How to choose an AI consultant: 7 questions to ask before signing

May 16, 20268 minAI, Consulting, SaaS, Hiring

Short answer (60 seconds): the 7 critical questions before signing with an AI consultant or agency: (1) can you show me code of yours that's in production?, (2) how do you handle scope creep?, (3) what happens if the underlying model deprecates mid-project?, (4) who, specifically, will be on my project?, (5) what's your handoff plan?, (6) how often and in what format do you communicate progress?, (7) what guarantees exist if the outcome isn't what was expected? Without clear answers to these 7, don't sign.

This isn't a guide on "evaluating AI consulting" in the abstract. It's the list of questions I would ask — and recommend asking — before committing USD 10,000-50,000 to a project.

It's based on patterns I see from the client side (quotes that arrive at LATAM SaaS founders and CTOs) and from the provider side (what other consultants I talk to get asked). If the consultant in front of you can't answer these 7 clearly, that's a signal, not a problem with the questions.

Question 1 · "Can you show me code of yours that's in production?"

This separates faster than any other question those who sold AI consulting in the last 12 months from those who delivered it.

What to ask specifically:

Public repository or real snippets (with sensitive data redacted).
Ideally something deployed and working — link to a product, not a localhost demo.
An explanation of the technical decisions behind it: why this model, why this architecture, what they tried before.

What does NOT count as an answer:

"Everything's under NDA, I can't show anything." Reasonable up to a point, but a good consultant has something public (open source, personal demos, technical posts) that demonstrates real capability.
"I'll show you a slide with client logos." Tells you nothing about work quality. Agencies collect logos; that doesn't translate to your project being well-implemented.

Positive signal: a consultant who writes technically on their blog, maintains open-source projects, or speaks at technical conferences usually has demonstrable level. What they publish for free is the minimum quality of the work they charge for.

Question 2 · "How do you handle scope creep?"

This separates professional proposals from aspirational ones.

What to look for in the answer:

A specific policy: "I include in the price any change that doesn't add more than X% to the original scope. Above that, we quote an addendum in writing."
A process for triggering it: "when I detect a request adds scope, I tell you before doing the extra work."
A price for additional scope: explicit USD/hour or USD/day.

Red flags:

"We'll discuss it when it happens." You'll pay more than double the original budget.
"Don't worry, I charge the same even if scope grows." Sounds good until the consultant starts cutting quality to maintain price.
Pure hourly with no ceiling. Scope-creep risk lands 100% on you.

Question 3 · "What happens if the underlying model deprecates mid-project?"

This is the question that separates consultants who have shipped AI to production from those who only did demos. OpenAI and Anthropic deprecate models every 6-18 months. If your project finishes just as your main model is deprecated, someone has to migrate it.

Good answer:

"I design with provider abstraction from the start (swappable LLM interface)."
"I have fallback to the alternative provider configured (Claude ↔ OpenAI)."
"If deprecation falls within the post-delivery guarantee, I migrate it. If after, you have it documented or I quote it separately."

Bad answer:

"That shouldn't happen."
"I don't know, OpenAI doesn't announce."
"That's not in my scope."

If they never thought about this, they'll learn it on your dime.

Question 4 · "Who, specifically, will be on my project?"

In agencies it's critical; with independent consultants it confirms you'll work with the same person who sold to you.

At agencies:

Ask for specific names of the team: senior tech, project manager, devs.
Confirm how much time each dedicates: full-time, part-time, "as needed" (red flag).
Ask what happens if the main person leaves mid-project.

With independent consultants:

"Will you be on this project?" Yes or no.
If they have assistants/extended team, how much of the work is them vs you?
If they subcontract parts, who's accountable for quality and for communicating with you?

Classic red flag: the person signing the contract is senior; the person implementing is junior. You pay for experience you don't receive.

Question 5 · "What's your handoff plan?"

What makes a consulting project valuable isn't only the deliverable — it's that your team can maintain it afterwards.

What to look for:

Technical code documentation (not marketing): how to add features, how to debug, how to deploy.
Pairing sessions with your team during the last 1-2 weeks.
A defined async support period (30-90 days post-delivery).
Access to resources: code lives in YOUR repo, services in YOUR accounts (OpenAI, Supabase, Helicone), keys in YOUR vault.

Red flag:

"I send you a ZIP at the end."
Services live in the consultant's account "for easier management."
"Team training: not included."

Without real handoff, you'll depend on the consultant forever — and that's a business model for them, not for you.

Question 6 · "How often and in what format do you communicate progress?"

AI projects have a lot of technical uncertainty. Without clear communication, you'll reach week 4 not knowing if the project is on track.

Good:

Written weekly update with: what's done, what's coming, risks/blockers, costs to date if applicable.
Sync call every 1-2 weeks, not more often (more interrupts than helps).
Access to Slack or shared channel for async questions.

