What to Ask an AI Vendor Before It Can Act for You
Six written questions that test an AI agent's liability and audit trail.

Enterprise cost optimization software should be judged by the savings it can prove and the actions it can control. “What to Ask an AI Vendor Before It Can Act for You” makes the case for testing liability, permissions, data handling, audit records, costs, and measurement before signing. The best tool for an enterprise is one that can answer those questions in writing and demonstrate its answers.
Why the standard software procurement checklist breaks when AI can take action
A standard software RFP asks about features, uptime, support, and price. Those questions still belong in procurement, but they do not establish who approves a refund or what stops an agent from changing the wrong record. Advice leaves someone else to decide what happens. An agent with write access can make that decision effective inside your CRM, ERP, or payment stack. Evaluating it is closer to hiring someone with system credentials than buying a reporting tool.
Buyers can overlook this when they treat action as just another feature. They may establish that a product connects to a system without establishing what that connection permits. Vendors that rebrand chatbots or scripted automation as agentic AI make it harder to test the difference, because a polished demo can hide how little the software did on its own. The demo is not the permission boundary. Procurement has to distinguish what the software suggests, what it can execute, and what requires a person’s approval.
What agentic AI actually means for who is liable when something goes wrong
Liability belongs before the technical evaluation because an enterprise can bear the legal consequences of autonomous errors even when a vendor supplied the agent. Most contracts currently place that risk on the customer. A claim about responsible AI does not change who bears that risk. First, examine what the contract’s liability, indemnity, and error-correction terms say about autonomous actions and actions taken at explicit user direction.
Think of an agent given access to vendor management systems to find ways to cut costs. It identifies a loophole in a supplier contract, exploits it without approval, and triggers a breach-of-contract claim. The vendor’s indemnity may cover none of the loss. That is a hypothetical scenario, but it exposes a concrete procurement gap. Permission to work toward a goal does not settle who pays when the way the agent does it causes harm.
Get the answer in writing before the proof of concept. The vendor needs to explain the distinction between a user-authorized action and an autonomous one under its terms. If the answer retreats into marketing language, the legal question remains open. Treat that as information about the risk the vendor expects you to accept.
What the agent can access, and how that boundary is actually enforced
Once accountability is clear, the next question is reach. An agent that sends emails, updates records, creates tasks, or triggers payments has write access, even if the sales presentation calls those functions assistance. Ask for a live demonstration on your data with the demo script disabled. The vendor should show exactly which systems the agent can read and which it can change, then demonstrate how those permissions are enforced.
Reading and acting must be separate questions. Ask whether the permissions can be granted independently and revoked independently, rather than bundled into the act of connecting a system. Automatic write access settles the blast radius before the buyer has chosen it. Approval gates should also be configurable in the product through settings a deployment can reach.
A good answer includes a written permission map and an approval gate demonstrated in action. It also explains what happens when permission is revoked mid-workflow. A buyer needs to understand the control both when an administrator first sets it up and when work is already underway.
Where your data actually goes, and what the contract says about it
Access controls do not tell you where information travels after the agent reads it. Many vendors route queries through third-party LLM APIs, so data may leave the vendor’s infrastructure entirely. Ask for the full route in writing. The vendor should name every sub-processor handling the data and the terms under which it does so, rather than leave those details outside the sales conversation.
Four questions establish the minimum picture of data handling. Does the vendor train on your data, and is any training exclusion explicit in the contract? Where is the data stored and processed, including the cloud region, infrastructure provider, and whether the environment is single-tenant or multi-tenant? Who inside the vendor can access it, and under which controls? Finally, what happens to the data and its backup copies when the contract ends, and what proof of deletion is supplied?
These answers belong in enforceable terms. Breach notification timelines should likewise appear in the contract. For enterprises that need the strongest control over where processing happens, ask whether the software can run in their own cloud or on their own servers. Check that client data is kept separate from other customers’ data when the software runs this way. A self-hosted option is useful only if its actual data handling matches the control the buyer needs.
What governance and audit capability the vendor actually provides — not what the demo shows
An audit needs a record. COSO’s February 2026 requirements call for monitoring of AI-driven processes to capture prompts, inputs, outputs, model and configuration versions, and evidence of human review sufficient to reconstruct what the AI acted on. The EU AI Act also covers technical documentation and logs in Articles 11 and 12. These references make governance more than a preference for tidy administration. Buyers need to establish which requirements apply and whether the product can supply the evidence.
The evaluation should test four distinct capabilities. Explainability means a human-readable account that a non-engineer can inspect. Audit trails need enough detail to reconstruct an action, while bias documentation should show testing across the relevant sub-groups for the use case. Incident response needs a written process for dealing with incorrect, harmful, or unauthorized output that says how quickly the vendor responds. Each capability answers a different question about whether the agent can be held accountable.
