Scope: what does it actually own?
Scope questions establish whether you are buying capacity or a tool, and they must come first because the answers determine which control questions are even relevant.
Ask what the role is accountable for, what it does unprompted, what it decides without asking, what it escalates, and how it knows what to work on next. A vendor selling function ownership answers all five in operational terms.
- What is this role accountable for, and where is that visible?
- What does it do on Monday morning without being prompted?
- What does it decide for itself, and what does it escalate?
- How does it know what to work on next?
Control: what can we stop, cap, and reverse?
Control questions establish your ability to contain a mistake, which is the difference between a manageable deployment and an unbounded one.
Ask which actions are gated by default, whether the list is configurable, whether autonomy is granted per class of action, whether it can be reduced again, what the spend caps are, and whether those caps are hard limits or alerts. The distinction between a cap and an alert is not pedantic; it is the difference between cannot and did not.
- Which actions are gated by default, and can we change that list?
- Is autonomy granted per class of action, and can it be reduced?
- Are spend limits hard caps or alerts?
- How do we pause everything immediately?
Data: where does it live and who can reach it?
Data questions cover both what the system stores and what it can reach, and buyers often ask only the first half.
Ask how credentials are stored, whether your data trains anyone's models, how tenant isolation works, whether access is scoped per property, and how you revoke a connection. Then ask where the work product lives: a system that writes into your ledger and publishes to your site leaves you with the output when the contract ends.
- How are integration credentials stored, and are they encrypted at rest?
- Is our data used to train models?
- How is one customer isolated from another, and how is access scoped?
- Where does the work product live — our systems or yours?
Failure: what happens when it goes wrong?
Failure questions are the most revealing and the least often asked, because they require the vendor to talk about the product at its worst.
Ask what happens when a long-running job fails at two in the morning, how the system signals that it is uncertain, what happens when a source cannot be verified, how you reconstruct a decision six months later, and what the last incident was. A vendor with a production system has specific answers; a vendor with a demo has reassurance.
- What happens if a multi-step job fails halfway overnight?
- How does it signal uncertainty, and to whom?
- How do we reconstruct what happened and who approved it, months later?
- What was your last incident, and what changed afterwards?
Commercials: how does the cost actually behave?
Commercial questions should establish how cost behaves as usage changes, not just the headline number, because usage-based models can surprise in both directions.
Ask what drives the bill, what happens when the allowance runs out, whether it stops or continues charging, whether unused allowance carries over, and what the exit terms are. On ftes.ai the model is a block of work metered in tokens with a hard stop at zero, which is deliberately unsurprising, but every vendor should be able to describe their equivalent precisely.
- What drives the bill, and what makes it go up?
- What happens at the end of the allowance — hard stop or overage?
- Can we start with one role and one property?
- What are the exit terms, and what do we keep?
Governance: who decides, who can stop it, and what can never be automated?
Ask who approves each action and how: is every action gated behind a named person until that specific class of action has earned autonomy from a record, or is there a single trust setting that grants everything at once? Ask what the never-automated list contains, meaning the actions that stay human at every level, such as rejections, offers, refunds, pricing, and contract terms, and ask to see it rather than hear about it. Ask how autonomy is stepped back, how long it takes, and whether the step-back is logged.
On ftes.ai the answers are the approval gate in the Approvals tab, the autonomy ladder per class in Setup, the always-escalate list in each role's remit and the never-automated list fixed in the platform's design, and a one-click step-back that takes effect immediately and appears in the audit trail with who made it. A vendor who answers the governance questions with adjectives has not answered them; a vendor who answers with a tab you can open has. The difference is the whole evaluation.