Guide

What to ask before you hire an AI automation firm

20 min read

By . Published . Last updated . Reviewed by Mavrin Labs on . How this guide was researched and checked.

A card reading "What to ask before you hire an AI automation firm" beside the Mavrin Labs logo
The card that shows when this guide is shared or linked.

An AI automation firm should answer a short list of questions in writing before you sign. Ask what result it will measure, who does the work, what happens to your data, who approves what goes out, who owns the build and how you leave.

Promises of results, vague answers about data and a contract that says nothing about ownership are the warning signs.

The guide is for the owner or manager of a small business or nonprofit who is comparing firms that offer to automate work with AI. It gives the questions to ask, what a good answer sounds like and the signs that should end the talks. It is not legal advice, and it recommends no firm, tool or AI provider.

Sources last checked Monday, November 16, 2026, or on the later date a citation gives. Mavrin Labs reviews this guide every quarter, and each source's own page wins over it.

What is an AI automation firm?

An AI automation firm is a company that designs and builds software, often around an AI model, to take a repeat task off your staff and hands over a working system that your own staff then run, check and keep in use.

Ownership needs a contract: software is not among the commissioned works that can count as made for hire, so ask the contract to transfer the custom work to you (Circular 30, revised August 2024, checked Wednesday, November 18, 2026).

Some firms build and leave, some run the system for you afterwards, and some resell a tool they did not make. Knowing which kind you are talking to tells you which of the questions below matter most.

Some firms build an AI agent, software that decides each next step within limits you set, and the plain definition of an AI agent explains how one differs from a fixed workflow. Others call themselves consultants, which raises the question what does an AI consultant do, set out in the FAQ, and how advice differs from a build.

Why it matters: most small firms are buying AI for the first time

Most small businesses are new to buying AI, which makes the questions below easy to skip.

The Census Bureau's business trends and outlook survey put AI use at between 17 and 20 percent of U.S. businesses from mid-December 2025 to early May 2026, and under 20 percent among firms with four or fewer employees (the Census Bureau's story on AI use by firm size, checked Monday, November 16, 2026).

Handing a firm your data and passwords also carries a measurable risk. The FBI's 2024 internet crime report logged $16.6 billion in reported losses from 859,532 complaints, a 33% rise on 2023 (the FBI's 2024 internet crime report, checked Monday, November 16, 2026).

A firm that holds your access becomes part of that picture, so its answers on data belong in writing.

Why ask before you hire an AI consultant or automation firm?

Ask these questions before you sign, because the answers become the contract, and a contract is hard to change once work starts. Asked early, they fit into talks you are already having; asked after launch, the same questions can leave you without the system, the data inside it or the right to take both to another firm.

AI makes big claims easy to sell. On Wednesday, September 25, 2024, the Federal Trade Commission (FTC), the federal agency that acts against deceptive business claims, announced a sweep of cases over AI claims it called deceptive.

Several promised buyers steady income from tools sold as AI-powered. The FTC's point was that using AI gives no business an exception from the law (the FTC's announcement of the sweep, checked Wednesday, November 18, 2026).

How buyer guidance on AI suppliers developed

Buyer guidance on AI suppliers is recent, and it arrived in three steps. NIST released its AI risk management framework on Thursday, January 26, 2023, asking organizations to plan for software and data that come from outside (NIST's page on the framework, opened again the day this guide was last updated).

The FTC's sweep of September 2024 then showed that AI claims answer to the ordinary law. In October 2024, NIST's quick-start guide on supply chain risk turned supplier checks into steps a small buyer can follow (NIST SP 1305, checked Wednesday, November 18, 2026). The questions below draw on all three.

The sections below take the questions in order: the result, the people, your data, approval, testing, ownership and leaving.

At the first meetingThe resultThe peopleYour dataBefore the buildApprovalTestingIn the contractOwnershipLeaving
The seven question groups in the order a firm should answer them, from the first meeting to the exit. The filled groups are the ones a contract must settle.

What result will the firm agree to measure?

A good firm agrees to measure one named task against a baseline, a record of how long the task takes and how often it runs today, and to check the result on a date you both accept. It says what it will build and what it will not.

A firm that promises a figure before it has seen your work has nothing yet to base it on.

Ask the firm to write the scope down: the task, the systems it touches, the result you will measure and the review date. Then ask what happens if the result falls short on that date. A clear answer names a next step, such as a fix or a change of plan, rather than a new promise.

  • A good answer: names the task, the baseline and the review date, in writing.
  • A warning sign: hours or revenue promised before anyone has watched the work.

Who will do the work, and who else touches it?

Ask who will plan, build and support the system, and who you call when it stops. A good firm names the roles and says whether any of the work goes to a subcontractor, another company it hires. It also lists the outside services the system will run on, such as an AI model provider or a hosting company.

