How to Roll Out Claude Across an Investment Firm: A 90-Day Playbook

1. Buying Seats Is Not a Rollout

A firm buys forty Claude seats, sends an email with the login link, and calls it an AI rollout. Ninety days later six people use it, thirty-four forgot their password, and the partner who approved it is wondering what the money bought.

That is the default outcome, and it has nothing to do with the tool. Buying seats is a procurement event. A rollout is a change in how people work, and people do not change how they work because a new icon appeared on their screen.

This playbook is the ninety-day version of doing it properly: decide the data rules, pick one workflow, prove it, then spread what worked. It is deliberately narrow. A rollout that tries to change everything changes nothing.

2. The Mistake That Wastes the First Quarter

The most common mistake is breadth. The firm announces that everyone should use AI for everything, runs a one-hour training, and waits. Nothing happens, because "use AI for everything" is not an instruction anyone can act on Monday morning.

Breadth feels ambitious and produces nothing. Depth on one workflow produces a result you can point to, and a result is what creates the next ten users. People adopt a tool when they watch a colleague save a day on something they also hate doing. They ignore a tool when they are told it is important.

Pick one thing. Make it work. Let the result recruit.

3. Decide the Data Rules First

Before anyone logs in, decide what data is allowed where. This takes an afternoon and prevents the incident that ends programs.

Two rules cover most of it. First, the firm runs on a commercial plan (Team or Enterprise), never personal Free or Pro accounts, because the commercial plans do not train on your data and the consumer ones can. Second, write a one-page list of what can go into the tool and what cannot, in plain language with examples, not a ten-page policy nobody reads.

The full version, including the vendor questions and the access controls, is in the is Claude safe guide. Do not skip it, and do not let it become the reason you never start. An afternoon is enough to begin safely.

4. The 90-Day Playbook

The shape of a good rollout is the same at every firm. Build a small foundation, win one workflow, then standardize and spread. Ninety days is enough to get through all three and have a real result to show.

The rest of this guide is each phase in detail. The temptation at every stage is to go wider. Resist it.

5. Days 1 to 30: Foundation and a Beachhead

The first month is plumbing and a single beachhead.

Set up the commercial plan, turn on single sign-on and the admin controls, and add the people who will actually use it first. Do not add everyone. Add the team that owns the one workflow you are going to win.

Then build the firm's context once. Put your investment criteria, your house memo format, and a starter knowledge base into a shared Claude Project, so the tool knows how your firm works instead of starting blank every time. That is the difference between a generic chatbot and something that feels built for you, and it takes a day.

Pick the beachhead workflow now. The best first workflow is high-volume, reading-heavy, and hated: CIM screening, data-room review, monthly portfolio reporting. Something the team does constantly and dreads.

6. Days 31 to 60: Prove One Workflow

The second month is proof on one workflow.

Run the beachhead for real, every day, with the people who own it. Sit with them. Watch where Claude helps and where it gets in the way. Fix the prompts, improve the Project instructions, build a shared template so everyone does it the same way instead of each person reinventing it.

Measure the before and after on one number. How long did screening a CIM take before, how long now. How many hours did the monthly portfolio review take, how many now. One honest number on one workflow is the asset that funds the rest of the rollout, because it converts the skeptics who do not respond to enthusiasm.

By day 60 you want a workflow that is faster, a team that prefers the new way, and a number you can put on a slide.

7. Days 61 to 90: Standardize and Spread

The third month is standardize and spread.

Standardize first. Write down how the beachhead workflow is now done, and turn the good prompts into a shared, reusable form (a Project, a saved prompt, a custom setup) so a new person inherits the capability instead of building it again. An undocumented win that lives in one analyst's head is a single point of failure, not a firm capability.

Then spread to the next workflow, using the same pattern: one workflow, the right people, prove it, standardize. Do not try to do five at once now that you have momentum. The discipline that worked in month one is the discipline that works in month four.

At ninety days a well-run firm has one workflow fully in production, a second one starting, a documented way of working, and a number that justifies going further. That is a real foundation. Most firms never reach it because they tried to start everywhere.

8. A Champion, Not a Committee

Every rollout that works has a person, not a committee. Someone whose job, official or not, is to make this happen: the one who builds the Projects, fixes the prompts, answers the "how do I" questions, and shows the partner the number.

This is usually not the most senior person and usually not IT. It is someone respected by the team who is genuinely interested. Find that person, give them the time and the air cover, and the rollout has an owner. Spread the responsibility across a committee and it has none.

If you do not have that person internally, this is the role an outside partner plays, and it is the core of how we work as an embedded AI partner.

9. Training Is Hours, Not Months

Firms overestimate how much training a rollout needs, which becomes a reason to delay.

Andrew Ng's long-standing guidance holds up: most people need a few hours of practical AI literacy, the leaders carrying out projects need somewhat more, and only the small group building need real depth. For an investment firm that means the partners and associates need an afternoon of how-to-use-it-well, not a course.

The training that matters is not a lecture but a working session: showing the team the firm's own Projects and prompts on the firm's own work, so they leave with something they will use on Monday, not a certificate.

10. Measuring Whether It Worked

Measure two things, and resist measuring twenty.

Usage and value. Usage tells you whether people actually adopted it (Enterprise gives you the admin view; otherwise ask). Value is the time saved on the workflows you targeted. If usage is high and value is real, expand. If usage is low, you have an adoption problem, not a tool problem, and adding features will not fix it.

Vanity metrics like "messages sent" tell you nothing. The number that matters is hours returned to the work that needs a human, on a workflow you can name. Why adoption stalls, and how to fix it, is the whole subject of why AI rollouts fail.

11. Where to Start

Decide the data rule this week: commercial plan only, a one-page allowed-list, no personal accounts for firm data. That alone makes starting safe.

Then pick the one workflow where reading eats the most hours and commit to winning it in sixty days. Not five workflows. One.

If you want a partner to run the ninety days with you (set up the plan and controls, build the firm's Projects, win the beachhead, and stand up the system the rest of the firm grows into) a Discovery Sprint starts it and the AI Operating System is where it goes.