How we work with firms to maximize adoption

Most firms that buy an AI tool can tell you how many people signed in during the first week. Fewer can tell you how many still use it two months later. Our customers are consulting and finance firms, which audit every subscription at renewal, and a tool nobody opens is the first one cut.

What we have learned is that adoption at a firm comes from three pieces of work done together. 1) We learn the firm's real use cases before anyone is given access. 2) People are trained on those use cases. 3) We work day to day with the firm's own AI team, the people inside the firm whose job is to close the gap between their colleagues and the technology. What connects all three is fit: Verra has to work with the files, templates, apps and habits people already have. Most of our work is finding out where it doesn't, yet.

In October, Verra's AI models read and wrote about 316 billion tokens for people at the firms we work with. This post describes how we got there.

Learning the work before the rollout

Before a firm gives anyone access, we work through their use cases with the firm's heads of AI. Then we sit with the people who will use Verra, in person where we can, and go through what they actually do. The engagement types that come up every week, the documents each one starts from, the deliverable it ends in, and the conventions a new hire would be taught.

We build the first demonstrations on the firm's own material wherever its security review allows it. A demo on the firm's own case files, or on a clean folder of its best past work, tells us far more than a demo on public documents. People can see whether the answer looks like their work, and we can see where it doesn't.

When a firm starts with a pilot, we keep it to a few use cases in one practice, usually over about four weeks. Before it starts we agree on what success means, with numbers attached, and on the data each use case needs, with an owner and a date for every item. We also record how the work goes today, so the readout at the end can say how many hours and days were saved.

This is also where we learn the firm's limits. Some firms want Verra to answer only from a knowledge base someone chose, never from everything a person can reach. Departments within a firm often want different controls, and a team with sensitive client work may need a feature switched off that another team wants switched on. We set those rules up before the rollout, so the first experience matches what the firm expects.

Starting small and widening access where it takes hold

We don't turn Verra on for a whole firm on day one. The first people with access are usually the firm's own AI team, because they will answer their colleagues' questions later. Firm leadership comes next. People are far more willing to change how they work when they see their leaders doing it, and leaders who use the tool themselves are better at judging where it should go next. Practice groups follow one at a time.

What has worked best is widening access as groups take it up. When a group's first users come back week after week, the rest of that group gets access. When they don't, the firm finds out why before adding more people. The firm decides who gets access through its own directory groups, and Verra follows those groups, so widening a rollout is a change the firm's IT team makes in a place it already manages.

Training on the tasks people already do

In our experience, the trainings that work best are the ones a firm's own people run, built around the tasks their colleagues do. A firm's AI team knows which partner needs a deck rebuilt in the house template and which analyst spends Friday on the same workbook. In practice, training at the firms we work with looks like recurring office hours that anyone can join, short one-on-one sessions with the people in a group who show the most promise, and recorded walkthroughs of real tasks.

The shift we think matters most, and one firms' training teams are making now, is moving those sessions from "here is what the product can do" to "here is how to do this task you already do." People remember a session about a task they have to do next week far better than a tour of features.

At the start of a rollout we build the first trainings and the first skills together with the firm's AI team, and walk that team through everything we know about running Verra well. After a week or two they take it over and teach their colleagues themselves, and we stay close while they do. We also run some sessions for a firm's users directly.

After that, we support their work in a few ways. New releases come with a short guide inside Verra that walks people through what is new, with a worked example they can replay and a prompt to try it themselves. New features that change how people work can ship switched off, so the firm's trainers can learn them and prepare a session before their colleagues see them. We also give trainers the practices we see work across firms, such as pointing Verra at the right knowledge base before asking a question about a client.

Working with the firm's AI team

Where firms have adopted Verra widely, a small group inside the firm has owned the rollout. Sometimes it is an AI or innovation team, sometimes a practice leader who has made AI part of the job, and usually it includes someone who runs training. That group is our main partner, and most of what we do runs through it.

Day to day, that means a direct line between their people and our engineers. It means a regular check-in every other week, where they bring what their colleagues are asking for and we bring the patterns we see at other firms. And it means giving that team the controls to run the rollout themselves: admin roles that match how the firm is organized, group skills that a practice's own leads look after, and usage reports by group, so they can see which teams are taking to it.

A lot of the partnership is with the firm's IT and security teams too. Verra follows each file's permissions, so when a site's permissions are set up in a way we can't read, we find the affected files, send the list, and work with IT on a fix and on a way to catch the same problem next time. We answer security reviews, data protection questionnaires and subprocessor approvals as they come, and when a security tool flags something in our part of the firm's environment, we trace it and explain it the same day.

Turning feedback into the product

The fastest way to lose a firm's users is to hear the same complaint twice. Early in a rollout we ask for every small thing that doesn't feel right, write it down with the firm's team, and work through the list. After a large release we hold a call just to hear where the product falls short.

Many of the changes in Verra this year started that way. Feedback on projects changed how existing chats are added to a project and opened inside it. When a firm's trainers wanted a new feature held back until they had prepared a session on it, we put it behind a switch the firm controls.

Some feedback says more about the rollout than about the product. As people write their own skills, they can end up with many of them switched on at once, some better than others. The answer to that is partly in the product and partly in training, which is why we work through it together with the firm's AI team.

What the numbers show

In October, 91.2% of the people our customer firms had given access to used Verra, counting everyone with an account that month and everyone in the firms' synced directory groups who had not signed in yet, and counting as use anything beyond signing in, such as sending a message or making a file.

A token is a small piece of text, the unit AI models read and write in. In October, Verra's models read and wrote about 316 billion tokens for people at the firms we work with. Nearly all of that is reading: the documents, files, conversation and instructions the models take in to do the work, including text they read again from a cache. The total includes processing documents when they are first added and running searches. It is approximate, because some of those calls are estimated from request counts.

Most of the work is document work. In October, Verra handed people more than 39,000 files, and nearly half of them were decks or spreadsheets.

Verra also filled in more than 2.8 million rows in tables, where each row is one document read against a set of questions.

These totals show how much work people give Verra. They don't show the reason, which is that Verra is built around each firm's own work.

If you are rolling out AI at a firm

Start with the few tasks your people already do every week, and give the tool the firm's own material for them: the templates, a set of good past examples and the conventions a new hire would be taught. Find the people inside the firm who will own the rollout and give them the time and the controls to do it. Widen access to the groups that use it, train on their tasks, and fix what they tell you quickly. Then measure the share of people with access who use it each week, and watch that number for longer than the first month.