Responsible AI

A manager opens the advisor from their dashboard and picks a question from a fixed list — where profit is leaking in F&I, which salespeople need coaching, how the store is tracking against goal. The product assembles the relevant metrics, sends them with the question to a language model, and displays the answer as structured suggestions. Each answer is stored so that the next one can build on it.

It is a reading of numbers you already have. It does not have access to anything a manager could not already see on their own dashboard, and it cannot take any action in the product.

This is the section people actually want, so it is specific. Every time someone asks a question, the following leaves our infrastructure:

• Aggregated dealership performance figures: deal counts, front and back gross, PVR, product penetration, and the trailing three months of the same
• The names of individual salespeople and finance managers, each with their own deal count, gross, average, and product penetration — the model is explicitly asked to name top performers and people needing coaching
• The dealership’s name, brand, and type, on the general-manager view
• The question you selected, from a fixed list — the interface has no free-text box
• A short summary of previous advice, so answers build on each other

Verified by reading the code that assembles the payload, not by policy. The interface builds an explicit object field by field; it never passes a raw deal record through, so these fields have no path to the model:

• Any vehicle buyer’s name, VIN, telephone number, or any other consumer detail — none of these fields exist in the payload the interface assembles
• Any customer’s financing terms
• Login credentials, session tokens, or payment details
• Data belonging to any dealership other than your own

The code supports three providers — Anthropic, OpenAI, and xAI — selected from a configuration record. Only Anthropic is configured. No key exists for the other two, so those code paths cannot run, and the administrative screen that exists changes only which Anthropic model is used, not which company receives the data. If that ever changes, this page and the Privacy Policy change in the same release.

• It is generated text. It can be confidently wrong, and it will sometimes produce a plausible-sounding explanation for a number it has misread.
• It is not financial, legal, tax, compliance, or employment advice, and our Terms say so as a term of the contract.
• It must not be the sole basis for a decision about someone’s pay, discipline, or employment. This is not boilerplate: a system that ranks named employees and suggests who needs attention sits close to a line that data protection law draws around automated decision-making, and the thing that keeps us on the right side of it is a human manager reading the output and deciding. The product cannot act on its own — it has no mechanism to change a pay plan, a schedule, or a role.
• Each answer is stored against the dealership that requested it, along with the metrics it was based on, so a suggestion can be traced back to the numbers that produced it.

One other place a third-party model touches data, disclosed for completeness even though no customer of ours is involved: our own sales team can photograph a business card collected at a trade show and have the text read into a contact form. The image goes to OCR.space. The cards belong to dealership staff we are selling to, not to your customers and not to your employees — nothing in a dealership's workflow reaches it. Access requires a login and is rate-limited to twenty scans per user per five minutes.

Every dashboard carries a small question mark in the bottom corner on a desktop or tablet screen — it is deliberately not shown on a phone, where it would cover the figures it is there to explain. It answers questions about how the product works — where a setting lives, what a status means, how to do something. It is a different feature from the AI Advisor above, with a different provider relationship and, most importantly, a different diet: it is sent none of your data.

That is the complete list. In particular it is never sent:

• Any deal, unit count, gross figure or other performance number
• Any pay plan, commission or earnings figure
• Any vehicle buyer’s details
• The name or identifier of your dealership or dealer group
• Any employee’s name, including your own

The reason we can state that flatly rather than as a policy: The code that assembles the request has no access to any dealership data, and an automated test asserts that the message sent contains only those four items. Adding anything to that message would fail the test and would require this page to change in the same commit — which is the point of writing it that way.

The guide it answers from is written by us and shipped inside the software. The model is instructed to answer only from that guide and to say so plainly when the guide does not cover something, rather than inventing a screen or a button. It cannot see your data, so it cannot answer “how many did I sell” — it will tell you which screen shows you that instead. It cannot take any action in the product.

It goes to Anthropic, the same provider as the AI Advisor, under the same terms described in section 5. It is capped at 25 questions per person per day and your question is capped at 400 characters. If it is unavailable, the panel falls back to searching the same guide on your own device, with no network request at all.

It is deliberately absent from screens where the person at the keyboard is not the account holder — during a live demo, a support impersonation session, or while an administrator is viewing an employee's dashboard — and from every public page, including this one.

Both features are optional. If your dealership would prefer either disabled, email support@thedasboard.com and we will disable it for your tenant. Questions about the data handling on this page go to privacy@thedasboard.com.