Trustworthy AI for the work that matters.
We build AI systems that carry real business workflows — client intake, document handling, follow-up, filing-season operations — with the domain expertise encoded and a named human gate on every consequential step. Delivered where your business already works: on the web, in WeChat, and inside ChatGPT and Claude.
- A human approves every consequential step
- One platform, four surfaces
- Demonstrations labelled as demonstrations
Demonstration · synthetic data, not a client
WeChat group
A client sends a photo
"Here's my W-2" — 11:42pm, in the chat the firm already uses.
Agent
The document is read and placed
Classified, filed to the right client and year, and checked against what that return still needs.
Agent
A reply is drafted
"Received your W-2. Still outstanding: 1099-INT and last year's return."
Human gate
The accountant approves before it sends
Nothing reaches the client until a person with the licence releases it. This step is architecture, not policy — the draft is held until it is approved.
What we do
Three ways in, one method underneath
Learn the method in a workshop, apply it to your own workflow through an assessment, or use a system we already built for your industry.
Proof
Products
Two systems we build, run, and stand behind. They are how you can judge the method before you buy any of it.
- Tax Aide — practice operations for CPA firms, working inside the chat their clients already use
- NestPilot Retirement — a transparent second opinion on the handful of retirement decisions you cannot undo
- One platform underneath both: web, WeChat, ChatGPT, and Claude
Your workflow
Applied AI services
A structured path from "we think AI could help here" to a working, measured workflow — or to a clear answer that it should not be built.
- Readiness assessment — your process, data, permissions, systems, and risk, examined before anything is built
- Measured pilot — one bounded workflow with a baseline, a target, human gates, and a stop condition
- Custom build — only where configuring or buying an existing tool genuinely will not do
Start here
Workshops
Vendor-neutral education for small businesses, delivered with community organizations. You leave with a workflow you designed, not a vendor list.
- "AI From Prompt to Process" — 75–90 minutes, built around your own recurring work
- Every business maps one real workflow and leaves with the card it built
- No product pitch: the session teaches a decision method, and some workflows should not use AI at all
Built and running
Two products that prove the method
We do not ask anyone to take the approach on faith. These are the systems we build, run, and answer for.
For CPA and accounting practices
Tax Aide 税务管家
Practice operations for CPA firms
Client intake, document collection, missing-item chasing, and filing-season coordination — handled in the group chat clients already use, with the accountant approving anything that reaches them.
Explore Tax AideFor households, advisors, and communities
NestPilot Retirement
Your Decision Copilot
Medicare, Social Security, Roth conversions, and withdrawal order modeled together, with every assumption inspectable — built for the client to verify the advice, not for the advisor to control the screen.
Explore NestPilot Retirement
The difference
Not a portal, and not a chatbot
The distinction that matters is not how clever the model is. It is where the work happens, whose knowledge is encoded, and who signs off.
| Dimension | Traditional software | Chat assistants | What we build |
|---|---|---|---|
| Who does the work | A person, one click at a time | A person, with drafting help | The system runs the steps; a person decides |
| Where it happens | In a portal someone has to open | In a chat window, separately | In the channel the work already arrives in |
| What it knows | Whatever was typed into the form | General knowledge, no context | Your process, your rules, your history |
| Who approves | Whoever clicked | Nobody — output goes straight to you | A named person, before anything leaves |
| When it is wrong | The form rejects it | You find out later | It holds and escalates rather than guessing |
How we talk about results
You should be able to check what a vendor tells you
This industry runs on numbers nobody can verify. We hold ourselves to two rules, and you are welcome to hold us to them.
A demonstration is labelled a demonstration
When we show you a workflow built on synthetic data, the slide says so. You will never have to work out whether what you are watching is a real customer.
Results are published only with permission
We publish a client result when the rollout has been accepted and that organization has approved the wording — on their timetable, not our marketing calendar. Anonymizing does not substitute for asking.
It follows that some things you would expect to see here are missing. We have no case studies published yet, because none has cleared both gates.
Start with the workflow that is costing you the most.
Come to a workshop and map it yourself, or ask us to assess it with you. If the answer turns out to be configuring a tool you already own, we will tell you that.