iTromsø · IBM, VC2
10 minutes per day — time spent reviewing potential stories
Independently reportedMedia & entertainment · Document processing · NLP · In production
Document processing
Finch · Parallel
Vendor-reportedProfessional servicesUnited StatesIn productionAI agentsLive 2025
Vendor-reported. The customer is named and the numbers are quoted from the source page, but the account comes from the vendor. No independent confirmation.
Reported by parallel.ai (vendor self-report). Checked against the source page on 2026-09-06. 3 of 3 figures below appear on that page word for word.
The problem
Finch's AI agents for legal research relied on an unreliable third-party search API, leading to inconsistent data, significant defensive engineering overhead, and a complex three-step search-extract-validate workflow.
What was deployed
- Vendor
- Parallel
- Products
- Parallel Task API, Lite Processor, Core
- Technique
- AI agents
- Build or buy
- Bought off the shelf
- Deployment
- API
- Data used
- case-relevant facts from the web
- Scale
- runs all web research through it
- Timeline
- 3 months
What changed
Each row is quoted from the source. Figures we could not find on the page in those words are marked — they are kept, not deleted, so you can judge them.
per-query costs
“cutting per-query costs by 96%”
Quoted word for word from the source
per-task cost
“A 90% drop in per-task cost.”
Quoted word for word from the source · round percentage, no baseline given
Zero5xx errors or timeoutsthree-plus months
“Zero 5xx errors or timeouts across three-plus months in production”
Quoted word for word from the source
defensive engineering overhead
“eliminating the defensive engineering overhead that came with an unreliable vendor.”
Quoted word for word from the source
product featuresengineering time spent on product featuresfrom defensive validation layers
“The team that previously built defensive validation layers around an unreliable API now builds product features instead.”
Quoted word for word from the source
“We went from spending engineering cycles on defensive validation to spending them on product. The Task API gives us structured outputs with citations in a single call, which is exactly what production legal AI needs.”
Sources
- parallel.ai — vendor self-report, 2026-04-20
Similar deployments
Adyen · Rossum
70 — automation rate
Vendor-reportedBanking & finance · Document processing · NLP · Scaled · 2023
Amplitude · Wolfia
4X — throughput of security questionnaires
Vendor-reportedTechnology & software · Document processing · AI agents · Scaled
Skaylink · Moterra
dropped by around 80% — NIS2 preparation time
Vendor-reportedTechnology & software · Document processing · LLM · Scaled
OM Pharma · DrugCard
3 minutes — review time per article
Vendor-reportedPharma & biotech · Document processing · NLP · In production
AMPREP (Amalgamations Repco) · Cimplify.AI
100% — invoice automation
Vendor-reportedManufacturing · Document processing · NLP · In production