Evaluating vendors

How to Choose an AI Consulting Firm: 7 Criteria and the Questions to Ask Before You Sign

By Juanma (Juan Manuel Ortiz de Zárate), co-founder and CTO of Data Voices. Master's and PhD in Computer Science (UBA). · 11 min read

Choosing an AI consulting firm comes down to checking four things: cases in production, not demos; a method that goes from assessment to implementation; clear ownership of the code and the data; and how the result is measured. If a firm can't show you all four with evidence, keep looking.

This guide covers how to choose an AI consulting firm when you already have two or three proposals on the table, or are about to ask for them. You'll find seven criteria, each with the concrete proof worth asking for.

You'll also find ten questions you can paste into an email before you sign, and what an initial assessment should deliver. It all starts before the first meeting, with the problem you want to solve.

Before you search: what problem do you want to solve?

Bring a concrete, measurable process to the first meeting, because without one no firm can propose something you can compare. That holds for a small business and for a large company alike.

A concrete process fits in one line: who does it today, with what data and how long it takes. “Two people answer WhatsApp messages by hand and enter each order into the system” is a starting point. “We want to use AI” is still an intention.

With a defined process, proposals can be compared side by side, and the assessment can focus on feasibility from day one. If it isn't clear yet, look for the place where a person acts as a bridge between a channel and a system: that's often where the first use case is.

The 7 criteria to evaluate an AI consulting firm

Evaluate an AI consulting firm on seven criteria; each one can be checked with proof you can request in writing. The four above (production, method, ownership and measurement) are the baseline; the other three weigh more or less depending on your project.

Criterion	Proof to ask for	Red flag
1. Cases in production	A system running today, with sector, problem and current status	Only demos, pilots or screenshots
2. Method	Written stages, from assessment to implementation and follow-up	Starts with the tool before understanding the process
3. In-house technical team	Names and profiles of who builds and who maintains	Everything is subcontracted, or nobody owns maintenance
4. Ownership of code, data and accounts	A clause stating who owns each item at the end	Vague answers, or accounts in the vendor's name
5. Integration with what you already have	Which data and systems it connects to (databases, documents, WhatsApp, CRM)	Requires migrating everything before starting
6. Measurement	The business metric that will move and how it's reported	“You'll see the impact”, with no number to watch
7. Security and sensitive data	Permissions, usage logs and what the system does when it doesn't know	Nobody asks which data is sensitive

1. Cases in production

A case in production is a system that real users operate today, with a result someone can describe. Ask for the status of each case: in use, technically validated, pilot or archived.

2. Method

The method is the sequence that takes a project from assessment to a running system, with deliverables at each stage. A solid sequence first maps processes, data and permissions, then prepares the team that will operate the system, and only then implements.

Data Voices' AI implementation method, for example, publishes its stages: Maturity, Upskilling, Implementation, “Report to whoever's paying” and Follow-up. If a proposal jumps straight to the tool, ask who checked whether your team will be able to use it.

3. In-house technical team

The technical team defines who builds the system and, above all, who maintains it once it's running. Ask for names and profiles, and ask which part is subcontracted: models, data and processes change after delivery.

4. Ownership of code, data and accounts

At the end of the project, the code, the data and the accounts need an owner defined in the contract. That includes the repository, the knowledge base, the usage logs and the keys for the model providers.

If the accounts stay in the firm's name, changing vendors means starting over.

5. Integration with what you already have

A good firm works on the data and systems your company already runs: your own databases, documents, WhatsApp or the CRM. Requiring a full migration as a condition to start turns one project into two.

6. Measurement

Every project should name, from the start, the business metric it will move and how it will be reported. It can be response time, hours of manual data entry, requests resolved without intervention or adoption of the system by the team.

7. Security and handling of sensitive data

The firm should explain what data the system sees, who accesses what and what gets logged. Ask to see how permissions are configured, whether answers cite their source and what the system does when it doesn't have the information.

The questions to ask before you sign

Before you sign, send these ten questions in writing and compare each firm's answers side by side. They're the practical version of what to ask before hiring an AI consultant.

Can you show me a system that's running in production today? What status is it in?
What stages does the project have, and what do I receive at the end of each one?
Who builds the system, and who maintains it after delivery?
Who owns the code at the end? And the data, the repository and the model provider accounts?
Which of our systems does it connect to, and what access do you need?
Which business metric will we watch, and how will you report it?
What happens if the model gives a wrong answer? Who detects it, and how is it fixed?
In which cases does the system hand the conversation to someone on my team?
Which model or platform do you propose, and why that one? Can we change it later?
What happens to the system if we stop working together?

For question 7, ask for concrete mechanisms: human review before an answer is published, a log of every execution or an explicit criterion for the system to say “I don't know”.

Red flags

Six signals justify discarding a proposal or, at the very least, demanding a clear explanation before you go further.

It promises full automation or replacing teams. Systems that work in production define when a person steps in.
It guarantees a return on investment before knowing your processes and your data. Without an assessment, that figure is a sales estimate.
It only shows demos. A demo proves the idea is possible; a case in production proves someone sustains it over time.
It sets a price without having seen your data. The scope depends on what data exists, what state it's in and which systems it has to integrate with.
It depends on a single platform and doesn't explain why. Choosing it before the assessment ties the project to it.
It presents cases with no name or verifiable detail. Without a sector, a system status and one concrete data point, a case looks a lot like a promise.

