Building your own AI procurement tool works fine for internal dashboards. It breaks on supplier communication — unstructured, chaotic, high-stakes inputs where the prototype works but the production system never does.
The conversation we keep having
A VP of supply chain at a fast-growing pet food company described a plan: hire a head of AI, build data infrastructure, connect it to Claude or ChatGPT, and automate workflows — reporting first, then procurement automation, then possibly supplier communication.
The logic is sound in the abstract. The team knows the business better than any vendor. The tools exist. Why not build exactly what you need?
Every company that has tried reports the same outcome: the prototype worked. The production system didn’t.
Where build works
Build makes sense when the inputs are structured and the outputs are obviously verifiable:
- Internal reporting and dashboards. Connect an ERP to an LLM for summary reports. If the chart is wrong, someone notices immediately.
- One-off automation scripts. Reformatting POs, drafting emails, cleaning up spreadsheets.
- Outputs a human can eyeball. The cost of being wrong is low because wrongness is visible.
Where build breaks
Supplier communication is a different problem. The inputs are chaotic, unstructured, and high-stakes. Real supplier communication looks like:
- PDFs containing 19 concatenated documents.
- RFP responses split between an email body and a spreadsheet with mismatched column headers.
- Mixed-language responses with unit conversions buried in footnotes.
- Promotional offers confused with firm pricing quotes.
- Specification changes embedded in pricing threads.
Roughly 30–40% of real supplier communication is messy, ambiguous, or structurally weird in ways you can’t anticipate from a sample.
The data network effect you can’t replicate
Waystation has processed over 47,000 supplier emails across dozens of food, beverage, supplement, and pet food companies. Each new customer surfaces edge cases the others didn’t — how flavor houses structure quotes differently than commodity suppliers, how co-manufacturer pricing contains embedded spec changes, how some suppliers send certificates as inline images rather than PDF attachments.
Every new customer makes extraction better for every other customer. An internal build starts at zero and only learns from its own supplier communications.
The maintenance tax
Beyond the initial build, ongoing maintenance becomes a permanent headcount cost:
- LLM API changes and prompt behavior drift with each model update.
- Supplier format changes break parsers silently.
- Engineering attrition creates knowledge concentration — the system ends up living in one engineer’s head.
- Multiple internal teams need updates and can’t all be served by a single engineer moonlighting on the project.
The same institutional knowledge problem that procurement has with suppliers, the build-it-yourself team now has with the parser.
The real decision framework
Build if:
- Data is already structured (ERP, database, spreadsheet).
- Output format is predictable.
- Wrong answers are obviously wrong.
- One team uses the tool for one purpose.
- Engineering capacity is permanently available.
Buy if:
- Data is unstructured (email, PDFs, attachments, zip files).
- Inputs are chaotic and unpredictable.
- Accuracy is high-stakes — wrong extraction means wrong pricing or a missed certification.
- Multiple teams need the same data.
- The system needs to improve from exposure to more suppliers than any one company has.
Supplier coordination is firmly in the “buy” column.
What “AI-native” actually means
Bolting AI onto an existing workflow is digitization. An AI-native system is built entirely on AI extraction — the AI is the product, not a feature. Every email processed improves extraction. Every edge case resolved helps future customers. Each new document type expands capability.
Foundation models getting 10× better reduces cost. It doesn’t generate the data. The data that runs CPG has always existed — it just lived in supplier emails. LLMs made it readable.
