We sell AI software, so vague reassurance is not good enough. This page states exactly what the AI does inside Top Builder AI, the one thing it is architecturally forbidden from doing, who approves what, and what happens to your data. Where a claim below is enforced in code rather than by policy, we say so.
Top Builder AI runs advisory agents across booking, dispatch, finance, inventory, workforce, and document handling. Inside that product, a large language model (Anthropic's Claude) does a specific and limited set of jobs.
Notice what is not on that list: deciding, calculating, and committing. Those are handled elsewhere on purpose.
Every number the platform shows you comes from deterministic, unit-tested code, never from a language model. The model is given the finished figure and asked to explain it. It is never asked to produce it.
This is architectural, not a promise. The separation between the calculation layer and the language layer is a documented design rule in our codebase, and the learning system that improves the agents over time is wired only into the narration side. The figures cannot be reached by it, so they cannot drift.
The reason is simple. A language model can produce a wrong number that reads perfectly, and in a contractor's finances that is worse than no number at all. So we removed the possibility rather than trying to prompt our way around it.
Agents recommend. They do not reach into your ServiceTitan or QuickBooks and change things on their own. Any consequential action follows a gated path: the agent proposes, an authorized human approves, and only then does the system execute. Actions are off by default at the platform level and enabled per action type.
Approval and execution are restricted to admin users, and sensitive actions additionally require two-factor authentication. An agent cannot approve its own proposal.
Every agent has an autonomy setting that you control, and the cautious option is the default.
| Setting | What the agent may do |
|---|---|
| Off | Nothing. The agent is inert. |
| Approve | Analyze and propose. A human releases every action. |
| Auto | Act within the specific action types you have allowed. |
You can change this per agent at any time, and turning everything off is always one action away.
Automated agents do not make legally or similarly significant decisions about individuals. They do not hire, fire, discipline, or evaluate your employees for you. They do not determine anyone's credit, insurance, or eligibility for anything. They do not replace your CPA's professional judgment, your licensed trades' expertise, or your own call as the operator.
Where an agent surfaces something about a person, for example a workforce recommendation, it is decision-support for a manager who remains responsible for the decision.
Your business data is not used to train shared foundation models. We use AI providers under business API terms where the provider does not train on the data we send. We do not sell your data, and we do not pool one customer's data into another customer's results.
Tenant isolation is enforced at the database layer with row-level security, and the application connects as a non-superuser that cannot bypass it. The model receives only the context needed for the request it is answering, inside your tenant. The full list of companies that process data on our behalf is on the privacy policy.
Every action an agent proposes or executes is recorded with who approved it and when. Executed actions carry a compensating undo path, so a mistake is reversible rather than permanent. Sensitive operations are written to a tamper-evident audit trail.
If you want to know what an agent did last Tuesday and why, that is a question with an answer, not a shrug.
It will be, sometimes. That is the premise the whole design is built on, which is why the numbers are deterministic, the actions are gated, and everything is undoable. Using AI does not move responsibility from us to a vendor. If our product did it, it is ours to answer for.
Tell us what happened and we will show you the log, correct it, and fix the cause. We would rather hear it from you early than lose your confidence quietly.
We use AI tools in producing our written, audio, and video content, including blog posts and educational material. Claims about security, data handling, or how the product behaves are written and verified by a human before publishing, because those are commitments rather than commentary. Marketing content may publish on an automated schedule with human spot-checks.
Questions about how we use AI, or a request to review a specific agent action:
Salisbury Bookkeeping, LLC
7835 N. Escalante Dr, Eagle Mountain, UT 84005, USA
AI, privacy & security: support@salisburybookkeeping.com
Related: Privacy policy · Security · Terms
If we make material changes to this policy we will update the "Last updated" date above. If a change would weaken a commitment on this page, we will say so plainly rather than quietly editing the wording.