On This Page
AI Quote Ingestion — Zero-Template RFQ Processing for Enterprise Sales
A multimodal AI ingestion workspace that turns unstructured emails, handwritten notes, and non-standard spreadsheets directly into active quotations.
My role: Lead UX & Interaction Designer · Duration: 4 months
Team: Product manager, ML / NLP engineer, 2 full-stack engineers, Lead UX designer
My contribution: End-to-end design of the zero-template ingestion flow, side-by-side extraction verification workspace, and the batch conflict resolver for handling low-confidence catalog matches.
// From Customer email attachment (no fixed columns):
1) Need 450 units of SKF-6205-2RSH bearings ASAP for line 4.
2) Also quote 120 pcs 6204-ZZ Deep Groove Ball Bearing.
3) 15 sets of TIMKEN LM11749/10 (verify suffix).
Proprietary part numbers and customer client names are anonymized under Non-Disclosure Agreement. .
The Problem, at a Glance
Fragmented manual entry & rigid templates: Sales specialists copied part numbers one-by-one from emails, handwritten notes, and WhatsApp screenshots, or spent 45 minutes reformatting customer spreadsheets to match a rigid bulk template.
Zero-template multimodal ingestion: Drag and drop any Excel file (regardless of column format), screenshot, image, or email paste. AI extracts item codes and quantities, fuzzy-matches catalog databases, and flags ambiguities in an isolated batch resolver.
What I Learned from Shadowing Sales Engineers
During user research and contextual shadowing across quotation specialists, I noticed that sales inquiries rarely arrived in clean, structured formats. In reality, a single request for quote (RFQ) was assembled from multiple chaotic touchpoints:
Users toggled between Outlook, phone notes, and ERP search screens, copying 12-character alphanumeric SKUs one-by-one. A 40-item RFQ took 45+ minutes just for basic line entry.
The existing bulk upload required exact header keys like Material_Code_V1. Because customer spreadsheets never matched, users abandoned bulk upload because fixing column names took longer than manual re-typing.
- Alphanumeric error risk: Misreading a handwritten
0(zero) for anO(letter) or missing a suffix like-2RSresulted in generating quotes for incorrect engineering parts. - Unforgiving validation: Traditional forms failed completely when a single row in an Excel file was invalid, discarding the entire upload instead of importing valid rows.
- Trust requires provenance: Users needed to see where every extracted number came from. "Magic" AI that directly altered pricing without clear trace links caused immediate skepticism.
Zero-Template Interaction Model
Rather than demanding that users adapt to our software's schema, we inverted the mental model: the software adapts to whatever artifact the customer handed over.
-
01
Universal Dropzone
Accepts any spreadsheet, photo, screenshot, or raw email text.
-
02
Pattern Extraction
Extracts part numbers & quantities regardless of column layout.
-
03
Batch Verification
Review ambiguous items in one grouped step before commit.
Users can drop sheets with merged cells, random headers, or multiple tabs. The engine scans columns using entity detection rather than static key matching.
Snips from PDF spec sheets, customer chats, or photos of scribbled part tags are parsed through OCR with bounding-box provenance.
A quick keyboard shortcut allows pasting conversational email threads like "please quote 50 of part X and 200 of part Y" without saving files.
The AI Extraction & Catalog Matching Pipeline
The technological core combines multimodal optical character recognition, large language model entity parsing, and a deterministic catalog fuzzy matcher. This tri-tier architecture ensures that generative creativity never introduces fictional SKU numbers into an enterprise quotation.
Normalizes unstructured input streams (PDF tables, camera photos, .xlsx cells) into clean spatial text chunks.
Identifies SKU patterns, quantities, unit of measure (e.g. PCS, METERS, BOXES), and target delivery dates.
Deterministic Levenshtein & semantic alias matching against master product database to prevent invalid SKUs.
The Batch Conflict Resolver (Core UX Innovation)
In enterprise data tables with 50+ rows, scattered inline validation errors create cognitive fatigue. If row 4, row 19, and row 38 have missing quantities or ambiguous catalog codes, users are forced to endlessly scroll and hunt for red borders.
I designed the Batch Conflict Resolver: an isolated, dedicated triage step that surfaces all items needing human attention before anything is injected into the active quotation.
By resolving ambiguities upfront in a focused triage screen, the sales engineer enters the primary quoting environment with a clean, 100% verified table ready for margin analysis and pricing approvals.
Working with Developers & ML Engineers
Designing for AI requires a fundamentally different collaboration model than traditional deterministic SaaS. With ML models, the interface must gracefully absorb latency, non-deterministic extraction formats, and network timeouts.
-
Async Streaming States
Instead of a generic spinner, line items stream into view progressively as the parser processes each block.
-
Full Undo Checkpoints
Users can instantly revert an AI ingestion batch without damaging existing rows in an active quote.
-
Visual Provenance
Hovering any extracted row displays the source bounding box or cell in the original document.
-
ERP Schema Fallbacks
If ML extraction fails on an obscure brand, the system falls back to regex part-number heuristics.
The Result & Measurable Impact
The zero-template AI ingestion engine transformed how quotation specialists handle complex enterprise bids. What was previously an agonizing exercise in data entry became an effortless, two-minute verification flow.
Average 50-line RFQ turnaround reduced from ~45 minutes to under 3 minutes.
Zero hours spent restructuring customer spreadsheets to match rigid bulk schemas.
Rated "Indispensable" by sales engineers across initial pilot enterprise deployments.
What I Would Carry Forward
- Forgiveness in inputs beats strictness in validation: Demanding that customers or internal teams format their data to match an arbitrary engineering database creates immense user friction. When the system absorbs the mess, user delight skyrockets.
- Human-in-the-loop is an asset, not a compromise: In high-stakes B2B pricing, 100% automated black-box execution is dangerous. Designing an elegant, high-speed triage interface (the Batch Conflict Resolver) gave users total confidence that no mistaken part would slip into production.
- AI UX requires visual provenance: Whenever an AI extracts data, always retain the interactive link back to the source document. Grounding the extracted entity in physical source reality removes user paranoia.
Explore how the downstream pricing simulation, margin feedback matrix, and multi-tier approval workflows were architected for global sales teams.
Discuss This Project
I can walk you through the end-to-end Figma workflows, prompt engineering boundaries, and interaction blueprints.