Buyer's guide
12 questions to ask before hiring an AI implementation partner
Every AI provider says they deliver results. These 12 questions, grouped by proof, ownership, operations, money and people, show you which ones actually will, and what a good answer sounds like.
Short answer
Before hiring an AI consultant or implementation partner, ask whether they will show a working prototype on your own data before you commit, which specific process and number will change, who owns what they build, where your data goes, who monitors the system after launch, how exceptions are handled, what the full first-year cost is, and what happens if it doesn't deliver. Clear, written answers to those questions matter more than the proposal itself.
Key takeaways
- Ask to see a prototype working on your real data before you commit any money.
- Insist on a single named process and a measured baseline, not an organization-wide assessment.
- Confirm in writing that you own everything built and that your data isn't used to train models for others.
- Find out who monitors the system after launch and how exceptions reach a person.
- Get one full first-year cost, and ask what happens if the result doesn't materialize.
Why the questions matter more than the proposal
Most AI proposals look alike: a discovery phase, a roadmap, a build and a handover. The differences that decide whether you get value show up only when you ask specific questions about proof, ownership, data, operations and money. A good partner answers them plainly and in writing. A weak one changes the subject.
The 12 questions below are grouped into five areas. Ask all of them, take notes on the answers, and compare providers side by side using the scorecard at the end. They work whether you are talking to a large consultancy, a boutique agency or a freelancer.
Questions about proof
The first job of these questions is to separate people who build working systems from people who sell strategy decks. Ask to see the thing working.
1. Will you show me a working prototype on my own data before I commit?
A prototype that processes your actual invoices, emails or job records is the fastest test of competence. Generic demos on clean sample data prove very little, because real business data is messy. Good answer: "Yes. Give us 20 to 50 real examples and we'll show you what it gets right and what it flags." Red flag: a paid multi-week discovery phase before anything is built.
2. Which specific process will you start with, and which number will change?
You want a named process and a named metric: hours per week on draw requests, response time to new leads, dollars of duplicate payments recovered. Good answer: a single process, a baseline you'll measure together, and a target. Red flag: "We'll start with an AI readiness assessment across the organization."
3. Can I see something you've built running in production?
Ask to see a live system, even with sensitive details hidden, along with its logs, its exception queue and its monitoring. Slides and recorded videos are not the same thing. Good answer: a screen-share of a real, running workflow and an explanation of what goes wrong with it. Red flag: only case-study PDFs and vague claims about results.
Questions about ownership and data
These questions decide whether you end up with an asset or a dependency. Get the answers in the contract, not just in conversation.
4. Who owns what you build: the workflows, prompts, code and documentation?
If the provider keeps ownership, you are renting your own process. Good answer: "You own all of it, and it runs in your accounts." Red flag: a proprietary platform where your workflows can't be exported or understood without them.
5. Where does my data go, and is it used to train models for anyone else?
Ask where data is processed and stored, which AI model providers are involved, what their data terms are, and who at the provider can see your information. Good answer: specific, written answers, and a clear statement that your data isn't pooled, resold or used to train models for others. Red flag: "It's all secure," with no details.
6. What happens if we stop working together?
Ask what you keep, what stops working, how handover works and how access is revoked. Good answer: documented systems that keep running in your accounts, with a defined handover. Red flag: everything stops on the day the contract ends.
Questions about operations after launch
Building an automation is the easy part. Keeping it working for years, as your vendors, software and processes change, is where most small business AI projects quietly fail.
7. Who monitors the system after it goes live, and how will I know if it breaks?
Logins expire, vendors change their invoice layouts, and software updates break integrations. Good answer: named responsibility for monitoring, alerts on failures, and a regular report you can read. Red flag: monitoring sold as an optional extra, or no answer at all.
8. How are exceptions and mistakes handled?
Every real process has edge cases. Ask what happens when the AI is unsure, how exceptions reach a person, and which actions need approval. Good answer: a clear human-in-the-loop design where anything uncertain, or anything touching money or customers, goes to a named person. Red flag: "The AI handles everything automatically."
