Skip to content
roadmark

Questions to ask a TMS vendor about AI and your data, and what a good answer looks like

By Roadmark Team5 min read

  • ai
  • security
  • buying a tms

Most TMS vendors now ship AI that drafts quotes, answers track-and-trace emails, matches invoices or plans loads. The demo shows what it can do. The questions below are about what it can't do, and what happens to your data. They're written to be vendor-neutral: use them on anyone, Roadmark included. Our own answers are at the end.

A good answer has three things: it's specific, it says where the control lives (a setting, a database rule, a contract clause), and it's the same in the contract as in the demo. "Our AI is secure" is not an answer.

1. Can the AI act without a person approving it?

Why it matters. An AI that sends a quote, tenders a load or posts an invoice on its own is making commitments in your company's name.

A good answer: the default is that the AI drafts and a person approves. Acting alone is something your team turns on deliberately, for a defined scope (one customer, one lane, one kind of task), and the vendor can show you the setting and who's allowed to change it. Ask what the AI does with a customer on credit hold or a carrier that failed compliance: the answer should be "the same as a person: nothing."

Watch for: autonomy that's on by default, or set once for the whole company.

2. Can you switch it off, and does everything still work?

Why it matters. You may need to pause the AI during an incident, a dispute, or a model provider's outage.

A good answer: it can be switched off per person, per team and per company, plus one switch that stops everything, and every screen still works by hand while it's off. Ask the vendor to show it live.

Watch for: AI that can't be disabled, or workflows that only exist as AI features.

3. Does the AI see only what the user can see?

Why it matters. If an assistant can answer questions across companies, customers or rates the person asking isn't allowed to see, your permissions don't really exist.

A good answer: the AI runs as the person who invoked it, with exactly their access and limits, never a broader service account. Dollar limits and approval thresholds apply to it the same way.

4. How is each company's data kept apart?

Why it matters. A multi-tenant system holds many customers' data, and a group of companies may have data that must stay separate even internally.

A good answer: isolation is enforced in the database, for example with row-level security on every table that holds company data, not only by the screens deciding what to show. That way the same rule applies to the screens, the API, integrations and the AI.

Watch for: "each customer has their own login" offered as the isolation story.

5. Is every AI action in an audit trail you can export?

Why it matters. When a customer disputes a quote or an auditor asks who approved an invoice, you need the record, not a recollection.

A good answer: every action is recorded with who took it (a person, a named agent, an AI assistant or an integration), what changed, before and after, and when. The record can't be edited after the fact, and you can export it, for any date range, as a file.

6. How do you get your data out when you leave?

Why it matters. Loads, rates, documents and history are yours. A switch is only realistic if they leave with you.

A good answer: a complete export in a usable format, a stated deletion schedule for what remains, and both written into the contract.

7. Where is the data hosted?

Why it matters. Customers, insurers or privacy law may care about where data lives. In Quebec, for example, a business must assess privacy factors before communicating personal information outside Quebec (Act respecting the protection of personal information in the private sector, s. 17).

A good answer: the provider and region in writing, including backups, and an honest statement of what's processed elsewhere, such as AI inference, email or support tools. "Data stays in Canada" is only true if every one of those does.

8. Which subprocessors see your data?

Why it matters. Your data reaches every service the vendor relies on: hosting, email, support desks and AI model providers.

A good answer: a current list, including which AI model providers process prompts and results, and notice before the list changes.

9. Is your data used to train AI models?

Why it matters. Your rates and customer lists are commercially sensitive, and nobody wants them shaping a model that serves a competitor.

A good answer: a clear statement of whether the vendor uses your data to train or improve models, for its own models and for the model providers it uses, and what governs a provider's handling of it (the vendor's contract with that provider, or your own plan with it). Get it in the contract.

Roadmark's short answers

QuestionRoadmark
AI acting without approvalEvery agent starts at Draft: a person keeps or rejects what it prepares. Your team raises autonomy one lane or customer at a time, with dollar limits; credit holds and compliance blocks stop agents as they stop people. See AI controls.
Switching it offPer person, team or company, and one switch pauses every agent. Every screen, quote and load works without AI.
Seeing only what the user seesAn agent, or a question asked in Claude, ChatGPT or Gemini, runs as the person who called it, with exactly their access.
Per-company isolationRow-level security in the database, forced on every table that holds company data. See Security.
Audit trail of AI actionsEvery action is recorded: who (a person, an agent, an assistant or an integration), what changed, before and after, and when. Database triggers write it, and it can't be edited afterwards. Export as CSV or JSON for any date range.
Export on exitYou keep an export of your records, and what remains is deleted on the schedule set out in your contract.
Hosting regionWe confirm the hosting region and the scope of data residency, including any external processing, for your deployment before you commit.
SubprocessorsAsk us for the current list, including the AI model providers we use.
Training on your dataRoadmark doesn't use one customer's data to serve another. What a model provider does with a conversation in Claude, ChatGPT or Gemini depends on your plan with that provider.

Roadmark is priced from $499 USD a month, with unlimited users. If you're comparing vendors for a Canadian fleet, the Canadian TMS page covers French, CAD and USD billing, and the cross-border paperwork.

Questions and answers

What should I ask a TMS vendor about AI?

Whether the AI can act without a person approving it, whether you can switch it off per company and keep working, whether it inherits the user's permissions and limits, whether every AI action is in an audit trail you can export, and which model providers process your data. Then ask about data: isolation between companies, export on exit, hosting region, subprocessors and training.

What does a good answer on AI approvals look like?

The default is that the AI drafts and a person approves, and acting alone is something your team turns on deliberately, for a defined scope, with limits. The vendor should be able to show where that setting lives and who can change it.

Should a TMS work with the AI turned off?

Yes. Every screen and workflow should keep working by hand when you switch the AI off or when a model provider is down. If turning AI off breaks quoting, dispatch or billing, the AI isn't optional.

What should a TMS audit trail record about AI?

Every action, with who took it (a person, a named agent, an AI assistant or an integration), what changed, the before and after values, and when. The record shouldn't be editable after the fact, and you should be able to export it for any date range.

Does Roadmark's AI act without approval?

Not by default. Every Roadmark agent starts at Draft, so a person keeps or rejects what it prepares. Your team can raise an agent's autonomy for one lane or customer at a time, with dollar limits, and credit holds and compliance blocks stop agents the same way they stop people.