# How Finch is scaling plaintiff law with AI agents that research like associates

## **Key highlights**

• - **90% cost reduction:** per-query cost dropped after migrating to the Task API
• - **Zero 5xx errors or timeouts** across three-plus months in production
• - **Single-call structured outputs** replaced a three-step search, extract, and validate chain

**Tiered processing** lets Finch match cost and accuracy to query complexity at the individual task level

## **About Finch**

Finch builds the back office that plaintiff law firms never had. The company pairs an in-house legal team with AI agents to handle the research, case preparation, and administrative work that bogs down attorneys. Lawyers focus on settlements and litigation; Finch handles everything upstream. Their current focus is personal injury pre-litigation, with plans to expand into other plaintiff law areas.

The problem Finch targets is a supply-side bottleneck: 75% of Americans who need legal help can't get it. Law firms hit capacity limits and refer cases out rather than hiring more staff to handle the work.

## **The problem**

Finch's AI agents run structured research workflows at scale, pulling case-relevant facts from the web, enriching records, and feeding structured outputs into the firm's data models. Before Parallel, Finch powered these workflows through another search API provider.

That vendor's API contract was unreliable. Fields appeared and disappeared between responses with no warning. Finch's engineers wrote defensive validation layers and permissive schemas to keep production from breaking. The search quality compounded the problem: results came back as free text that required a second extraction pass before the data could enter Finch's structured pipelines. Every query became a three-step chain: search, extract, validate.

The engineering overhead of maintaining defensive wrappers around an inconsistent API consumed time that should have gone toward Finch's core product.

## **The solution**

Finch migrated to Parallel's Task API in November and now runs all web research through it. They break these two into two key workflows:

• - **Deterministic workflows.** These are repeatable queries with versioned schemas, the bread-and-butter of legal back-office research. The same categories of questions run against different cases, pushing structured data directly into Finch's data models with a stable output contract. Finch started here because the work is high-volume, well-defined, and predictable, which made the Lite Processor a clean fit for cost and latency.
• - **Production agents that write their own Task API specs.** These agents construct their own structured output schemas at runtime with variable depths, depending on the complexity of the research question in front of them.

Finch is also exploring Parallel's Monitor API for passive web monitoring as a future addition to their pipeline, so that agents can kick off research themselves, without human intervention.

## **The impact**

A 90% drop in per-task cost. Finch now runs higher volumes than before at a fraction of the spend.

> **"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."**

> **— Ben Weems, CTO, Finch**

After three-plus months in production, Finch has seen zero 5xx errors or timeouts from Parallel. The team that previously built defensive validation layers around an unreliable API now builds product features instead.

Parallel's tiered pricing lets Finch match compute to complexity across their entire operation. Deterministic workflows run on Lite. Production agents use Core. The tradeoffs between speed, accuracy, and cost are explicit at each tier, so Finch's engineers can make informed decisions at the query level rather than paying a flat rate for capabilities they don't always need