9. What systems will it connect to, and what access do you need?
Ask for a list of every system the automation will read from or write to, and the permissions required for each. Good answer: dedicated accounts with the minimum access needed, read-only wherever possible. Red flag: a request for your admin password or a shared login.
Questions about money
Price matters, but the structure of the price matters more, because it decides who carries the risk if the project doesn't deliver.
10. What is the total cost for the first year, including everything?
Ask for build or setup fees, platform and model usage costs, monitoring and maintenance fees, change requests, and an estimate of your team's time. Our guide to what AI implementation really costs a small business lists the hidden costs to ask about. Good answer: one number you can hold them to. Red flag: a low build price with everything else "billed as needed."
11. What happens if it doesn't deliver the result we agreed on?
This is the most revealing question on the list. Good answer: fees tied to the measured result, or a clear remedy if targets aren't met. Outcome-based models are explained in how shared-savings AI pricing works. Red flag: "Every project is different," followed by an hourly rate.
A question about your people
12. How much of my team's time will this take, and from whom?
Small teams don't have spare hours. Ask who needs to be involved, for how long, and at what stages: process walkthroughs, collecting examples, testing and reviewing exceptions. Good answer: a realistic estimate by person and by week, front-loaded in the first month. Red flag: "Hardly any," from a provider who hasn't seen your process yet.
Scorecard: good answers vs. red flags
Use this table to compare providers after your calls. A provider with more than two or three red flags is usually not the right fit for a small business, however impressive the pitch.
| Area | Good sign | Red flag |
|---|---|---|
| Proof | Prototype on your data before you commit | Paid discovery before anything is built |
| Scope | One process, one measured number | Organization-wide "AI transformation" |
| Ownership | You own workflows, prompts, code and docs | Proprietary platform, no export |
| Data | Written terms; no pooling, reselling or training for others | "It's secure" with no detail |
| Operations | Named monitoring, alerts, exception routing | Handover and goodbye |
| Access | Least-privilege, dedicated accounts | Asks for admin passwords |
| Money | Full first-year cost; fees tied to results | Open-ended hourly billing |
| Your team | Realistic time estimate by person | "It'll take no time at all" |
What to prepare before you talk to anyone
You will get far better answers if you arrive with specifics. Spend an hour on this before your first call.
- Pick the one or two processes that cost you the most time or money.
- Estimate hours per week and who does the work.
- List the software involved: accounting, CRM, project management, payroll, email.
- Gather 10 to 20 real examples, including a few messy ones.
- Write down the three most common exceptions and how your team handles them.
- Decide which actions must always require your approval.
If you're still deciding which process to start with, our practical guide to AI implementation for small businesses has a simple scoring method. If the vocabulary is new, start with what an AI agent is and how it differs from a chatbot. Tarvora is happy to be asked every question on this list.
Frequently asked questions
What should I ask an AI consultant before hiring them?
Ask whether they will show a working prototype on your own data before you commit, which process and metric they will improve, who owns what they build, where your data goes and whether it trains other models, who monitors the system after launch, how exceptions are handled, what the full first-year cost is, and what happens if the project does not deliver.
What are red flags when hiring an AI agency?
Common red flags include a paid discovery phase before anything is built, an organization-wide scope instead of one measurable process, proprietary platforms you cannot export from, vague answers about data security, no monitoring after launch, requests for admin passwords, and open-ended hourly billing with no link to results.
Should I own the AI workflows a consultant builds for me?
Yes. You should own the workflows, prompts, code and documentation, ideally running in your own accounts. If the provider keeps ownership or runs everything on a proprietary platform, you are renting your own process and may lose it if the relationship ends.
How do I compare AI implementation providers?
Ask every provider the same questions about proof, ownership, data, operations, money and your team's time, write down the answers, and score them side by side. Weight heavily toward providers who will show a working prototype on your data and tie fees to measured results.
Is a paid AI discovery or strategy phase worth it for a small business?
For most small businesses, a long paid discovery phase is less useful than a working prototype on one real process. Discovery that ends in a roadmap leaves you with a document; a prototype on your own data shows whether the provider can actually deliver before you spend significant money.