The National Institute of Standards and Technology (NIST), the federal standards agency, treats this as supply chain risk. Its quick-start guide on the subject, published in October 2024, lists checking critical suppliers before you buy as an outcome to aim for.

It suggests judging how critical a supplier is by how much the work matters, how sensitive the data is and how much access the supplier gets (NIST's quick-start guide to supply chain risk).

The same guide suggests asking suppliers for evidence of their security practices, such as a statement of those practices, a certification or an inspection. A firm with no certification can still answer well, if it says so plainly and puts its practices in the contract.

  • A good answer: named roles, one contact, and the subcontractors and outside services listed in writing.
  • A warning sign: work passed to companies the firm will not name, or no answer on which AI provider the system uses.

What evidence of past work can the firm show?

Past work is the first evidence to ask for, and the hardest to check. Ask to see work like yours, such as a design plan with private details removed or a demonstration on cases you supply. A firm may not be able to name clients without their permission, and that is a fair answer.

Be wary of praise you cannot check: in the same FTC sweep, one company sold a tool that wrote detailed customer reviews from very little input, and the FTC alleged that many were likely false. A sample of real work you can read proves more than a page of glowing quotes.

What will the system do with your data?

Ask the firm to list the data the system will touch, where that data will be kept and who can see it. A good answer also says whether your data trains any AI model, and gives the firm only the access the job needs.

The FTC's guidance on vendor security for small businesses, dated September 2025, says to put security terms in the contract, covering how the vendor may use, share, sell, keep and delete your data.

It also says to confirm that the vendor follows them, instead of relying on its promise (the FTC's cybersecurity guidance for small businesses).

The FTC's guide to lessons from its own security cases adds that a contractor should get no sensitive data it does not need (the FTC's guide to starting with security). CISA, the federal cybersecurity agency, suggests asking whether multi-factor sign-in, a second check beyond the password, is on by default (CISA's guide for software buyers).

  • A good answer: the data, where it is kept, who can see it and whether it trains a model, all in writing.
  • A warning sign: a request for full administrator access before there is a written plan, or no clear answer on training.

What happens if something goes wrong with your data?

Ask how soon the firm will tell you if your data is exposed, and what help it gives you afterwards. A good answer gives a notice window in writing and says who drafts any notices your customers need. State law may set its own deadline, and the firm's window should fit inside it.

Florida's law is one example: a company hired to keep, store or process personal information for another business is a third-party agent under it. A third-party agent must tell that business no later than ten days after it finds a breach or has reason to believe one happened.

It must also shred or erase customer records once they are no longer to be kept, so that nobody can read them (Florida's law on the security of personal information).

Other states have their own rules, so ask counsel which ones reach your business.

  • A good answer: a notice window in writing, and help with the notices you have to send.
  • A warning sign: a contract with no breach clause at all.

Ask for that window as a number of hours or days, not a vague word such as promptly. Under Florida's law the ten days is the outer limit for a third-party agent, so a firm's window can sit well inside it.

Ask the same of support and exit: a named list of what support covers, and the date the firm's access ends.

Who approves what the system sends?

A person at your business should approve what the system sends until you decide otherwise, in writing. Ask the firm which actions the system may take alone, which need a person and how you change that later. A good answer is a written list of allowed actions, with approval on anything that reaches a customer.

The Small Business Administration (SBA) gives the same advice for AI tools in general: have a person check every message and campaign the tools produce before it goes out. Its page calls these tools super intelligence, but the advice fits any system that writes for you (the SBA's page on AI for small businesses).

Ask, too, what the system records, since CISA's guide suggests asking whether security logs come with a product as standard. A log of what arrived, what the system did and who approved it lets you trace and correct a mistake.

  • A good answer: allowed actions listed in writing, approval on customer messages and a log you can read.
  • A warning sign: a promise that the system will run with nobody checking anything from the first day.

How will you know the system works before you rely on it?

You will know the system works when it passes acceptance tests, written cases that say what goes in and what result you will accept, drawn from your real work before the build starts. Ask the firm who writes the tests, when you sign off and what happens when a test fails.

A demonstration on the firm's own examples shows that a tool can work, not that it works on your cases. Ask for the hard cases too: the angry customer, the request you do not offer and the message that needs a person. The guide to testing an AI agent sets out what to test in detail.

  • A good answer: tests written from your cases before the build, and a sign-off step before handover.
  • A warning sign: a polished demonstration offered as the only proof.

Who owns what is built?

You should own what is built for you, and the contract should say so in writing. Ask who owns the custom code, settings and content, who owns your data, and what the firm keeps for itself. A good answer separates the custom work, which becomes yours, from the firm's general tools, which you may use under a license.

Ownership does not follow payment on its own. The U.S. Copyright Office explains that a commissioned work counts as a work made for hire, owned by the party that ordered it, only in a short list of categories and only with a written agreement signed by both sides.

