4 signals running unattended · an estimated 12 to 20 hours of account research down to a thirty-minute review · every project on the account covered by default
The Legacy Operating Model
The client sells impact analysis to renewable energy developers and corporate buyers. Their product is strong and their market is real. Finding the right moment to reach someone was the constraint.
The moment matters more than the list here. A developer becomes a buyer during a narrow permitting window. A corporation becomes a buyer in the weeks after signing a power purchase agreement. Miss the window and the same conversation goes nowhere.
So the work was research. Someone opened a federal filing, looked for projects that had entered permitting, and started digging.
The Glue
That digging was the whole job, and it never appeared on anyone's roadmap.
Federal generator filings list roughly 2,300 planned projects with size, location, technology and status. What they do not list is who owns the project. Filings name shell companies formed for a single site.
So a person searched. Press releases, interconnection queue references, state utility commission dockets, developer websites, until something confirmed the real owner. Then they went looking for the right human inside that company, worked down a bench of titles, and verified an email. Then they ran the impact model against the project's actual coordinates. Then they wrote the email.
The filing names a shell company. Finding out who actually owns the project is the job.
Every step was public information. None of it was fast. Our estimate, built task by task from the process itself, puts a full account map at twelve to twenty hours of focused work. Every project behind a developer's shell companies, the opposition picture on each one, and a named contact with evidence for every project.
Nobody spends those hours. They cherry-pick the three biggest projects, take one name each and move on. The rest of the account goes dark.
The Shift to AI-Native
The question that reframed the build was not how to speed the research up. It was what a person should be doing with those hours instead.
Research is retrieval, cross-referencing and judgment applied in a fixed sequence. Retrieval and cross-referencing scale. Judgment does not, and it should not have to. So the design put the machine on the sequence and left the human on the decision at the end of it.
That split is what makes the system worth trusting. Nothing goes out without a person approving it.
The System
Four detection pipelines run continuously. Two watch the developer side of the market, two watch the buyer side. All four converge on the same shape.
Detect. Each signal watches a different public surface. A monthly federal filing, compared against last month, catches projects entering permitting and projects falling backward through it. A live search agent catches organized public opposition to a specific project. A third watches for corporate power purchase agreements above a volume floor. A fourth watches for corporate projects facing public scrutiny.
Qualify. Scored gates decide whether anything continues. Thresholds are tuned so a single generic news article cannot fire a signal, and stale events decay out.
Identify. Shell company to real developer to a named person, each step carrying a source. Titles are ranked, so the outreach reaches the person who owns the decision.
Quantify. The client's impact model runs against the project's real coordinates, producing the numbers that make the message worth reading.
Draft. A written email lands in a shared drive with its source material attached.
Approve. A person reads the card, then approves, revises or skips. Approval is a button. Revision happens by commenting on the document.
Three rules hold the whole thing together. Nothing is invented, so every claim carries a clickable source or its confidence drops below the review line. Spending only happens on leads that will ship, so enrichment and model runs are gated behind the decision to send. And every failure explains itself, so a quiet week is diagnosable rather than mysterious.
Example In Practice
A 150 MW solar project appears in the monthly filing at permitting stage. It scores 5 of 5 on fit.
The system resolves the shell company to its parent developer and cites where it found the answer. It surfaces a VP of Development with a verified email and an evidence link. It checks the same project for organized opposition and reports none. It runs the impact model at the project's coordinates and returns the health, water and economic figures.
A card lands in a channel. The draft is written, the sources are attached, and three buttons sit at the bottom.
Total elapsed time, start to finish, is ten minutes.
How We Made It Happen
The build ran in slices, one signal at a time, each shipped and used before the next one started.
Autonomy was earned rather than granted. Every capability ran supervised first, and the supervised phase became the training data for the autonomous one. Contact resolution was validated against real permitting-stage leads before it was trusted.
Training happens where the work happens. A user comments on a draft in plain language and the agent either revises the document or files the note as a lesson, then confirms in the thread. No prompt engineering is asked of anyone.
The system was handed over with its documentation, its runbook and a knowledge transfer session. The client operates it.
The Impact
Research time per account. An estimated twelve to twenty hours of manual mapping becomes a thirty-minute review of finished research. The machine's ten minutes run unattended. The baseline is our estimate, built task by task from the same process the system replaces.
Coverage. Every project on an account researched by default, where the manual version cherry-picked the biggest three. Four signals watch the whole market continuously.
It repeats for free. The manual version was a one-time push that went stale. The map reruns on demand, ten prospects in a morning when the week calls for it.
Trust before autonomy. Every capability was validated against real leads and ran supervised before it earned autonomous status.
Where attention goes. The team reviews finished research instead of performing it.
What This Means
The reachable gain in most revenue operations sits in the connective work between systems. Looking things up, cross-referencing them, moving them from one place to another. It rarely has an owner and it never appears on a roadmap, which is exactly why it survives.
Ask the question that started this build. If this workflow did not exist today, how would it be designed?
The answer is usually that a machine handles the sequence and a person handles the decision. Getting there takes a working system, evaluated before it is trusted, installed where the team already works.
Ready to redesign a workflow?
Thirty minutes, one workflow of yours, timed end to end. We come back with what it costs today and what it would look like rebuilt.
No deck.