Overview
FMI is a consulting and investment banking firm dedicated to the built environment. More than 200 professionals across four offices work in two units, FMI Consulting and FMI Capital Advisors. The firm's value is its people: decades of institutional knowledge, deep relationships, and hard-won judgment. That knowledge lived in individual relationships, scattered documents, and disconnected systems. FMI's leadership knew AI could help their people work faster and with more of the firm behind them, but they wanted more than a tool. They wanted to understand how knowledge actually moved through the firm before deciding where to invest.
The Challenge
FMI's leadership believed most of their people were not using AI. The survey said otherwise: 92% were already using it every day or every week. The gap was not adoption, it was quality. Senior professionals were finding work from junior staff that included AI output no one had verified or went straight to a deliverable without a second read. An earlier training led by an outside vendor had not landed with FMI's experienced practitioners.
Underneath the quality problem sat a structural one. Decades of institutional knowledge existed across the firm, but finding it depended on knowing who to ask or where to look. A firmwide survey of roughly half the organization put a number on the cost: more than $250,000 a year lost to searching for knowledge the firm already had, and that figure left out duplicated effort and work redone because no one knew it existed. When 61% of respondents described, unprompted, a tool that could surface the firm's accumulated knowledge, the mandate was clear.
FMI had already tried the obvious version of that tool. They connected AI directly to their document and CRM systems, and the results were poor. It surfaced drafts instead of finished work, could not tell what was current or most relevant, and gave answers that needed so much checking they were hard to act on.
The lesson was not that AI could not help. It was that a direct connection was the wrong shape. Plugging each person's AI tool straight into the firm's systems put the cost, the setup, and the risk of a wrong answer on every individual, over and over.
Start with Discovery
Those findings came from a Discover Phase, the deliberate first phase of the engagement, not a byproduct of building. Two workstreams ran in parallel: an AI enablement assessment, with interviews across both units and the firmwide survey, and a data evaluation that mapped FMI's systems to see whether AI could reliably reach the right information. Together they produced a maturity assessment, a quantified business case, and a working prototype that proved the approach end to end. RoleModel presented it to senior leadership with a sequenced path forward.The Solution Process Explained
"We had already tried plugging AI straight into our systems, and it didn't work. What changed with RoleModel was the order of operations. They spent a full phase understanding how our people actually find and use knowledge before writing a line of code, and that's why what we built is something our practitioners trust."
Jericho Cruz
Chief Information Officer, FMI Corporation
The Harness
What FMI needed was not a connection but a harness: one place that sits between the firm's systems and the AI tools its people already use, and does the hard part well every time.
Atlas is that harness. Someone asks a question in plain language from the tool they already work in, and Atlas answers from the firm's real records, with a citation on every response so the answer can be checked. Behind that simple experience, the harness carries the weight. It runs its own agents and skills to carry out a task, reaches across the firm's several CRM and document systems to gather what the question needs, and does the work that belongs in code rather than forcing everything through a prompt. A person can still go straight to a system when that is the better path. The harness is the standard way, not the only way.
The AI tools FMI's people use
e.g. Claude and ChatGPT
Atlas
the harness
OBSERVABILITY
watches every operation
THE FIRM'S EXISTING SYSTEMS
Figure 2 · One shared layer between the firm's people and the firm's systems
Because the work runs in one place, three things become properties of the system rather than problems each person solves alone:
Reliability. The harness gathers, checks, and sources its answers the same way every time, which is what FMI's earlier direct connection could not do.
Cost that scales. Because every request runs through one place, the harness can match the model to the task, and the firm makes that choice once instead of rebuilding the setup for every person.
Flexibility through standardization. Improve a workflow once and everyone benefits. Because the harness watches its own operations and turns them into reports, the team can see what is working and refine the models and the workflows over time, instead of each person tuning in the dark.
WHO'S ASKING
A Strategy partner
ASKED IN ATLAS
“What's the full scope of FMI's relationship with a key building-products client?”
WHAT ATLAS REACHES INTO
15 engagements since 2018 and 18 opportunities, across 5 service lines
Signed proposals, final reports, and workplans from prior engagements
FMI market research and benchmarking reports relevant to the client's sector
WHAT SURFACES
The full client picture: every engagement, deliverable, and the firm's own research on their market, before the call.
38 records · sourced & cited
The same instinct shaped how skills spread through the firm. Individual staff had already built effective ways of working with AI, but those stayed with the person who figured them out. Cookbook, a shared skills library inside the harness, gives the firm a managed place to publish, browse, and reuse those skills with version control and company endorsement. What one person works out becomes something the whole firm can run.
Making it Stick
A harness only matters if people trust it and use it well, and earlier attempts to roll out AI at the firm had not landed with senior staff. So the work included a practitioner training program built on FMI's real workflows, delivered in four sessions across practice-specific cohorts. RoleModel's CRAFT and OWN frameworks gave the firm a shared language for structuring what they ask AI and judging what it gives back. The sessions taught people to use AI as a thought partner rather than a search engine, giving it the information to work from instead of expecting it to recall. And they covered working across tools, prototyping in one and handing the right task to the right tool and model, so the firm could match how they use AI to the work in front of them.
The discovery work delivered value before anything was built. The survey and interviews gave FMI's leadership a data-driven picture of how AI was really being used and what the firm's knowledge gap was costing: more than $250,000 a year in search alone, and several times the initial investment once duplicated effort and missed reuse are counted. Every decision that followed rested on that foundation.
The training gave the firm a shared standard for working with AI, addressing the specific quality concerns discovery had surfaced rather than AI literacy in the abstract.
Atlas is now live across the firm, and practitioners are using it to save hours on research. FMI’s CIO championed Atlas internally through rollout, and FMI and RoleModel continue to work together as it grows.
A System that watches itself
FMI's first attempt at AI search lost practitioners' trust because nobody could see when it was wrong. Atlas doesn't wait for someone to complain before it learns something is wrong. Every day, it reviews the questions people asked the day before. For each answer, it checks the cost, the time it took, and the searches that came back empty or ran in circles. It also collects every thumbs-up and thumbs-down people gave. Each morning, the team reviews a report of what went wrong and what to fix.
Once a week, a health check pulls those daily reports together with error alerts and the messages people posted about Atlas. It looks for the same problem showing up in more than one place, ranks the issues, and recommends which to fix first.
When the consulting or investment banking agent misses a file or gets stuck, the team replays that exact question. They change the agent’s instructions until the replay passes, then save it as a test so the same problem does not come back.
Atlas gets better on a schedule, not by chance. Each fix starts with a question a practitioner really asked.
Share this page