Cloisa AI
Trust
Client identifiers never leave the firm.
You are accountable for every SIN, SSN and account number in a client file. Cloisa is built around that fact, and the control model below is enforced in the system, not promised in a policy page.
Where the data goes
Names, SINs, SSNs and account numbers become tokens inside the firm's environment. The model sees tokens only. Real values are restored locally, in front of your tax professional.
The controls
Every control is enforced in the system, visible to the firm, and evidenced in the audit record.
-
A person commits
The agent proposes. Nothing reaches a return without a tax professional's approval.
-
Abstains when unsure
Below the confidence threshold it asks rather than guesses.
-
Fails closed
If identifier detection is uncertain, the document stops. It does not proceed.
-
Least privilege
Scoped access and segregation of duties on anything that touches client data.
-
Versioned, visible policy
Thresholds and rules the firm can read and change, with every version kept.
-
Complete audit evidence
Every extraction, score, rule applied and human override is recorded.
-
Kill switch
Suspend an agent, a workflow or an action type at once.
A claim we will not make. Because the token mapping is reversible, this is pseudonymization rather than anonymization. It reduces exposure. It does not remove the firm's obligations, and we will not tell you otherwise.
Sovereignty and security
Where things run matters as much as how they run. These are the commitments the product is being built to, stated carefully: a commitment is not yet a certificate, and we will not blur the two.
-
A Canadian company, Canadian infrastructure
Cloisa AI Inc. is a Canadian company based in Toronto. The product is being built to run its infrastructure in data centers operated in Canada, with the identifier vault, client documents and audit log living inside the firm's own environment or a dedicated Canadian tenant.
-
What crosses the border, if anything, is tokens
By the time any text reaches a reasoning model, names, SINs, SSNs and account numbers are already tokens, and the mapping back to real values never leaves the firm. That design holds wherever the model runs.
-
Client files never train models
Our policy from day one: client documents and drafted returns are not used to train AI models, ours or anyone else's. The product is improved against anonymized cases graded by practitioners.
-
Defense in depth
Encryption in transit and at rest, scoped access, segregation of duties, and the full audit record described above. Security is layered so that no single failure exposes a client file.
-
SOC 2: a commitment, not yet a certificate
We are building to SOC 2 control standards from the first line of code, with certification planned ahead of live client files at scale. Until the audit is done we will say exactly that, and the certificate will be announced in Notes when it exists.
Why the stakes are high
Cross-border filing mistakes are not rounding errors. The penalties for a single missed information form are set by statute.
-
US$10,000
Penalty per missed US information form, such as Form 5471 or 8938, with continuation penalties up to US$50,000, and the statute of limitations stays open until the form is filed.1
-
US$16,536
Maximum penalty per non-willful FBAR violation in 2026, inflation-adjusted from the US$10,000 statutory base. Willful violations run far higher.2
-
C$25 a day
CRA late-filing penalty for Form T1135 foreign income verification, to a maximum of C$2,500 per year, before gross-negligence penalties.3
1. IRC ss. 6038(b), 6038D(d), 6501(c)(8). 2. 31 U.S.C. s. 5321(a)(5), as inflation-adjusted for 2026. 3. ITA s. 162(7). Penalties depend on the facts; this is context, not advice. A system built for this work has to assume every field can carry this kind of consequence, which is why a tax professional approves every one.
The questions we expect
Asked by the practices and filers we talk to. Honest answers, current as of today; as commitments become facts, the answers get stronger in Notes.
Where does my client data live?
In the deployment we are building: inside the firm's own environment or a dedicated tenant in Canadian-operated data centers. The identifier vault and the audit log never leave that boundary. What reaches a reasoning model is de-identified tokens.
Is my data used to train AI models?
No. Client documents and drafted returns are not used to train models, ours or anyone else's. That is policy from day one, and it will be written into every agreement.
Are you SOC 2 certified?
Not yet, and we will not imply otherwise. We build to SOC 2 control standards and plan certification ahead of live client files at scale. When the certificate exists, it will be announced in Notes.
What happens when the AI is unsure or wrong?
Below its confidence threshold, the system asks instead of guessing. If identifier detection is uncertain, the document stops entirely. And nothing reaches a return without a licensed tax professional approving it field by field, with every figure linked to the exact page it came from, so an error is caught at review, not after filing.
Who sees my client's identity?
The firm, and only the firm. Identifiers become tokens inside the firm's boundary, the reasoning model sees tokens only, and real values are restored locally in front of the tax professional who signs.
I am a filer, not a firm. What should I do?
Your return should still be prepared and signed by your own licensed professional; Cloisa is a tool for them, not a replacement. If someone prepares a coordinated Canada–US return for you, send them this site.
This site practices what it sells
The control model starts with our own website.
This site sets no cookies, loads nothing from third parties, runs no scripts and keeps no analytics. Nothing about your visit reaches anyone, including us.
Questions about the data boundary, deployment, or how the controls are evidenced: hello@cloisa.com.