We build and host practical business agents that remember the work, respect hard boundaries, and stay maintained over time. The model underneath is rented and getting cheaper for everyone. The operating memory, the judgment, the upkeep, and the relationship are the durable part.
Most AI hands you another thing to manage. We do the opposite. At Petrichor takes the recurring jobs that make a week heavy and runs them quietly in the background, so they stop living in your head.
Every inquiry gets logged, followed up, and surfaced when it needs you. Nothing slips.
Job details gathered, a draft quote prepared, waiting on your one approval.
Past customers, open jobs, and quiet weeks get the nudge they need, on schedule.
Start with one. Add more as they prove themselves. They begin to hand off to each other, and the week gets lighter.
Most AI tools talk well for a week and then drift. At Petrichor builds the working layer around the model: memory, rules, review gates, receipts, and maintenance. The point is not a flashy demo. The point is a business system that still makes sense next month.
Customers, jobs, follow-ups, receipts, and owner preferences are kept in a durable operating record.
No rogue spending, sending, scheduling, quoting, or destructive moves. The agent stops where the owner should decide.
Important work leaves a record: what fired, what changed, what stayed blocked, and what needs review.
We host it, tune it, and keep the boring parts alive after launch, so the business does not inherit an unsupported experiment.
Every build is scoped after we understand the workflow, the records involved, the approval gates, and the maintenance load. You see the shape before we build.
A useful agent needs more than a prompt. It needs a protected home, scoped tools, review gates, logs, and a human owner above anything sensitive. You reach the agent through one watched door; private records stay in their lane. Hover the diagram.
Requests pass through a brokered route where keys, spend caps, and tool limits can be reviewed instead of hidden inside a prompt.
Customer details, working notes, receipts, and job records live in approved private destinations, not public proof pages.
The agent can prepare and organize. It does not send, quote, schedule, invoice, or spend unless the owner approves that route.
Pick the job that makes your week heaviest. We scope it, build one bounded agent around it, and you approve every customer-facing move. Small first, useful fast. Add more as they prove themselves.
Start with a real workflow, not a generic demo.
No. Agents prepare and organize. They do not send, quote, schedule, invoice, spend, or make destructive changes unless you approve that route. The boundary is the operating model.
In approved private destinations. Customer details, lead notes, and receipts stay in their lane, not on public pages. A dedicated server can put the work on a machine scoped for your company alone.
We start with the real workflow, the records involved, the tools the agent may touch, and the actions that must stop for owner approval. The result is written clearly before the build begins.
We host it, tune it, and keep the boring parts alive so the business does not inherit an unsupported experiment. Maintenance and review are part of the model.
The working layer around the model does the work: memory, rules, review gates, and receipts. Claims are marked verified or flagged before they leave.
With one loop, usually a Lead-to-Job pilot around a single real workflow. Add more once they prove themselves and begin to hand off to each other.