AI agents for business operations.

It stays in your systems. Agents read through views you control and cannot write. The AI account is yours, with retention off and no training on your inputs. Anything we store is identifiers, timestamps, and counts, never narratives, case facts, or student records. The policy on who may see what is decided before the second wave begins and written into the setup.

It will be, sometimes. That is why a person decides every time, why every flag shows the numbers behind it, and why "I don't know" is a first-class output that routes to a human and records its reason. Being wrong in the open is the design. Being confidently wrong in silence is the failure we build against.

No new system of record. The fleet reads what you already run and writes drafts, lists, and reports into the places your people already look: a shared folder, a mailbox, a channel. Model usage is billed by your AI provider to your own account, at cost, never marked up by us.

Against the avoided hire, never the hour. Discovery is a paid product that credits toward the build. After that, a fixed build fee and a flat monthly that covers every agent on the list, support, monitoring, a monthly proof report, and a set number of build hours. Anything past them is a written change order with a fixed quote, agreed before work starts.

Anything that contacts a customer, client, student, or parent on its own. Autonomous changes to a live account. Outreach built on public arrest, court, or accident data, for anyone, at any price. Screen-scraping automation against software with no interface, which dies by a thousand UI updates. And we do not run unpaid pilots; a pilot can be discounted, not free.

You do. The engagement is work-for-hire. Code and configuration live in a repository you own, templates and rules are built from your practice, every record is exportable at any time, and a documented handoff means it keeps working if we part ways. Students own their portfolios; the school owns what was built for the school.

A chatbot waits to be asked. A fleet runs on a schedule against your own systems and hands a named person a finished piece of work before they open a tab — a brief, a reconciliation, a flagged list, a drafted letter. Each agent holds one job, the only tools that job needs, a written cap, and a log. Nobody has to remember to prompt it, and nobody has to check whether the answer came from your data or from the model's imagination.

No, and we will not sell it that way. The work we take is the repeated, checkable work nobody wanted: the list somebody rebuilds by hand every morning, the reconciliation, the chase. Your people check a list instead of building it, and the hours returned are counted by your team, not estimated by us. If headcount reduction is the goal of the engagement, we are the wrong firm.

Discovery is two to three weeks and ends with a written map of what you run and the first task worth taking. The first agent is usually live in assist mode four to six weeks after that, reading real data and handing its output to a named person. It is promoted out of assist mode only when it clears the accuracy bar written into the agreement — a date does not promote it.

The Claude Agent SDK and Managed Agents, with deny-by-default tool policy, spawn-depth and budget caps, schema-enforced outputs, and a watchdog that enforces the ceilings. It lives in a repository you own, in plain code and plain configuration, with a runbook. Your engineers can read it on day one and take it over on any day after that — we build for the successor, not for lock-in.

Yes, and that is the part we have run longest on our own book. Numbers come out of a warehouse rather than a platform UI, so a platform's modelled conversions never quietly become the truth: leads are matched to spend at the source, and every figure in a report cites the query behind it. The agent writes the report; a person sends it.