How CommonLit Cut Partner Research From 30 to 5 Minutes While Improving Quality
The education nonprofit uses Mastra to combine deterministic TypeScript with LLM steps, turning CRM records, internal data, and public research into first-call notes.
30 → 5 min
Research time per new partner
15+
Workdays saved every year
CommonLit is an education nonprofit that helps elementary, middle, and high school teachers teach reading and writing. Working with almost 3,500 school partners, its platform has reached 42.3 million students and teachers. And every year, they add over 300 new partner school districts.
Before a first call with a prospective school, CommonLit's team spent the half hour it could spare reading notes and gathering public information. Now, a Mastra-powered workflow puts more thorough research directly into Salesforce in five minutes, giving the team back at least 15 workdays every year.
From quick research to high-quality sales intel
Generic sales tools were not always a good fit for CommonLit's nonprofit partnership process. The company talks to schools and departments of different sizes, with the schools' purchasing process and a mix of formal requirements and informal decisions. It was difficult to templatize the approach.
Before the first meeting, the partnerships team spent roughly half an hour preparing a brief. That work could wait 2-3 days before someone even had time to do it. They reviewed public information about the district and its town, and CommonLit's data about existing customers. The outcome was useful, but it could have been much better if the team had had more time.
The team experimented with tools like Claude, but the team of former teachers had little patience for AI-smelling writing. And Geoff Harcourt, CommonLit's CTO, wanted a cloud process that could connect existing systems, run on a trigger, and send its output to Slack (not just locally).
First he tried Anthropic's managed agents offering, but found it difficult to set up, debug, and manage credentials. When a colleague recommended Mastra, it checked all the boxes.
"Mastra had all the tools sitting right there and it was absurdly easy to stitch it together. The set of pre-existing tools and adapters is really good." — Geoff Harcourt, CTO
CommonLit's main application runs on Ruby on Rails, and the team was already familiar with TypeScript. Mastra gave them a straightforward SDK and a managed deployment option. They could run a service with explicit connections to the necessary systems without setting up another service in their Kubernetes cluster.
How the AI partner research works
The research starts when a school reaches out through CommonLit's contact form. The lead passes through HubSpot and Salesforce. When it converts to an opportunity, a webhook sends Mastra the Salesforce opportunity ID.
This is what starts the workflow. It pulls the contact form submission from HubSpot and existing information from Salesforce, including earlier conversations with the district. A Salesforce MCP connection provides access to CRM information. The workflow also gathers data from Looker and CommonLit's data warehouse.
Exa searches add public information about the district and its schools. Depending on what is available, the research can include school calendars, school board minutes, and local context.
Harcourt estimates that about 80% of the workflow is deterministic code. LLM steps read the gathered material and produce summary notes. With such a setup, CommonLit limits where model uncertainty enters the workflow, and controls the model costs.
The workflow writes the report to a large custom field on its own tab in the Salesforce opportunity page. The partnerships team gets its usual notification that an opportunity has been qualified. A few minutes later, the research report appears in the same record.
"This used to take half an hour, and now the workflow does it in five minutes and does it significantly better. I have not had a single case where someone has looked at these results and said, 'I wish I'd done this myself.'" — Geoff Harcourt, CTO
Expanding to onboarding and other workflows
CommonLit built two more workflows with Mastra: one for tracking school district onboarding and another for lead generation updates.
When a school adopts CommonLit's curriculum, the work is just beginning. Teachers need to get set up, learn the curriculum, and complete training. CommonLit's district success team has to keep track of all of this. That matters because implementation affects CommonLit's renewals. If a school implements the curriculum poorly, it may conclude that CommonLit doesn't work.
The onboarding report brings the Salesforce records together so the head of the District Success team can see where each account manager's schools stand. That has saved them hours spent chasing updates and combining the data.
Another report provides an update on lead generation three times a week. It combines deterministic processing with LLM steps to produce a consistent format. The process began in the marketing lead's Cowork setup, sending updates to Slack, but the team moved the process into a cloud deployment in Mastra to make it more reliable and accessible.
And CommonLit has more ideas for automating their work. The team is now exploring a way for the fundraising team to ask questions across several separate data sources, and a system to help prepare board materials with consistent metric definitions.
"Mastra has made it really easy to combine deterministic processes with LLM-driven synthesis to build reliable, load-bearing workflows for our teams." — Geoff Harcourt, CTO
