How we built private AI in ~8 weeks and eliminated $1.7M a year in cloud-AI spending.

The breach made the hosting decision urgent.

This was a firm that had already spent years investing in AI. It had its own knowledge-management and innovation teams. Leadership had backed the tools, paid for them, and told attorneys to use them.

Then came the LexisNexis breach in February 2026.

LexisNexis confirmed that an unauthorized party had accessed servers holding sensitive data, including customer names, business contact details, and support requests. The company claimed customer search queries, client information, and contracts hadn't been accessed.

Harvey AI, for example, hosts its environment in Microsoft Azure and processes documents, prompts, and responses through its cloud service. It provides encryption, access controls, and says it limits what's used for training and what's retained. But this wasn't enough for our client. They wanted to ensure that sensitive AI processing stayed inside their own environment.

That requirement brought them to us to build a private system. We agreed with the direction, but hosting alone wouldn't tell us whether the tools helped attorneys finish their work. So before rebuilding them, we studied what happened when their attorneys actually used AI.

Attorneys were using AI, then doing the work again.

For one week, we studied workflows inside their corporate practice. We included regular AI users and attorneys who weren't using their tools.

The observation ran on their own servers. The system removed case details and grouped activity by type of work. Our team received the reports, not raw captures of privileged matters.

The pattern was hard to miss. An attorney would ask the AI for a draft, read it, check the cases it cited, and then rewrite the output. In their corporate practice, 83% of AI outputs needed edits or a redo.

Their attorneys still carried responsibility for what left the firm, but an AI's answer appearing on a screen didn't make it trustworthy. That left them checking and repairing the very work the tool was meant to help them finish.

And because they were tracking use, the mandate made that problem harder to see. Attorneys could comply with the instruction to use AI and still spend their time redoing its work.

More pressure would've reinforced the wrong behavior.

When we walked their leadership through our findings, they pointed to usage dashboards that didn't look so bad. They had been tracking adoption, but the number didn't show how much work came after an AI's answer.

One partner's instinct was to speak with the team and push adoption more directly. But more pressure would produce more use of the same tool, without giving attorneys a reason to trust its output. That's why we asked them to hold off.

Their frustration was that the problem had been there while they were watching the wrong measure. They'd set a standard and followed it. Their standard hadn't captured whether the AI helped attorneys finish their work.

So we laid out a different plan in a memo and workflow maps. Bringing AI in-house would meet their data-control requirement, but useful adoption also required answers their attorneys could verify and use.

Our diagnosis changed what we built: private AI designed around the work their attorneys needed to finish, with checks that made each answer easier to trust.

We made the system prove itself on their own work.

We built their private AI system on their servers, starting with M&A inside their corporate practice. The work took ~8 weeks of engineering.

Because trust was the barrier, we tested the system against historical matters their team already understood. The system had to work from information available before the deal, without seeing the test matter's own documents.

To make that trust checkable in daily work, we built checks into each AI answer: the cited source had to exist, the reference had to resolve, and quoted text had to match. The system could decline an answer it couldn't support. Their partners kept responsibility for reviewing the legal judgment.

Their security team connected their existing authorization system, so document access was checked before retrieval. Sensitive documents and AI processing stayed inside their environment.

The goal was practical. Give attorneys an answer they could trace to its source, check, and put to work.

Our first handover plan asked too much of busy attorneys.

We initially planned for attorneys across the firm to build their own workflows. But tool-building took time away from billable work, so our plan asked busy attorneys to do something their targets discouraged.

We therefore changed the plan and trained their knowledge-management team and named champions as the main workflow builders. That gave their attorneys useful workflows without making each of them responsible for building one.

But the people doing that improvement work still needed time for it, so we also proposed counting approved build and adoption work toward billable targets. Their leadership reviewed our proposal and adopted a credit of up to 100 hours per attorney each year.

That decision gave improvement work time in the operating model. It addressed the conflict we'd missed in the original plan: asking attorneys to help improve the firm while their targets rewarded doing something else.

The mandate ended. Attorneys kept choosing the system.

Their practice's innovation lead removed the AI-usage mandate for the group. As a result, their attorneys could decide from launch whether the system we built was worth using. Continued use would have to come from the work it helped them finish.

~3 months later, we repeated the one-week study in their corporate practice. Rework had fallen from 83% to 24%. Usage records also showed an increase from 32% to 72%, after the mandate ended.

For their leadership, there was now a concrete answer to whether their investment helped: more voluntary use, with less work being redone. In a follow-up conversation, an M&A attorney described the burden of repeated checking easing. That was the change their usage dashboard alone couldn't show.

The private system we delivered also let them eliminate ~$141,000 a month in cloud-AI expense, ~$1.7M a year, while keeping sensitive AI work in-house.

We gave their team the skills to keep building on it.

We taught their knowledge-management team and practice champions how to build workflows and how to find the next bottleneck. They learned to review time spent on past work and study live weeks to see where AI could help.

With that training and the system we delivered, their team extended it beyond M&A into other practice areas.

That left their leadership with useful AI, a team that could build on it, and a way to judge where further work would matter. Their attorneys had more reason to use the system because it helped with the work they remained responsible for.

Technical details

Read how we handled deployment, access controls, evaluation, and handover.

They came for private AI. We made it worth using.

They came to us to bring sensitive AI work in-house, but they hadn't realized how much their existing tools left attorneys redoing. Our diagnosis changed the build, and we solved both problems: we kept their sensitive AI processing inside their environment and built a system their attorneys chose to use without a mandate. Usage rose, rework fell, and ~$1.7M in annual cloud-AI spending went away.