adaptto.ai
AI for the back-office
of Industrial Businesses
For Manufacturing & Distribution companies
Case Study

How our AI system for RFQ is built

Industry
Lighting products distribution
Geography
+6 countries across Europe
RFQs
+5,000 per month
Process
ERP · Excel · Email
The Challenge
Each RFQ required 1 to 2 hours of manual work.
  • Search the product catalog
  • Cross-reference pricelists
  • Handle competitor references
  • Build the quote line by line in the ERP
At +5,000 RFQs per month, the volume made this unsustainable.
Our Approach

About our AI system

Receive
  • Email
  • Excel
  • PDF
  • Image
→
Parse
  • Extract line items
  • Quantities
  • Classify type
→
Match
  • Find SKU + price
  • Apply rules
  • Expand to ERP lines
→
Propose
  • Draft ERP quote
  • Flag exceptions
  • Await approval
Human in the loop
  • System produces a draft — not a final quote
  • Sales rep reviews, resolves flagged items, approves
  • Only then does the quote get created in the ERP
Parsing

How we read any input

Each line is classified before processing. Different type — different parsing strategy.
Type Description Example Strategy
A Own SKU reference ILAR-01691 x20 Exact catalog lookup
B Product name Berg Trimless 10W Search by product description
C Competitor reference Philips BN126C ou equivalente Spec extraction + cross-reference
D BOQ / spec code Tipo A / Lum_A01 Decoded against project spec
E Unstructured something for an 8m corridor LLM extraction + confidence flag
Edge Case

Low confidence match

Customer writes: "Berg Trimless Downlight"
ILAR-04821 Berg Trimless 10W recessed 74% confidence
ILAR-04830 Berg Trimless 15W recessed 61%
ILAR-04819 Berg Surface 10W 55%
≥ 90% Auto-matched — no human needed ✓ auto
60 – 89% Rep confirms or picks an alternative ⚠ review
< 60% No match — routed to senior sales ✗ no match
Edge Case

Competitor reference

The majority of RFQs contain competitor product codes — Philips, iGuzzini, KATOA — followed by "ou equivalente". The system reads the specs from the competitor code, then uses those specs to search our own catalog for the closest equivalent.
Step 1 — Read specs from competitor code
Input: "Philips BN126C LED 36W ou equivalente"
36W IP65 ← hard constraint 4000K 3600lm
Some specs are hard constraints — the system will never propose a product that falls below them (e.g. a lower IP rating).
Step 2 — Find our product matching those specs
ILAR-02204 36W · IP65 · 4000K ✓ all specs
ILAR-02198 33W · IP65 · 4000K ~ close
ILAR-02210 40W · IP44 · 4000K ✗ IP too low
Edge Case

Hidden dependencies

Some products require additional components that the customer does not specify. A DALI-dimmable luminaire always needs a separate driver — this never appears in the email.
Customer specifies: ILAR-02750 × 20 (DALI luminaire)
System detects DALI · adds automatically
ILDV-0048 DALI driver 40–80W × 20
How rules are learned
Rules are not hardcoded. They are mined from historical orders: when product A consistently appears alongside product B across past quotes, the system learns to propose both together.
Edge Case

SKU format inconsistency

Customers write the same product code in many different ways. The system normalises the reference before lookup — no manual correction needed.
Customer writes System reads as
ilar 1691 ILAR-01691 ✓ matched
ILAR01691 ILAR-01691 ✓ matched
IL-AR-01691 ILAR-01691 ✓ matched
ILAR 01 691 ILAR-01691 ✓ matched
Human in the Loop

5 gates where a human is required

1 Client not recognised
2 Line item unclear
3 Match confidence 60–89%
4 No match found
5 Full quote ready
Gate 5 is unconditional. No quote enters the ERP without a human signing off.
Continuous Improvement

How the system keeps improving

Every operator decision teaches the system.
Operator action System learns
Confirms a match Raises confidence for that pattern
Picks an alternative New competitor → own SKU mapping stored
Adjusts a quantity Client preference recorded
Flags a wrong match Confidence threshold recalibrated
  • New product A + B co-occurrence → dependency rule added automatically
  • New competitor code confirmed → cross-reference stored for future RFQs
  • Over time: fewer flags, more of the queue resolves without human touch
Next Steps

What we'd like to do next

We've shown you what we built for a similar business. Before we can give you a concrete proposal, we want to understand your process in depth. There are a few things we still need to figure out:
How orders arrive today — central inbox or distributed across reps?
How long it takes from order received to SAP — that's the baseline we'd improve on
How your SAP instance is managed and what the integration path looks like
What success looks like for you in 90 days
Based on this, we'll come back with a concrete quote and a proposed way of working — scoped to your actual setup, not a generic estimate.
Full list of questions →
1 / 11