From Point-and-Click to Escalation Center: What Agentic Software Means for Deal Teams
We built Rubi the wrong way first. Software for deal teams is becoming an escalation center: agents run the work on rails, pull from your firm’s context, and come to you only when they need a decision.

Key takeaways
- Point-and-click software with an LLM underneath still makes you do the work.
- Agentic software flips it: agents do the work and escalate to you when they need a decision, in plain language.
- Agents need rails, meaning guardrails and the right tools, to do firm work reliably.
- Context is the moat: your emails, meetings, documents and CRM history, in a form agents can use.
- Chat tools are great for individuals but stall at the firm level without that context.
For a long time, AI products were point-and-click software with a model underneath. That's not where this is going. Software is becoming an escalation center: agents run the work on rails, pull from your firm's context, and come back to you only when they need a decision.
We built it the wrong way first
I’ll be honest about our own path. The first version of Rubi was point-and-click software with a model under the hood. You still clicked through screens, pressed buttons and moved things along. The AI made each step faster, but you were still the one doing the work.
We decided that was a mistake and paused growth to rebuild, because it isn’t where software is going.
The escalation center
In the new model, agents run the work, and you’re the escalation point. An agent logs the teaser, screens the CIM, updates the record and drafts the note. When it hits something it can’t or shouldn’t decide alone, it comes to you, in plain language, by email or chat. Which banker should this deal be attributed to? You answer in one line, and it carries on.
The shift sounds small but changes the job. Instead of opening software to do work, your software tells you where you’re needed. A deal team can have several agents running at once, the way an engineer now runs multiple coding agents in parallel.
Agents need rails
Firm work isn’t a demo. It needs to be right, repeatable and safe. So agents need infrastructure around them, which I think of as rails: the right tools, guardrails on what they can touch, approval gates before anything is sent or written, and a clear record of what they did and why.
That’s also why this has to be designed from the start. Every part of the system has to be usable by an agent through tool calling, not just by a person through a screen.
Context is the moat
The model is not the differentiator. Everyone has access to roughly the same frontier models. What makes an agent useful at your firm is your context: the emails, meeting transcripts, teasers, CIMs, memos and CRM history that make up how your firm actually works.
The building blocks are unglamorous. Make that context reachable by agents, in a structure they can use, tied to the right deals, companies, people and bankers. Meeting transcripts become memory. Every pass and pursue becomes a signal about what you like. Over time it all forms a context graph of the firm’s deals, owners, intermediaries and decisions, which is effectively the brain behind every agent.
An AI-native system ends up holding far more about your firm than a traditional CRM ever could, because it reads every document and conversation instead of waiting for someone to type a summary.
Why “just use ChatGPT” stalls
A line I hear from partners: ChatGPT is great for me. It’s not great for the firm. That’s right. A chat window is a fantastic personal assistant. But it starts from zero every time, it doesn’t know your criteria or your history, and its output lives in one person’s tab. Nothing gets logged, nothing compounds, and nobody else benefits.
Chat is still useful, and it’s a great place to ask an agent something. You should be able to reach your firm’s agents from Claude, ChatGPT, Outlook or Slack. But the work itself should run on rails, with your context, every time.
Closed loops compound
The last piece is how agents get good. Ours improve through tight loops with the deal teams using them: build, test, get feedback, repeat, every week. Each loop makes the agent better at that firm’s work, and the improvements stack. It compounds like capital.
That’s also why winning in this market is as much firm building as technology building. It’s a relationship business, and the software has to earn trust the same way people do.
Frequently asked questions
What are AI agents for private equity?
Software that does deal-team work on its own, such as logging deals, screening CIMs, updating the CRM and drafting outreach, and asks a person for a decision only when it needs one.
What is an agent harness?
The infrastructure around an agent: its tools, guardrails, permissions, approval gates and record of actions, which together let it do real work reliably and safely.
Why isn't ChatGPT enough for a private equity firm?
It works well for individuals, but it lacks the firm's context, doesn't log or share its work, and starts from scratch each time, so nothing compounds across the team.
How does a private equity firm become AI-native?
Make the firm's context reachable by agents, start with recurring workflows that have clear triggers, keep people approving key decisions, and improve the agents through regular feedback loops.