Return on investment
Model the cost and the payoff first. Only then pick the workflows worth handing to an agent.
Now streaming: Agents on Agents
Put AI to work
Put AI to work
where it pays.
Pacific Street advises companies on where AI agents belong in their operations, what they should cost, and where a person still signs off.
- Background
- Financial analysis & economics
- Focus
- Agent workflows in the enterprise
- Measure
- Return on token spend
Approach
Token spend is capital. We allocate it that way.
Deploying an agent with a job and a codebase, and nothing else, is what finance calls an unhedged position. We start from the return, then decide which workflows are worth restructuring. Every one is assessed against the same four criteria.
Fit with how teams work
Agents meet task owners in the tools they already use, such as Slack and Google Sheets. Changing their behavior doesn’t require a rebuild.
Safety & compliance
Before touching real work, every agent is tested in a sandbox we call the Treadmill. We measure whether it escalates when it isn’t confident.
Human sign-off
Clear rules on what an agent decides alone and what goes to a person: tight for anything that reaches the books, looser for a first draft.
What we mean by an agent
A model
The reasoning engine.
A substrate
The knowledge base and standing instructions it runs on.
Scoped tools
Only the ones it has been granted.
Work
Built and run in-house.
We operate the systems we advise on. Two of them, each built to stress-test one of the criteria above.
Return on investment
NativeMe
A daily planner maintained end to end by a multi-agent pipeline. The agents maintain the code that builds each schedule rather than being called on every request, so model spend stays flat as usage grows.
One-line support ticket → scoped requirement → reviewed pull request → one human sign-off.
Fit with how teams work
QuéCompré
A transaction categorizer for back-office teams. Provide your chart of accounts, sample fields and any house rules; the agent assigns each row to the right account.
Low-confidence rows go to the owner in Slack. The agent finishes the file once they reply.
Bohandra brings all four criteria together. Next: the Bohandra Agent Builder, where clients design their agents’ substrates with us.
About
Finance first.
Engineering second.
We came to AI from financial analysis and economics. So the first question we ask about an agent is what it returns for a business going AI-first, or AI-hybrid.
Miguel Arteaga
Co-founder
Former financial analyst. M.S. in Data Science, with development experience across the stack.
Miguel Ortega
Co-founder
B.A. in Economics, with hands-on development experience.
Insights · Webinar
Agents on Agents
Our first briefing: the questions a business should settle before writing its AI-agent policy, across return, fit, safety and sign-off.
Weighing an agent for a workflow?
Tell us what it is. We’ll tell you whether an agent belongs there, and what it should cost.