One workflow, owned by an agent.
One workflow that runs on a spreadsheet, a shared inbox, or somebody's memory, moved to an agent with a human approval gate, an audit trail, evals, and a dashboard your bookkeeper trusts.
For clinics, contractors, agencies, and logistics companies. Start with a map.
Thirty minutes with an engineer. One page back. Free.
- gate
- a person on money, deletion, outbound messages, anything irreversible
- audit trail
- every action the agent took, and why
- pricing
- per workflow, never per seat
Thirty minutes. One page. Which of your workflows an agent should own first, and why.
Owners of a clinic, a contractor, an agency, or a logistics company that still runs on spreadsheets and a shared inbox.
Most owners do not have an automation problem. They have a which-one-first problem. The map answers that with one page and no projection we cannot back up: which workflow an agent should own first, where the approval gate goes, and what we would need from you to build it.
- A one-page map of the workflows we can see from the call: what runs on a spreadsheet, a shared inbox, or a person's memory.
- A first-agent recommendation with the human approval gate drawn in.
- No dollar-per-month projection unless you give us the hours and the error cost; we do not invent them.
- Not a systems audit. One page, on purpose.
One workflow, owned by an agent, with a human approval gate and a dashboard your bookkeeper trusts.
An agent that owns a workflow is not an agent that runs your company. It does one job, it shows its work, and it stops and asks when the next step cannot be undone.
Quoted after the automation map or Sprint Zero, when we know which systems it touches. If the map says a scheduled script would do the job, we build the script and say so, because an automation dressed up as an agent is a thing buyers have learned to spot.
The on-ramp is Sprint Zero: a spec, a working slice of the agent, an eval harness that defines done, and a fixed price for the rest.
Tell us the workflow. A human engineer reads every message and replies within one business day, Pacific time. You get a scope, or the two questions we need first.
- An intake, support, scheduling, invoicing, or phone agent wired into the systems you already use: CRM, accounting, phone, email.
- Bounded autonomy: approval thresholds written down, and a person on anything irreversible.
- An audit trail of every action the agent took and why.
- Evals that run before every change, and a dashboard with a cost line.
- Tuning after it goes live, because the first month teaches everyone something.
- evals
- A test or eval harness that defines done in a way a machine can check, kept in the repo.
- observability
- Logs, metrics, and an error monitor you can open without us.
- runbook
- A written runbook: how to deploy, how to roll back, what pages whom, what the bill is.
- 30-day fix warranty
- Anything we shipped that breaks within 30 days, we fix without a new quote.
Stated once on the pricing page: evals, observability, a runbook, a 30-day fix warranty.
Agents propose. People dispose.
The dashboard your bookkeeper trusts is the one that shows every action the agent took, what it cost, and who approved the ones that mattered. The trust comes from the gate, not from the model.
The gates are written down at the start and they are yours to loosen as the evals earn it.
An agent can draft an invoice or a refund. A person releases it.
- the agent
- drafts an invoice or a refund
- a person
- releases it
An agent can propose. Nothing is deleted without a person, and everything deleted can be restored.
- the agent
- proposes
- a person
- deletes, and everything deleted can be restored
Templates the agent fills are sent. Anything it wrote from scratch waits for approval until the evals say it can stop waiting, and you decide when that is.
- the agent
- fills templates, which are sent
- a person
- approves anything written from scratch
If it cannot be undone, it has a gate. That is the whole rule.
- the agent
- stops and asks
- a person
- holds the gate
Companies that still run on a shared inbox.
The first agent is usually the same in each trade, because the workflow that costs the most is the one nobody owns. These are the ones we would map first. Yours may differ, which is what the map is for.
Clinics
The intake form, the reminder, the reschedule, and the no-show follow-up, with a person approving anything that touches a chart.
- agent runs
- the intake form, the reminders, the reschedules
- person on
- anything that touches a chart
Contractors
The estimate written from the site visit notes, the invoice sent when the job closes, the reminder when it is late. Payment release stays with you.
- agent runs
- the estimate, the invoice, the late reminder
- person on
- payment release
Agencies
The weekly client report, pulled from the systems you already have, and the shared inbox sorted by what needs a human today.
- agent runs
- the weekly report, the sorted inbox
- person on
- what needs a human today
Logistics
Dispatch and status. The order comes in, the driver is assigned, the customer is told, and the exception goes to a person before it becomes a phone call.
- agent runs
- dispatch and status
- person on
- the exception, before it is a phone call
Every document, ticket, and contract your company has, answerable with the permissions you already have.
Professional services, agencies, and any company drowning in tribal knowledge. Retrieval is a permissions problem before it is a model problem, so the system reads what you can already read and nothing else, and every answer says where it came from.
Often a component of a first agent. Quoted after we see the stores.
Tell us where the documents live. We reply with what we would connect first.
- reads
- SharePoint, Google Drive, email, ticketing, with the permissions you already have
- every answer
- carries a citation
- cannot cite
- says it does not know, and logs the question
- evals
- retrieval tested separately from generation
- Permission-aware retrieval over the stores you have: SharePoint, Google Drive, email, ticketing.
- A citation on every answer and a log of the questions it could not answer.
- Evals that test retrieval separately from generation.
- No chatbot that answers from memory. If it cannot cite a document, it says it does not know.
- No bypass of your existing permissions.
Refusals, in writing.
Bounded autonomy means the bounds are the product. These are ours.
- No agent for a thing that is just an automation. If a scheduled script does it, we build the script and say so.
- No autonomy on money, deletion, or outbound messages without a human gate.
- No per-seat pricing. You pay for the workflow, not for chairs.
- No projection of hours or dollars saved unless you give us the hours and the error cost. We do not invent them.
Map it first.
Thirty minutes with an engineer, then one page: which workflow, where the gate goes, what it would take. Free, and useful whether or not you continue.
Thirty minutes. One page. Free.
Fixed price, $3.5k to $7.5k. Credits to a retainer if you continue.