AI consultancy or implementation vendor?

An AI consultancy assesses, designs and implements with a business lens, while an implementation vendor builds the pieces or tools the client has already defined.

They play different roles. If you already know what you want to build and have someone to specify it, an implementation vendor may be enough. If you still have to decide which process to tackle, with what data and how to measure it, you need someone to do that work before building.

Nearshore (Latin America) or global firm?

A nearshore firm in Latin America gives you shared working hours and familiarity with the channels Latin American customers use; a global firm gives you scale and presence in several countries.

For a mid-sized company, day-to-day logistics matter most: meetings in your working hours and a team that knows the channels your customers use, from WhatsApp to Mercado Libre. The legal framework matters too: the system has to comply with the data protection law that applies to you, such as Ley 25.326 in Argentina.

A global firm makes sense when the project spans several countries under a single standard. In both cases, the same seven criteria apply.

What a good initial AI assessment should include

A good initial assessment ends with five deliverables: the process map, the available data, opportunities prioritized by impact and effort, the risks and a concrete next step.

Process map: who does what today and where it gets stuck.
Available data: what information exists, in what format and with what permissions.
Prioritized opportunities: what to tackle first, by impact and effort.
Risks: sensitive data, dependencies and team adoption.
Next step: a narrow scope, or the recommendation not to move forward yet.

That level of detail is built in stages. The first conversation works as a filter and defines whether there's a problem worth a project; the full mapping comes when the work starts.

That's how Data Voices approaches it: the free diagnostic is a meeting to see if there's a problem worth a project. When there is one, the Maturity stage maps “Which processes, which data, which permissions.”, and the “Report to whoever's paying” stage covers “Adoption and, when measurable, impact.”

Its published case studies show the kind of proof worth asking any firm for:

In the work with an international organization, the work included data auditing and inference pipelines covering on the order of a thousand organizational units and some fifteen years of projects.
The Real Time Interview Agent runs on Windows and macOS, with the main workflow in production, and processes a candidate's answer in about 6.45 seconds on Windows 11.
In Normsy, the replies the model proposes go through human collaborators before they're published; the study had 1,021 participants, and the project won the 2024 Anthem Award for Best Use of AI.
In Agent Launcher, every execution is persisted and audited, and the user can change model or provider without redoing the agents.

Frequently asked questions

How do I choose an AI consulting firm for my company?

Choose the firm that can show you cases running in production, a written method from assessment to implementation, a contract stating who owns the code, the data and the accounts, and the business metric that will measure the result. If any of that proof is missing, keep comparing before you sign anything.

What questions should I ask an AI consultant before hiring?

Ask whether they can show you a system running today, what stages the project has, who builds and who maintains it, who owns the code, data and accounts at the end, which metric will move, what happens if the model answers wrong and what remains if you stop working together. Get the answers in writing.

AI consultancy vs software development vendor: what's the difference?

An AI consultancy assesses the process, designs the solution and implements it with business judgment until it's running. A software development vendor builds the pieces or tools the client has already defined. If you still need to decide which process to tackle, with what data and how to measure it, that earlier work belongs to a consultancy.

Should I hire a local, nearshore or global AI consulting firm?

A local or nearshore firm simplifies the day-to-day: shared working hours, the same language and knowledge of the business and regulatory context. A global firm makes sense when the project spans several countries under a single standard. In every case, evaluate the firm with the same criteria: production cases, method, team, ownership and measurement.

How can I verify that an AI consultancy has real production cases?

Ask to see a system running today with real users, and ask what status it's in: in use, technically validated or pilot. A verifiable case has a sector, a problem, what was built and at least one concrete data point, such as response times or volume processed. If there are only screenshots or videos, it's a demo.

Who owns the code and data at the end of an AI project?

It should be written into the contract before you sign: who owns the code, the data, the knowledge base, the repository and the model provider accounts. The recommended setup is that everything stays in your company's name, or that you keep guaranteed access when the project ends, so changing vendors doesn't mean starting over.

What are common mistakes when hiring AI consulting services?

Four common mistakes: starting without a concrete process, choosing by the most impressive demo, signing without defining who owns the code and the data, and starting without an agreed business metric. All four can be prevented with an initial assessment and written questions sent to every firm before comparing the proposals side by side.

What should an initial AI assessment include?

A good initial assessment ends with five deliverables: a map of the current process, the available data and permissions, opportunities prioritized by impact and effort, the risks, including sensitive data and team adoption, and a concrete next step. That next step can be a narrow scope or a recommendation not to move forward yet.

How to decide between the proposals on the table

Go with the firm that answers the ten questions in writing and shows at least one case in production with verifiable detail. If two proposals tie, choose the one that best defines the metric and code ownership: those are the two points hardest to fix after signing.

The next step is to bring your concrete process to a first conversation and find out whether it's worth a project. Data Voices is an AI consultancy for businesses based in Buenos Aires, and its first commercial step is that conversation; meet the Data Voices team. If the first case runs on WhatsApp, how to choose a WhatsApp AI provider covers the criteria for that channel.

Get a free diagnostic

Have a specific process to review?

Bring it to a first conversation with the team.

Get a free diagnostic