A collections agency where busy season was won or lost in the inbox

For most of the year the email was manageable. Then the month-end and quarter-end surge hit, and the inbox swallowed the day. Agents spent a large part of every busy-season morning reading the same kinds of messages, finding the matching file, copying details from one document into a reply, and deciding who should actually handle each request. The volume, not the judgment, was the cost. The real work, the parts that needed a person, kept getting pushed to whatever hour was left.

They had already tried to fix it. During one peak season they brought in an outside service built on a public AI model. The results were not conclusive, and worse, it meant client information was being sent to a model the firm did not control. For a business that handles sensitive financial detail on every account, that was a line they were not willing to cross again.

We built the assistant to run on the firm's own infrastructure, not a public model. Client correspondence and account detail never leave systems the firm controls, which removed the confidentiality objection that had killed the earlier attempt.

On top of that sits the day-to-day work. Incoming mail is sorted by what it actually is, a payment question, a dispute, a document request, a new instruction, and routed to the person who owns that account. For the routine messages, the assistant drafts a reply grounded in the firm's own files, pulling the matching account and the right numbers so the agent edits and sends instead of writing from scratch. Anything it is unsure of it flags rather than guesses, so a person always makes the call on anything sensitive.

In peak season the inbox stopped owning the morning. Agents get a sorted queue with a draft already attached to the routine messages, and they spend their time on judgment, checking and adjusting, not on reading, searching, and retyping. That gave back roughly two hours per agent per day during the surge, the stretch where every hour was already spoken for.

Just as important, the firm got the help without the exposure. The assistant runs where the firm's data already lives, so the leadership that had vetoed a public-model tool could finally say yes to this one.