Summum a i
ES EN FR
Contact→
Home/ Cases

Cases told as model cards .

Structure drawn from public research: context, baseline metrics, intervention, before-and-after results . What we would ask of any provider who says "we did AI".

CASE-01 · ECOMMERCE · FASHION

Automatic returns triaging

Context

Thousands of returns per month. Customer ops team overwhelmed. High average SLA. Unmanageable peaks during sales periods. Frequent errors in classified reasons.

Intervention

Intent classifier + agent with tools (order query, stock, label) + refund signed off by human up to 0.92 confidence threshold. 8-week pilot.

Result

Most returns resolved without human involvement. Average SLA drastically reduced. Ops team now handles the complex queue instead of the repetitive one.

METRICS · before / after · 6 months

Agent v3

SLA · before
high
SLA · after
very low
notable reduction
Cost /case
high
Cost /case
very low
drastic reduction
Errors by cause
high
Errors by cause
minimal
drastic reduction
thousands of cases/month
high autonomy
team intact
CASE-02 · INDUSTRIAL · B2B

RAG over 14,200 technical documents

Context

Components manufacturer with 22 years of technical manuals in SharePoint. New engineer onboarding: 3 months to operational autonomy.

Intervention

Embeddings + BM25 re-ranking + Claude for synthesis and citation. Corporate auth. Team-level policy: R&D sees everything, workshop sees operational manuals only.

Result

Onboarding substantially reduced. High share of answers correct and cited. R&D team recovers productive hours each week.

METRICS · before / after · 3 months

RAG technical

Onboarding
long
Onboarding
short
notable reduction
Time /query
high
Time /query
very low
drastic reduction
Correct answers
partial
Correct answers
high
substantial improvement
thousands of docs
high query volume/wk
hours recovered/wk
CASE-03 · PUBLIC SECTOR

Computer vision for quality control

Context

Control centre of a public service. Manual inspection of a high volume of parts per hour. Human fatigue at end of shift translates into undetected defects.

Intervention

Several edge cameras with YOLO-v9 fine-tuned on thousands of annotated images. Triple human validation while the model was being calibrated. Conservative policy : when in doubt, escalate.

Result

Very high recall on critical defects. Low and acceptable false positive rate. Human inspectors move to qualitative review.

METRICS · before / after · 6 months

YOLO defects

Recall
partial
Recall
very high
substantial improvement
Rework
high
Rework
low
notable reduction
Throughput
baseline
Throughput
improved
substantial improvement
several edge cameras
thousands of training images
0 layoffs · redeployment

What would your case 04 be?

Design a pilot→ View services→