On almost every call with a new firm, someone asks the same question: how are other firms actually doing this?

Rolling AI out across an investment firm doesn't come with a playbook. Most firms have run a pilot or two, watched usage spike and fade, and are still working out what makes it stick.

We have an answer, because we've done this at firm after firm. Our playbook comes down to three principles, refined across every rollout we've run. This year we tested it: across every Metal customer, from January to October 2026, the data backs up each one.

1. Give everyone the firm's context, and adoption follows

Most firms assume the value of AI starts with automating a workflow. In practice, it starts earlier: with everyone able to reach what the firm already knows, from a company's history to every banker conversation to the memo from the last time the deal came through.

At one multi-strategy firm, we started small: a single deal team. Within weeks, a team from a second strategy had found Metal on its own. Nobody had rolled it out to them; the first team's context was simply useful for their work. More teams followed, and each one's knowledge made Metal more useful for the next. That firm now runs Metal across nine functions, from investment teams to operations.

The same pattern shows up across our customers:

• For three in four active users, Metal’s context layer sits behind every part of their work. They don’t run a prompt and leave. Before they write an email, draft a memo, prep a meeting, or make a call, they check with Metal for what the firm already knows, no matter which AI tool they use. That group has grown 64% since January.

• Usage nearly doubled. Active users across our customers grew 96% from January to October.

What we recommend: Open access from day one and bring people on in groups. No team should wait for its own workflow to use the firm's context.

2. Meet every person where they already work

The fastest way to stall an AI rollout is to ask busy people to open one more app. The fastest adopters did the opposite: they brought the firm's context to the tools each person already used.

That looks different for every role:

• A partner who lives in email gets an email agent that answers from the firm's history.

• A deal team or a function chats with the Metal agent directly in Slack or Teams, where its conversations already happen.

• People who live in Claude or ChatGPT reach the same context from inside those tools through Metal's MCP server.

• Anyone who wants to go deep uses the Metal platform, whether to run customized workflows or to get shared visibility into the deal pipeline and portfolio through dashboards.

This works because Metal isn't a screen. It's the firm's context layer: model-agnostic, tuned to each firm and delivered wherever the work happens.

The shift has been fast:

In February, none of our customers had connected Metal to an AI tool. By October, 78% of all interactions with Metal came through one. At one newer customer, 90% of interactions now run through Claude, and active users grew 4x after they connected it.

What we recommend: before choosing how to roll out, ask where each group already works. Then put the context there.

3. Trust is earned one run at a time

Every firm wants agents that do the work on their own: a morning digest of new deals that fit the thesis, portfolio KPIs tracked without anyone asking, every banker call logged to the CRM. Plenty of firms get there. None of them started there.

The pattern is the same everywhere. A team picks one workflow and runs it by hand, checks the output against its own judgment, and fixes what's off. Only once it trusts the workflow does it let it run on a schedule.

People ran more workflows by hand this year than ever: four times as many in October as in January. Over the same months, scheduled runs grew to more than two-thirds of all runs. Automation didn't replace people; it grew alongside them, built on their judgment. Agents do more of the work, and people still decide what that work should be.

What we recommend: give each team one anchor workflow and run it by hand until it earns trust. Teams shouldn't wait on each other, and nothing gets automated until someone has checked it.

Where it leads: better judgment, not just more usage

The three principles get a firm started. But the real measure of AI at an investment firm isn't how many people use it. It's whether the firm makes better calls.

During a visit to one customer, we watched their investment committee open on a Metal dashboard. Partners were querying it live, minutes before the discussion. They pulled up the history on a company the firm had looked at twice before: what they'd learned, what they'd passed on, and why.

Nobody asked them to change their IC process. They changed it themselves, because for the first time the firm's full context was in the room at the moment of decision. We now see the same shift in pipeline reviews and fundraising meetings.

That's what we build toward with every customer: adoption stops being a usage number and becomes how the firm decides, with everything it has learned in front of it when it matters most.

So, why not just use Claude?

At the start of this year, our industry was talking about little besides agentic workflows, and almost every prospect asked us: why can't I just use Claude?

Our answer: you absolutely should. Use Claude, use ChatGPT, use whatever model comes next. Metal works with all of them, and most of our customers' interactions with Metal already happen inside those tools.

But a model is only as good as the context it's given, and a firm needs more than a stack of individual subscriptions. Put firm-wide context on shared infrastructure underneath it, and every person and every model works from the same picture of every deal, company and relationship. The firm also gets what individual AI subscriptions don't provide: control over who sees what, permissions that match how the firm is organized, and clear visibility into usage and cost.

The work flows back, too. Every memo, call note and workflow run adds to the firm's collective knowledge, so the firm gets smarter with every interaction. That's why it compounds: at firms that have been on Metal all year, active users are up nearly 60%.

None of this starts with a big launch. It starts with one team, one workflow and the firm’s context in front of the people who need it, , and let it build from there. A year in, that context is their edge, and it grows every time someone uses it.