How Veracross expanded Harriet from HR to CS and beyond

Veracross builds software for schools. It offers a number of products that schools rely on day to day, each with its own users, its own questions, and its own internal teams supporting them.

Our partnership with Veracross didn't begin with a top-down rollout. It began with one team that was buried in repeat questions, and grew into something neither of us had fully mapped out in advance.

An HR team that needed capacity

The first conversation was with HR. The team was fielding the same questions over and over, the kind employees could plausibly answer for themselves if the right information were easier to find. Adding more people wasn't the answer they wanted. What they wanted was capacity, visibility, and consistency.

That's where Harriet came in. Instead of routing every question to a person, the team set Harriet up to handle the self-servable ones directly. HR got two things at once: time back to work on the questions that actually needed a human, and a clear picture of what was being asked across the company. Patterns surfaced. Inconsistencies in how questions were being answered surfaced. Decisions about what to document and what to change came out of data the team hadn't had before. Of all the places we could have started inside a company as broad as Veracross, HR turned out to be the right one, because the shape of the problem matched what Harriet does well: a high volume of repeatable questions, a small team, and a real need to see what was actually being asked.

Trust earned, then extended to customers

What started in HR didn't stay there. As the partnership deepened, the conversation shifted from “how do we solve this one team's problem?” to “where else does this pattern apply?”

The next experiment was Customer Success. Veracross CSMs spend a meaningful share of their time answering questions that have been answered before, often buried inside support cases that customers can theoretically search themselves. We started carefully: when a customer reached out, the CSM would ask Harriet first and use the answer to shape their reply. That gave the team a low-risk way to validate that Harriet was surfacing the right cases and giving the right answers.

The pilot worked. As the team got comfortable with what Harriet could handle, the feedback inside Veracross was direct.

The next step was bolder. We embedded Harriet directly inside the closed customer community where Veracross customers had previously been left to search for themselves. Customers stopped searching. Harriet read across the relevant cases and gave them one answer instead of asking them to read ten.

Early NPS came in around 9.4. From there, we extended what Harriet could do: if a question can't be answered from existing cases, Harriet now creates a new case so a CSM can step in.

Why it worked: a knowledge layer that was ready

Most AI rollouts fail not because the AI is bad but because the knowledge underneath it isn't ready. Veracross had spent six years investing in Knowledge-Centered Service (KCS), a discipline of writing knowledge into the act of solving customer problems, not as a separate documentation project. By the time Harriet arrived, the knowledge base had grown from around 1,500 product articles to thousands of Q&As that reflect how customers actually phrase questions.

That foundation is what Harriet sits on top of. As Ian Drummond, Veracross's Director of Knowledge & Instructional Design, put it in a recent piece, AI tends to amplify whatever's underneath it. Harriet works as a synthesis layer over years of disciplined knowledge work. At Veracross, it has been resolving questions at 98%+ accuracy across thousands of interactions over a recent three-month window. The few failures have mostly pointed back to gaps in the knowledge base rather than failures of the AI itself.

Sharing skills, controlling spend

Underneath all of this is Endpoint AI, the layer that lets Veracross scale Harriet across more teams without each one starting from scratch, and without losing control of cost.

The first piece is sharing skills. Someone inside Veracross who knows how to build a Harriet skill well doesn't have to build it once for HR, once for CS, once for implementation. Endpoint AI makes a skill built once available to whoever needs it, so teams aren't rebuilding the same patterns in parallel.

The second piece is spend. Veracross uses Endpoint AI to route queries to the model that fits the task, including cheaper models where they're a good match. That keeps token usage in check and keeps total spend below what a flat Claude Enterprise setup would cost, without asking teams to think about model selection themselves.

What this partnership has actually been

Looking back at how this has unfolded, it doesn't look like a typical software rollout. It looks like a series of small bets that paid off, each one earning the trust to make the next one. HR first. Then CS, internally. Then CS in front of paying customers. Implementation alongside it. And the systems underneath to make sure no team is building the same thing twice.