01Does it work without ERP integration?
Most procurement AI tools require a connection to your ERP before they can do anything useful. That means an IT project, API credentials, data mapping, and weeks of setup before you see a single result.
The best tools work from the data you already have flowing through email. ERP integration should be additive, not a prerequisite.
How Kynthar answers: Yes. The Pilot runs on forwarded email. No ERP integration required. ERP exports (NetSuite, QuickBooks, Epicor, CSV) are additive after the Platform contract.
02Can it read messy real-world documents, or does it require structured templates?
Real procurement documents are messy. Vendor invoices come in dozens of formats. Carrier emails bury delay notices in reply threads. Rate confirmations are PDF attachments to plain-text emails. Contract amendments show up as scanned PDFs with handwritten annotations.
Template-based systems break when the format changes. AI-native systems read whatever comes in.
How Kynthar answers: AI-native from day one. No templates, no per-vendor configuration. Reads PDFs, Excel, email threads (including quoted replies, forwards, and nested attachments) in any format.
03Does it cross-reference documents, or only extract from one at a time?
Single-document extraction is a commodity. Every LLM API can pull fields from a PDF. The hard problem is connecting documents: matching the invoice to the PO, comparing the invoiced price to the contract, checking the received quantity against the BOL.
This is the difference between a scanner and an analyst. A scanner reads one document. An analyst reads all of them and tells you what does not match.
How Kynthar answers: Cross-document intelligence is the core of the product. Every document is resolved to entities (POs, vendors, items, contracts, shipments) and validated against everything else in the graph.
04Does it surface anomalies automatically, or does it require you to know what to ask?
Some tools give you a search box and call it AI. You still need to know what to look for. The best tools surface problems you did not know existed: the overcharge buried in a routine invoice, the contract clause being violated on every order, the vendor whose on-time delivery rate dropped 20% last quarter.
If the system only answers questions, it is a search tool. If it proactively finds problems, it is intelligence.
How Kynthar answers: Automatic. Kynthar continuously scans for anomalies across all documents: price mismatches, contract violations, delivery failures, duplicate payments, and 70+ other patterns. No configuration required.
05Can the AI access cross-tenant data, or is it isolated at the database layer?
When a procurement system serves multiple companies, tenant isolation matters. If the AI can see data from other customers (even accidentally), you have a security and competitive intelligence risk.
The gold standard is database-level isolation: row-level security enforced by the database engine, not by the application code. Application-level isolation is one bug away from a data leak.
How Kynthar answers: FORCE ROW LEVEL SECURITY on every table, enforced by PostgreSQL. Every query requires a company context before any SELECT returns rows. Enforced database-side, not trusted from the application layer.
06What is the LLM provider's data-retention policy?
If your vendor sends your procurement data to an LLM API, you need to know: does the provider retain your data? Can they use it for training? Is there a contractual zero-retention agreement?
Ask for the actual contract, not a marketing page. The difference between "we don't train on your data" (which may still mean they retain it) and "zero retention" (they delete it immediately after processing) matters for compliance.
How Kynthar answers: Contractual zero-retention with every LLM provider. No data retained after processing. No data used for training. No human review of your documents. Auditable.
07How long is the implementation?
Enterprise procurement platforms routinely quote 6 to 18 months for implementation. Even modern tools often require weeks of configuration, data mapping, and integration work.
For a Pilot that is supposed to prove value, anything longer than a few days is a red flag. The vendor should be confident enough in their product to deliver results quickly.
How Kynthar answers: Forward your procurement emails to an address we give you. First findings within 48 hours. No IT project, no data mapping, no training sessions.
08What happens at the end of a trial?
Ask specifically: where does your data go when the trial ends? Is it automatically deleted? Do you have to request deletion? Can you get a confirmation that it has been purged?
The answer reveals how seriously the vendor takes data stewardship. If they cannot tell you exactly what happens, they have not built the infrastructure to handle it.
How Kynthar answers: Data is purged on request with row-count confirmation per table. You get a deletion receipt showing exactly what was removed. Automated, not manual.
09How does the system behave when documents disagree?
In real procurement operations, documents disagree constantly. The invoice says $52.50/unit. The contract says $48/unit. The PO says $50/unit. The carrier BOL says 480 received. The vendor invoice says 500 shipped.
A naive system picks one and ignores the rest. A good system flags the disagreement and tells you which source to trust.
How Kynthar answers: Cross-reference and flag. When documents disagree, Kynthar identifies the discrepancy, links it to the source documents, and surfaces it as an actionable finding with the evidence inline.
10Can it explain its findings, or are they black-box?
An anomaly detection system that says "this invoice looks suspicious" is not useful. You need to know why, what specifically is wrong, and where the evidence is.
Every finding should trace back to the source document. If you cannot see the invoice line, the contract clause, and the PO line item that disagree, the finding is not actionable.
How Kynthar answers: Every finding traces back to the source document with evidence quotes inline. You see the specific invoice line, the specific contract clause, the specific PO terms that disagree. Fully explainable.