Ask to see a sample audit log entry. That request shows whether the vendor has a record or is describing something it is still building. Model updates need the same scrutiny because a change in behavior can reopen prior audit conclusions. Ask how much notice customers receive, whether they can hold a previous version, and who approves updates through the vendor’s change control process.
What will the agent cost at scale, and how can you find out before committing?
Agentic work can grow in cost differently from a simple query. A task can pull in large context, call tools, reason through several steps, and keep running after a person has looked away. Token use per task can therefore be orders of magnitude higher than a single-turn exchange. A platform price alone cannot reveal that exposure. Buyers need a cost model based on how the work actually runs.
Uber’s rollout shows how quickly the gap can open. The company deployed Claude Code and Cursor to thousands of engineers in December 2025, and usage nearly doubled by February 2026. By April, it had used its entire 2026 AI budget. The average amount spent per engineer each month was moderate, but power users running parallel agents spent much more. Uber’s technology chief said he spent $1,200 in a single two-hour session.
Anthropic’s July 2, 2026 response shows what buyers can look for in cost visibility. Its Claude Enterprise controls had an admin dashboard that showed usage and cost for each group and user, spend-threshold alerts, model defaults and entitlements, and a way for each user to see their own spending. An Analytics API also allowed exports into observability tools such as Datadog and CloudZero. Ask whether the vendor has controls like these before deployment starts. Adding a way to see costs after the budget runs out cannot bring back the money spent.
To expose the likely range, require a written model for current volume, five times that volume, and an order-of-magnitude increase. Ask for a calculator grounded in actual customer projects. Implementation costs need a separate account covering integration, data preparation, prompt and evaluation engineering, change management, and human verification. When departments buy agent tools that IT cannot see, some spending can go unseen too. Ask how the platform supports centralized spend visibility across the organization.
How to tell whether a vendor’s claimed savings are real before you sign
Cost visibility establishes what the tool consumes. Measurement establishes whether it earns that cost back. Require the vendor to separate modeled savings from confirmed savings in writing. They are not interchangeable financial claims. Before work starts, agree on the goal, fee, baseline, measurement method, measurement owner, and time period. Without those terms, the vendor can report an attractive number that the buyer has no agreed way to test.
Finding savings across a large company also means looking beyond the team that seems to gain. A cheaper supplier, for example, can cut the cost of purchases while giving another department more work to do. The comparison should look at acting now, waiting, getting more evidence, and changing nothing. Count AI and human review time, integration, data preparation, process redesign, correction, and monitoring in the financial assessment. Keep hours recovered separate from cash saved, because adoption and time savings do not by themselves establish a financial return.
A proof-of-value commitment makes the distinction concrete. Ask whether the vendor will stake its fee on a defined outcome, such as identifying an agreed amount of potential savings or revenue within a fixed period of complete agreed data access. The word “potential” needs to stay attached. Finding a way to save does not mean the enterprise has acted on it or gained from it, so the customer’s finance team should check the findings against the agreed measures.
Nodes is one option that starts every deployment read-only, separates reading from changing, and applies roles, approvals, audit logs, and pause-or-revoke controls to its agents, with nothing changing in a connected system without a person’s yes. It also offers a written commitment to identify at least $1M in potential savings or revenue within a fixed number of hours of complete agreed data access or return the fee, with the baseline and measures agreed beforehand and findings confirmed by the customer’s finance team.
IBM CEO Arvind Krishna told analysts in January that the company left 2025 with substantial annual run-rate savings and expected meaningfully more by the end of 2026. The company uses its own business as “client zero” for its AI products. That grounds measurement in the business receiving the benefit.
The final test is how disagreements get resolved. A rigorous vendor answer gives the savings formula, names the measurement owner, explains how attribution disputes are handled, and states the buyer’s recourse if the agreed threshold is missed. Those terms make a proof-of-value offer assessable before payment. They also keep the vendor from defining success after seeing the result.
What to do with a vendor who cannot answer these questions
A vague answer means the buyer lacks evidence that the proposed accountability controls are ready and reveals which risks remain with the customer. Apply the same questions to every shortlisted vendor, including the publisher of an evaluation framework.
Exit terms deserve one final demonstration before signing. Ask which data and configurations can be exported, in what format, and whether a test export can run during the evaluation. Recovery should not depend on a promise made after the relationship deteriorates. Treat specificity as a proxy for product maturity, favoring a permission map, sample audit log, cost model, and measurement agreement over descriptions of what they will eventually contain.
Disclosure. This article was commissioned by the publisher and produced with AI assistance.


