Almost every software company now offers AI. Demos look similar, proposals use the same vocabulary and price ranges vary widely. The difference between a project that delivers and one that stalls usually shows up in the answers to a few direct questions. Use the twelve below in your next partner conversation. They are grouped into business fit, data and integration, security and control, and delivery and ownership.
A good demo shows what is possible. Good answers show what will actually work in your business.
To apply this workflow in your business, explore our AI enablement & readiness consulting service and implementation approach.
Business fit: questions 1 to 3
1. Which workflow would you start with, and why? A strong partner narrows the conversation to one process with a clear owner, available data and a result you can measure. Be wary of a plan that starts with a platform rather than a problem.
2. How will we know it is working? Expect specific measures such as exceptions resolved per day, time to answer a customer or report preparation hours, plus a baseline taken before the project starts.
3. Where should this not use AI? A trustworthy answer identifies steps that are better handled by ordinary rules, existing software or a person. Partners who apply AI to everything tend to create fragile systems.
Data and integration: questions 4 to 6
4. Which of our systems will you connect to, and how? Ask for the specific interfaces they expect to use with your ERP, CRM, helpdesk or production systems, and what happens if an interface is not available in your version.
5. What do you need from our team, and when? Good partners are explicit about subject-matter experts, test data, access approvals and decision makers. Vague answers here are the most common cause of delay.
6. How will you handle poor or inconsistent data? Listen for a plan to measure data quality early, agree definitions and show uncertainty to users rather than hiding it.
Security and control: questions 7 to 9
7. How do you protect against prompt injection and data leakage? The OWASP Top 10 for Large Language Model Applications lists prompt injection and sensitive information disclosure among the leading risks. A partner should be able to explain their mitigations in plain language.
8. What can the AI do on its own, and what needs approval? OWASP also highlights excessive agency: systems that can take more actions than they should. Ask to see the permission model and where people approve payments, record changes or customer commitments.
9. How will access follow our existing roles? Users should never be able to retrieve, through an assistant, information they could not see in the underlying system.
References: OWASP Top 10 for LLM Applications 2025
Delivery and ownership: questions 10 to 12
10. How will you test it before go-live? Look for an evaluation set built from real, representative cases, including questions that should be refused and actions that should be blocked, reviewed with your business owners.
11. Who owns the code, prompts and evaluation data? Agree this in the contract. You should be able to run, change or move the solution without starting again.
12. What happens after launch? Ask who monitors quality, how model or vendor changes are handled, how incidents are escalated and what support is included. A clear answer is worth more than a low initial quote.
Red flags and better answers
Use the table as a quick reference while you compare proposals.
| Red flag | A better answer sounds like |
|---|---|
| We can automate your whole department | We will start with one workflow, prove it and expand |
| Accuracy is guaranteed | Here is how we will measure accuracy on your own cases |
| The AI decides, so you do not have to | These actions need your approval, and here is the audit trail |
| Our platform connects to everything | These are the interfaces we will confirm in discovery |
| Ownership can be discussed later | You own the code, prompts and evaluation data from day one |
Should you start with a proof of concept or a pilot?
A proof of concept answers whether something is technically possible, often with sample data. A pilot answers whether it works in your business, with real users, real records and the controls you need. For most organisations a short discovery phase followed by a bounded pilot gives better information than a polished proof of concept that never meets production data.
The NIST AI Risk Management Framework, a voluntary framework widely used as a reference, organises this thinking into four functions: govern, map, measure and manage. Asking a partner how their approach covers each one is a useful final test.
References: NIST AI Risk Management Framework
Questions readers ask
What should an AI discovery phase include?
A shortlist of candidate workflows, an assessment of the data and systems involved, a recommended first scope with success measures, the controls required and a realistic plan for a pilot.
How do we compare quotes from different AI partners?
Compare the scope of the first workflow, what is included for integration, testing and support, who owns the deliverables, and the assumptions about your team's time. Two similar prices can describe very different amounts of work.
Do we need our own AI team first?
No, but you do need a business owner for the workflow and someone who can make decisions about data access. A good partner helps you build internal capability as the project progresses.
A practical starting checklist
- One workflow with a measurable outcome
- Named interfaces and team responsibilities
- A clear permission and approval model
- Ownership of code, prompts and evaluation data
Choose the partner whose answers are specific about your workflow, your data and your controls. That is the best predictor of a project that reaches production.
Original guide by the IOT Software and Solutions editorial team. External facts are linked to their sources; examples are illustrative and do not describe measured client results.

