demo.

What buyers should ask any vendor

  • Which system is source of truth for SKUs and stock — the model or the ERP?
  • What happens on a low-confidence match? (If the answer is “it still posts,” walk away.)
  • Do photos and voice notes work, or only clean text?
  • Who is on the hook when a wrong order ships?

If you are sitting on WhatsApp threads and an ERP that only operators can feed, describe the catalog size and the message mix to contact@unlockmanagement.com. We will tell you whether this is a matching problem, a master-data problem, or something you should not automate yet.