Bad:

"I'll let you know when there's news." Important news comes in the form "it doesn't work" and "I already spent your budget."
Sync calls daily or every 3 days. If that level of check-in is needed, the consultant doesn't know what they're doing or you don't trust them (both are problems).
"I'll send a final report when we're done." You'll hear about scope creep on day 60, when it's already too late.

Question 7 · "What guarantees exist if the outcome isn't what was expected?"

This is the question where you see who has real confidence in their work.

Types of honest guarantee:

Measurable uptime/performance guarantee: "The system responds in under Xms 99% of the time, or I work for free until we get there."
Free support for X days: "Any bug detected in the first 30 days I fix without extra charge."
Partial refund guarantee: "If the MVP doesn't hit the agreed criteria, I refund 30% of the project."

What is NOT a guarantee:

"I guarantee quality." Not actionable.
"I work until you're happy." Indefinite, dangerous for both sides.
"There's no guarantee, it's like any software project." Reasonable sometimes, but a good consultant offers something concrete.

Honestly: guaranteeing success in AI is hard because there are variables outside the consultant's control (model changes, client data quality, etc.). What should be guaranteed: that the work is documented, maintainable, and that technical decisions are justified.

How to apply these 7 questions in a conversation

You don't have to fire all 7 like an interrogation. Most come out naturally if you steer the conversation this way:

Ask for context: "Tell me about a recent project similar to mine and how it turned out." (Questions 1, 4)
Ask about process: "If we sign, how do we start? What are the milestones?" (Questions 5, 6)
Ask for honesty about risks: "What could go wrong in this project, and how would you handle it?" (Questions 2, 3, 7)

If in these three conversations the consultant answers with specifics, numbers, examples, and acknowledges risks without defensiveness, you're looking at a good candidate. If they answer with generalities, "it depends", or get uncomfortable with the questions — that's the information you needed.

One question that stands alone

If you can only ask one question:

"Tell me about a recent project where something went wrong and how you handled it."

The answer tells you more than 10 structured questions. A mature consultant has failure stories, tells them in detail, explains what they learned. A consultant with no failures to tell, or who has them but hides them, has no real production experience.

Let's talk about your case

If you're evaluating several AI consulting proposals and want a second pair of eyes on the quote you received, book a 30-minute call at no cost. 20 minutes is usually enough to detect red flags and validate whether the price aligns with the scope. If I'm not the right person for your case, I'll point you to someone who is.

Read also:

How much does an AI consultant charge in LATAM — the market ranges to benchmark against.
How much does it cost to implement AI in a SaaS startup — costs by project type.
More articles on consulting — guides and comparisons.
Back to the blog — all articles.

Frequently asked questions

What's the most common red flag in AI consulting proposals?

'Strategy' slides without a concrete technical deliverable. A proposal with no code, no automated process, no measurable metrics is paid networking. Real consulting delivers something that works in production. If the proposal has 80 slides and zero lines of code, ask for more detail before signing.

Which previous experience matters most?

Systems that actually shipped to production, not demos. Ask: 'can you show me a previous case, even anonymized, with before/after metrics?' A consultant who took AI to production knows the non-obvious problems (rate limits, surprise costs, model deprecation). A consultant who only did demos doesn't.

How long should the first contact last?

30 minutes max. If in 30 min a consultant can't understand your case and give you a scope/price idea, they'll take two months to deliver a proposal. The initial conversation should feel useful even if you don't hire them — the consultant understands enough to say 'your case needs X' or 'your case doesn't need AI yet'.

What does a good proposal include?

Five minimum: (1) specific scope — what's in, what's out; (2) enumerated deliverables — code, docs, training; (3) timeline with milestones; (4) price with currency, payment terms, what costs extra if scope grows; (5) what post-delivery guarantees exist. If any is missing, ask. If the consultant doesn't provide it, bad sign.

Specialist or generalist?

For bounded use cases (a RAG, an agent, an integration) — specialist in that. For broad diagnosis where you don't know what you need — generalist with experience across stacks. A clue: if the consultant recommends exactly the solution they dominate before understanding your case, they'll force-fit the tool. Bad sign.

How do I evaluate if prices are reasonable?

Compare against market ranges (LATAM independent consultant 2026: USD 80-200/h or USD 1.5K-30K per fixed project). If they quote 3x market, ask for justification; if they quote 1/3 of market, you'll likely pay the savings in rework. More detail in the [post on AI consultant rates in LATAM](/blog/cuanto-cobra-consultor-ia-latam-2026).

What do I do if an agency quotes me but I want an independent consultant?

Ask the agency to tell you who, specifically, will be on your project. If the initial contact is a senior partner but the implementer is a junior, you'll pay for the senior and work with the junior. With an independent consultant, the person on the project is the only person on the project.