Most invoice approval workflows aren't actually about approval. They're about catching extraction errors.
The approver isn't deciding whether to pay the vendor — that decision was made when the PO was issued. They're checking whether the data was pulled correctly. Whether the invoice amount matches what's in the system. Whether the line items are right. Whether the vendor name got parsed correctly from a PDF that was scanned sideways.
It's quality control for a tool or process that can't be trusted.
Lido eliminates the extraction errors that create approval bottlenecks in the first place. Its AI reads invoices accurately enough—line items, tax details, PO numbers, vendor names—that approvers can trust the data without manually re-checking every field. When extraction is accurate, approval workflows become what they were designed to be: a policy check, not a data verification step.
One operations lead at a gas distribution company put it bluntly:
"The approval is all about the accurate extraction of the data. It has nothing to do with the content."
They process over 20,000 invoices a month. Every single one goes through manual review — not because the business logic requires human judgment, but because their previous extraction tool wasn't reliable enough to skip it. The approval workflow exists because automation failed.
This pattern shows up everywhere. A government agency paid $30,000 for a document extraction contract. They were told it would be plug and play. Instead, they ended up reviewing every output because accuracy was inconsistent. Worse, they were charged for each extraction attempt, including the ones that failed.
When extraction accuracy is unreliable, organizations build manual review into the process as a safety net. That safety net has a cost: headcount, time, and the cognitive load of checking work that should have been done right the first time. But the cost is hidden because it's baked into "the process." Nobody questions whether the approval step should exist. They just staff for it.
The gas distribution company had already migrated from Docparser to Nanonets specifically to escape template maintenance. They ended up in a different version of the same problem — spending "a ton of time retraining the models" and still requiring manual approval on every extraction.
The tool changed. The approval workflow didn't.
This is the trap with template-based and model-trained extraction tools. You might get marginally better accuracy with a newer tool, but you're still in the same paradigm. Still training models, still retraining when formats change, still dealing with the fundamental limitation that these tools need to be taught what a document looks like before they can read it.
If your extraction were 100% accurate, would you still require manual review on every invoice? For most teams, the honest answer is no. The review exists because the tool can't be trusted. Fix the accuracy problem, and the approval problem solves itself.
Eliminating accuracy-driven approvals requires a different architecture — one that doesn't depend on templates or trained models that break when formats change.
Layout-agnostic extraction means the tool reads documents the way a human would: by understanding what's on the page, not by matching it against a predefined template. When a vendor changes their invoice format, nothing breaks. When you onboard a new supplier with a format you've never seen, it just works.
When extraction is reliable, approval workflows can focus on what they're supposed to focus on: business decisions that require human judgment. Exception handling. Spend policy compliance. Vendor disputes. The work your AP team was actually hired to do.
Lido uses a custom blend of AI vision models, OCR, and LLMs to extract data from any document — invoices, POs, receipts, statements — without templates or model training. Upload a document, tell it what to extract, and get structured data back. When extraction isn't perfect on the first pass, reprocess free for 24 hours until it's right.
ACS Industries automated 400+ POs weekly and avoided a hire. Relay processes 16,000 Medicaid claims in 5 days. Soldier Field cut 20 hours of weekly invoice work to minutes. These teams aren't approving every extraction anymore — because they don't have to.
If your approval process is really just an accuracy check, it's time to fix the accuracy.
%20(1).svg)
