How we handle
your clients’ data.
You are putting your name on what this produces. Here is what it reads, what it deliberately does not do, and what happens when it cannot answer.
What data does it read?
AgentLink reads the systems you choose to connect: bank transactions through Plaid, your accounting ledger, and connected business systems such as your client’s CRM.
Accounting connections include QuickBooks, Xero and others. New connections are added on request. Payroll information is read only where it appears in the accounting ledger; there is no direct payroll connection.
Nothing is connected unless you connect it.
What does it deliberately not do?
AgentLink does not post to the books in its ongoing controller work. Closing a period calls for professional judgment. That responsibility stays with you.
This is a deliberate boundary, not a limitation we intend to remove. AgentLink is built to work between periods. Closing them is a different job.
It does not file anything, move money or connect to payroll systems, so it never sees individual employee records.
Clean Slate is the onboarding module for clients whose books have not been reconciled in years. It does write to the books, and it is off unless you turn it on for a specific client. Everything else is read only.
Can one client’s data reach another?
No. Each client’s data is isolated and never processed together with another’s. Access is scoped per firm and per client.
Your clients’ data is never used to train anything
It is never used to train any model, by us or by anyone else.
Inference runs on AWS Bedrock. AWS does not store inputs or outputs, does not share them with model providers, and does not use them to train models. That applies to every model on Bedrock, including third party ones, and it covers prompts, completions and embeddings.
Who owns the data, and what happens if you leave?
The data is your client’s. The product does learn, inside a client. What it learns about one company is that company’s, and it is never pooled across companies or applied to another firm’s clients. There is no shared pool.
If a firm disconnects, it can ask for a copy and ask us to delete what we hold, and the learning goes with it.
That is what ownership means here: you can take it and you can end it.
What happens when the data will not support an answer?
It refuses. When the underlying data will not support an answer, it says so and routes the question to you rather than producing something plausible. Every layer that declines to compute records why.
Every answer traces to specific transactions, and you can open any one to see what it was built from. Forecasts are written down before the period and graded against actual bank movement afterwards.
We publish how often it refuses.
Who can see it?
Production access is limited to the founder. Credentials are encrypted. Data is gated by encrypted access tokens, so a firm reaches only its own.
AgentLink is built by Tublian LLC in Columbus, Ohio. Bootstrapped. Built with practicing CPAs and fractional CFOs.
Questions about any of this go to a person, not a form. Write to Nilanjan Raychaudhuri, founder, at nilanjan@agentlink.finance. The answer comes from the person who built it.