MuleSoft IDP Implementation Services
in the USA

We spent the first two weeks not writing code. We sampled 400 real invoices across every supplier and built a taxonomy of layouts. Turned out 84% of the volume came from just 22 suppliers. We trained the IDP model on those 22 templates first and left the long tail to a general-purpose model as a fallback. This meant we hit useful accuracy in week three instead of month three.

The pipeline: shared inbox and scanner output route into an S3 bucket, MuleSoft IDP extracts the invoice header, line items, and totals, an Anypoint flow does a three-way match against the PO and the goods receipt in JD Edwards, and clean invoices post directly. Anything the model is less than 92% confident on drops into a human-review queue with the extracted fields pre-populated. The reviewer confirms or edits and hits submit, one click instead of thirty.

After 90 days: exception rate at 6% (versus their internal target of 15%), zero missed payment discounts in Q3 (they had missed $47K worth in Q2). Cost per invoice worked out to about $0.19 all-in.

The problem was not just accuracy, it was routing. Their previous vendor built three separate integrations. When the CargoWise API changed, everything broke.

We rebuilt the whole thing with IDP handling extraction and Anypoint flows handling routing. The design decision that mattered: we treated the extracted document as a canonical shipment object inside Anypoint, then transformed it to whatever shape each downstream system needed. Now when CargoWise changes their API (which they do twice a year) we update one transformation, not three integrations.

Certificate of origin, the messiest of the four, sits at 88% with the balance flagged for review. The ops team reallocated 8 people into exception handling and account management. Customs broker submissions used to fail 4% of the time; now sit at 0.3%.

The project ran over 14 weeks. MuleSoft IDP handled document classification first (figuring out which document was which - critical when a client uploads six PDFs with unhelpful names like "scan_01.pdf") then field extraction.

We built specific validation rules for each document type: the EIN letter had to match the number on the incorporation docs, the address on the utility bill had to match the address on the beneficial ownership form, the driver's license had to be unexpired. Anypoint routed extracted data into Financial Services Cloud as a new lead, triggered OFAC screening via their existing vendor, and updated FIS Horizon once approved.

Complex packets that needed human compliance review dropped from 4.2 days to about 22 hours. The compliance officer's exact words after month two: "we finally have time to actually look at the risky ones."