Finance ops · Production

OffDeal powers every stage of M&A advisory with one Claude-based agent

The problem

OffDeal ran more than 12 separate agentic workflows that broke whenever SOPs changed, and hit a hard ceiling when deal context exceeded the million-token limit of their prior provider, causing API calls to fail outright.

First attempt

OffDeal had been using a long-context model from another provider, packing the full context into each call to preserve accuracy over RAG, but that approach failed when context exceeded the million-token limit.

Workflow diagram · grounded in source

Deal phase task initiated

trigger

“OffDeal uses AI across virtually every stage of a deal, from originating new clients to finding buyers, preparing materials, and running diligence”

Archie applies modular skills

ai_action

“the team now builds modular capabilities they call "skills," self-contained instructions that Archie can combine and apply flexibly”

Autonomous buyer sourcing

ai_action

“A long-running agent now handles the work autonomously for up to 4 hours, researching buyers across all 10 methods, cross-referencing findings, and recursively refining its search”

Personalized buyer emails drafted

output

“Archie, OffDeal's Claude-powered agent, drafts personalized chaser emails to 34 buyers from within the firm's Deal Portal”

Presentation decks generated

output

“Each deck takes about an hour with Archie, compared to the 30 to 40 hours a traditional team would spend on research, production, and review cycles”

Reported outcome

After migrating to the Claude Agent SDK, internal eval accuracy rose from 25% to 85%, context overflow failures dropped to zero, and each banker now manages five to eight concurrent deals; the firm closed eight deals totaling $91M in its first year.

Reported metrics

Internal eval accuracy (overall improvement)25% to 85%

internal eval accuracy after SDK switch only25% to 60%

internal eval accuracy after MCP and prompting improvements85%

context overflow API call failuresdropped to zero

Show all 17 reported metrics

internal eval accuracy (overall improvement)25% to 85%

internal eval accuracy after SDK switch only25% to 60%

internal eval accuracy after MCP and prompting improvements85%

context overflow API call failuresdropped to zero

concurrent deals per banker5 to 8

new potential client meetings per banker per day2 to 3

deals closed in first year8

total transaction value in first year$91M

average deal size increase (2026)$11M to over $20M

buyer list compute cost with agentroughly $200 in compute

buyer list cost with traditional teamupward of $12,000

deck production time with Archieabout an hour

deck production time with traditional team30 to 40 hours

agentic workflows consolidated12+

buyer sourcing agent autonomous run durationup to 4 hours

data sources connected to Archiemore than 20

prior provider context window limitmillion tokens

Reported stack

Claude Agent SDKArchieMCPDeal PortalClaude Code

Source

https://www.anthropic.com/customers/offdeal

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After migrating to the Claude Agent SDK, internal eval accuracy rose from 25% to 85%, context overflow failures dropped to zero, and each banker now manages five to eight concurrent deals; the firm closed eight deals…

What tools did this team use?

Claude Agent SDK, Archie, MCP, Deal Portal, Claude Code.

What results were reported?

Internal eval accuracy (overall improvement): 25% to 85%; internal eval accuracy after SDK switch only: 25% to 60%; internal eval accuracy after MCP and prompting improvements: 85%; context overflow API call failures: dropped to zero (source-reported, not independently verified).

What failed first in this deployment?

How is this finance ops AI workflow structured?

Deal phase task initiated → Archie applies modular skills → Autonomous buyer sourcing → Personalized buyer emails drafted → Presentation decks generated.