The Work Is in the Training, Not the Build
What several months of running three AI agents in production taught me about the work after deployment.
News flash: AI agents aren’t going to take over the world. (Yet, anyway.) Training them — “harnessing” them, in AI lingo — is a lot of work. But the productivity rewards are real. And we need real people to manage them — which, for me, ultimately meant more work and more productivity.
Here’s what that’s actually looked like inside our firm.
The first week after our underwriting agent was running, it miscategorized a line item. Payroll costs mapped to contract services. Not catastrophically wrong — both hit operating expenses — but wrong in a way that would skew NOI by category. We corrected it and noted why. The agent folded the correction in. That particular error hasn’t recurred.
What automation promises
Automation, in the traditional sense, promises a one-time effort in exchange for ongoing removal of work. You build the system, you turn it on, it runs. Rules apply consistently. Inputs produce predictable outputs. That’s valuable — and it’s exactly what breaks when inputs vary, which in real estate they always do. Every seller’s chart of accounts looks different, every operating statement is formatted its own way, and no two rent rolls tell the story the same.
Rule-based systems can’t handle that variation. They hit an edge case and a human takes over.
Agentic AI handles variation differently — through pattern recognition rather than rigid rules. It can do things traditional automation can’t, and it gets better through use rather than just through deployment. The work is in the training, not the build.
What training actually looks like
We have three coordinating agents running now. Each requires a different coaching discipline.
The underwriting agent processes deal documents and maps them to our model. It learns from category corrections — when it maps a line item incorrectly, we note why, and that mapping improves. After each deal, the agent understands our chart of accounts better than it did before.
The investor OS agent extracts intelligence from meeting notes and updates a structured record of each investor relationship. Its training is continuous in a different way — investor preferences change as fund cycles turn. A capital partner who was actively deploying three months ago may be in a quiet period now. The agent learns from those updates as they’re made. Preferences that used to degrade as they sat unrecorded in someone’s head now get captured and updated in real time.
The deal-matching component sits at the intersection of both. When our underwriting system analyzes a new opportunity, it queries our database to surface the most likely capital partners — by check size, geography, strategy, and the contextual signals captured from prior conversations. The quality of that match depends directly on the quality of both upstream systems. Better investor records produce better deal routing. Better deal analysis produces more relevant matches.
These aren’t independent tools. They’re a system. The value compounds in proportion to the consistency of training across all three. We built three narrow agents rather than one general one on purpose: each works within a context small enough to hold completely, and a specialist that fits its problem outperforms a generalist stretched thin across all the problems.
The management discipline this requires
Running these agents is closer to managing a fast-learning analyst than deploying software. The onboarding requires real effort — establishing the structure of your model clearly enough that the agent has something to learn from, coaching early mistakes carefully, building the habit of noting corrections rather than just overriding them.
The payoff is a team member who retains what it learns and applies it to every deal that follows. After three months of deal-by-deal correction cycles, the agent’s output is materially better than it was at launch. Not because the software updated. Because we coached it.
What an agent doesn’t have is a low tolerance for repetition. People get bored, and rightly so — boredom is what pushes us toward the creative, judgment-laden parts of the work where we can add the most value. The agent’s patience for the repetitive parts complements that; it doesn’t replace it.
What doesn’t change
The agent reflects the quality of its manager. What we correct, it learns from. What we overlook — a miscategorization we don’t catch, an LP preference we don’t update — it incorporates as fact. Judgment about what the system should optimize for still comes from the people running it. So does the decision about when the output is trustworthy enough to act on.
A firm that engages with these tools the way it would engage a new team member will find something that compounds. After a year of that, you have a colleague who knows your business — and a team whose attention is freed for the work only people can do.
That’s the trade I didn’t expect when we started. More work, not less. But the work is the kind worth doing.
Next: what the build process actually looked like — and what we’d say to a firm that hasn’t started yet.