Situation
CAC Group is a specialty and middle-market insurance brokerage that merged with The Baldwin Group in early 2026. Sketch engineers work with CAC’s product team on the internal platform CAC uses to manage policy data.
Every policy comes with paperwork. A broker enters the policy details in CAC’s agency management system and attaches the PDF, often a scanned binder that can include hundreds of pages. Those files hold details the platform needs, and getting them out was the hard part.
CAC’s existing process used an optical character recognition (OCR) tool on an always-on Windows server in Amazon EC2. It worked from keyword rules. When it found a keyword in a document, it looked for a nearby value and returned it, sometimes with a confidence score.
Manual effort, low adoption
Changing a rule meant someone on CAC’s product team had to log in to the Windows server and edit it by hand. The work was tedious and hard to sustain, and the tool never saw wide use. Some information CAC wanted from its documents was out of reach entirely.
What CAC needed
-
Identify and classify documents as they arrive
-
Pull policy details from long PDFs that don’t follow one format
-
Let CAC’s team change what gets extracted without editing rules on a server
-
Keep the cost per document low
Have documents your systems can’t read?
From proof of concept to production
Getting to production fast has always been important because that's how you get user feedback. It's especially vital today as many AI projects stall out in the experimentation phase. It's why we offer the 30-day AI MVP to new clients, and why the CAC team elected to get into production in five steps:
1. Prove it in a week
The product team agreed to a short spike to test the idea. By then, Sketch engineers were building with agentic coding tools, so the proof of concept took about a week.
Cost control was the main concern. Testing the approach and analyzing token rates revealed a low cost per document. Caching brought it down even further.
2. Build on Amazon Bedrock
Sketch is a leading AWS partner in the USA, and CAC had strict security requirements, so running LLMs on Amazon Bedrock made sense.
Users can select from multiple LLMs. Processing runs on short-lived functions like AWS Lambda, so nothing sits idle on an always-on server.
3. Identify, classify, extract
When a user uploads a document, the platform extracts the text and sends it to the model with task-specific prompts. The model identifies the document, classifies its type, and pulls out policy details.
It also reads the full binder and fills in policy information that manual entry might miss. A user verifies each value before it becomes part of the platform’s policy records.
4. Show where each value came from
A custom document viewer lists every value the AI found. Selecting one jumps to the right page and draws a box around the source text, so users can check the result against the original document. That’s far more transparent than the old tool’s single value and confidence score, and eliminates fear of hallucination.
5. Put the controls in CAC’s hands
An admin area the team calls the playground lets CAC try new things:
-
pick models
-
write and test prompts
-
map documents to coverage types
-
track metrics
Prompts replace keyword rules. CAC’s product team, and even brokers, can tune extraction themselves.
.png)
Putting future enhancements on autopilot
As newer models arrive on Bedrock, this platform gets even better. There's no need for additional coding to get the benefits of improved models.
Let's talk about how to stay current with AI advancements.
Looking for something similar?
If your team has documents full of data your systems can't use, we can help you build a solution fast. We did it for CAC, and this isn't the first time we've built a custom AI solution that created a measurable competitive advantage. Let's talk about how to control costs while delivering results.
