How we work
First the work. Then the AI.
The method, the Diagnostic and the rules every agent works inside, with two samples you can try.
01 The method
From your head to a working agent.
The drawing: one workflow, a client report, drafted in four passes. 1, Capture: where the work lives today (your CRM, SharePoint, email, spreadsheets and in your head) feeds one page, written down. 2, Structure: the page becomes steps, each with a check: pull the numbers, draft the summary, flag anything odd, send to the client. 3, Run: each step is tagged with what runs it: a script, a light model, a strong model, and for the last, you approve. 4, Review: before it is sent, a person says yes, and a receipt is kept for the run: 4 steps, 1 approval, every step logged.
02 The Diagnostic
Before we build, we learn how you work.
We sit with the people who run each part of your company, map how the work and the data really move, and show you where AI fits.
A sample map of a made-up company. Session 1: Operations, 45 minutes. Session 2: Client service, 45 minutes. Session 3: Finance, 45 minutes. Session 4, last: Leadership, 60 minutes, because it sees how everyone's work rolls up. New request, from Client service to Operations: typed twice. AI can do this. Hours, from Operations to Finance: lives in a spreadsheet. AI can do this. The numbers, from Finance to Leadership: waits on email. Your processes today: written down, checked and updated; or in people's heads, put on paper. Written down or not, you end with the same thing: your processes, ready for an agent. What you get: 1, the map. 2, first workflows worth building, with sample numbers: new request to job setup, 6 hours back a week; hours to invoice, 9 hours back a week. 3, your processes, written down.
03 Try it: capture
Tell us how it's done. Watch it become a workflow.
You're the COO of a made-up 60-person CPA firm. Tap or type your answers and watch each one become a step. Then one client file stops for your sign-off.
Sample Made-up firm · about 60 seconds · no sign-up · no AI in this sample
The sample: a project file for Wrenfield & Co. CPAs, a made-up firm of 60 people, in stages: capture, map, workflow and sign-off. Its chat opens: I'll help you map one task your team does over and over. It takes about a minute. First: what eats the most time in tax season? You pick one: chasing clients for missing tax papers; typing numbers from PDFs into our tax software; or answering “where's my return?” emails. You can also type your answer. Your answers draw the map of who does each step: the client, outside; Maria, admin; Dev, staff; Tom, manager; and the AI steps. Sample numbers for a made-up firm.
04 Try it: tune
Same work. Smaller AI bill.
Most teams run every AI step on the priciest model. Give each step Light, Medium or Heavy, and test every change on real files.
Sample Same made-up firm · sample prices and test scores · the math is under the demo
The sample: Wrenfield & Co.'s missing tax papers workflow, a made-up firm, 1,500 runs a month, seen as the architect or the owner. Its steps: 1, collect what the client sent: pulls new files from the portal. 2, read each page and label it: W-2, 1099, K-1, brokerage statements. 3, check it against last year's list: what came in vs. what we expect. 4, spot anything odd: a new rental? A K-1 that changed? 5, write the client a reminder: kind, clear, in your firm's voice. 6, manager signs off: before anything goes to a client. A person, always: not a model step. 7, mark the file ready for prep: updates your practice software. Each model step runs on a tier: Light, Medium or Heavy, or No AI where a plain rule does it. Every step passes the 98% bar on 200 past files. Week 1, every step runs on the heaviest tier and re-reads the whole client file. Architects trim it to the pages that step uses.
It all runs on Loup, our in-house workflow system. Clients get it by working with us.
05 Built for trust
Autopilot still has a pilot.
The agent works inside rules it can't change. We build every system this way, ours included.
If AI won't help yet, we'll tell you.
The six rules. 1, Your rules live outside the AI. The software enforces approvals and permissions, so no instruction can talk the agent past them. 2, Read-only until you say so. Agents start with the least access the job needs, and get more only when you decide. 3, If it doesn't know, it stops. When the agent is missing something, it stops and asks for it instead of guessing. 4, Nothing irreversible without a person. Anything that reaches a client, moves money or can't be undone waits for a person to say yes. 5, Every run leaves a receipt. Each one records what the agent did, which model it used, what it cost and who approved it. 6, Passwords stay out of workflows. Logins and passwords stay locked in your vault and never go into the steps an agent reads. The drawing: the agent's room, numbered by the six rules. 1, Rules outside the AI: the walls are your rules; the agent pushes on one and it can't move. 2, Read-only to start: at the door it may read, and write is struck out. 3, Stops when unsure: at a stop sign it asks for what's missing, a signed form, and gets it from you. 4, A person says yes: the gate out opens only when you sign. 5, A receipt every run: a sample receipt prints on the way out: 5 steps, signed: you, cost $0.61. 6, Passwords stay in your vault: the vault sits outside the walls, and the agent's path never reaches it.