New York fintech, 30 employees · implementation · Client details anonymized. Adoption and time figures are the client's own measurements over the first weeks.
Situation
The company sells a data-science score to institutional investors, analysts and lenders: a measure of the risk that a company is manipulating its earnings, heading for fraud or heading for bankruptcy. Thirty people. A sales and solutions team that spends its days on proposals, client questions, due-diligence questionnaires and follow-ups.
The team had already discovered AI on its own. That was the problem. Client names, portfolio details and pricing were going into public chatbots on personal accounts. Nobody could audit what had been shared. Nobody could reuse what had been learned. And the compliance posture of a company that sells risk detection to banks could not survive a serious question about it.
What we built
Self-hosted LLM tooling inside the company's own environment. The models run where the data lives. Nothing leaves. Every prompt and every answer is logged and reviewable.
Tools for the sales and solutions team, built on the workflows the team already ran: first drafts of proposals and client answers from the company's own approved material, questionnaire responses from past submissions, meeting preparation from the CRM and the inbox.
One shared knowledge base. The answers the company had already given, validated once, reused everywhere. Knowledge that used to live in individual chat histories now stays in-house.
Rollout with the team, not to the team. Power users first, two short training sessions, then the rest of the team, with the tools already working on real cases.
Results
- 90% team adoption in four weeks.
- 30% fewer admin tasks for the sales and solutions team.
- Shadow AI eliminated. No company data in public chatbots.
- Knowledge stays in-house, self-hosted and auditable.
What this means for a regulated business
Shadow AI is a signal, not a sin. The team was already telling you what it needed. The job is to give it a version you can audit.
Self-hosted is not slower. With the right tooling, the secure version was adopted faster than the public one, because it knew the company's own material.
Adoption is designed, not hoped for. Power users, working tools on day one, two trainings. Ninety percent in four weeks is a process, not luck.