Software is not named on that list (the Copyright Office's circular on works made for hire, revised August 2024).

So ask for the contract to transfer ownership of the custom work to you, and have counsel read that clause. Ask, too, which outside code, models and data the build uses, and under which licenses.

NIST's AI risk management framework asks organizations to plan for third-party software and data, including the risk of infringing someone else's rights (the framework's core functions).

  • A good answer: custom work transferred to you in writing, your data always yours, and a list of outside parts and their licenses.
  • A warning sign: the firm keeps the code and rents the system back to you, or the contract says nothing about ownership.

What happens after launch, and how do you leave?

After launch, a good firm watches the system, fixes faults and updates it when a provider changes something underneath. Ask what support includes, how you report a problem and how the firm tells you about changes. Then ask the question that is easiest to skip: how do you leave, and what do you take with you?

AI models change on their provider's schedule, and a new model can change how a system behaves even when nothing else moved. NIST's framework asks organizations to include the pre-trained models they build on in their regular monitoring and maintenance (the framework's core functions). So ask how the firm will test the system again when the model changes.

What should a clean exit include?

A clean exit hands you everything you need to run the system yourself or move it to another firm, and it removes the firm's access to your systems. Ask for the exit terms in the contract before you sign, while you still have the most say. The list below is a fair starting point.

  • the code, settings and documentation for the work you paid for;
  • the accounts the system runs in, held in your name;
  • every password and key, handed over and then changed;
  • the firm's access removed on a date the contract sets;
  • your data returned or deleted, with written confirmation.

Which warning signs should end the conversation?

End the conversation when a firm promises a result before it has seen your work, will not put its answers in writing, or cannot say where your data goes. Any one of these means you would be trusting the firm on the points that matter most. The full list of warning signs is short:

  • hours or revenue promised before anyone has watched the task;
  • pressure to sign quickly, such as an offer that expires this week;
  • no written scope, or a scope that names technology but no result;
  • vague answers on where your data goes or whether it trains models;
  • a contract that says nothing about ownership, breach notice or leaving;
  • claims about AI the firm cannot show you on your own cases.

What does a simple question sheet look like?

A simple question sheet is one table: the question, what a good answer includes and the warning sign to listen for. Take it to each meeting and write each firm's answers beside it, so the differences between firms are easy to see.

The rows below are a starting point. Add questions about the rules of your own sector, such as health or financial records, and ask counsel which rules apply.

  • Questions on the result, the people and your data

    The first six questions cover the result you will measure, the people who do the work and the data the system touches. Ask them at the first meeting, because the answers shape the scope and the data terms of the contract.

    The FTC says those data terms should cover how a vendor may use, keep and delete your data (the FTC's cybersecurity guidance for small businesses).

    Questions on the result, the people and your data
    Question to ask A good answer includes A warning sign
    What result will we measure, and when? One task, a baseline and a review date A figure promised before the work is seen
    Who will do the work? Named roles, one contact and any subcontractors Work passed to companies the firm will not name
    Which outside services will the system use? The AI and hosting providers, named in writing Tools described but never named
    What data will the system touch? The data, where it is kept and who can see it Full access requested before there is a plan
    Will our data train any AI model? A written yes or no for each provider No clear answer
    How soon will you tell us about a breach? A notice window in hours or days, in writing, and help with notices No breach clause, or only a vague word such as promptly
  • Questions on control, ownership and leaving

    The last five questions cover control of what the system does, ownership of what is built and the way you leave. Ask for each answer in writing: the U.S. Copyright Office notes that software is not on its short list of works made for hire (the Copyright Office's circular on works made for hire).

    Questions on control, ownership and leaving
    Question to ask A good answer includes A warning sign
    What may the system do without a person? A written list of allowed actions A promise that nothing needs checking
    How will we know it works? Acceptance tests from our cases, and a sign-off step A demonstration on the firm's own examples only
    Who owns what is built? Custom work transferred to us in writing Code kept by the firm and rented back
    What happens after launch? A named list of what support covers, including new tests when a model changes Support with no stated scope
    How do we leave? Code, accounts, passwords and data handed over, and a date the firm's access ends Systems that run only in the firm's accounts

How does Mavrin Labs answer these questions?

Mavrin Labs puts its answers to several of these questions in writing, as terms of its services agreement and in the written design plan a build includes. After a first conversation, an engagement has three stages: a workflow assessment, a build and a care plan. The agreement you sign is what governs.

  • Who does the work: a team in mixed roles.
  • The result: a design plan, acceptance tests and monthly baseline reviews.
  • Your data: never used to train AI models.
  • Approval: by default, nothing is sent without your approval.
  • A breach: prompt written notice, on the terms set out below.
  • Ownership: the custom work once paid for, and all of your data.
  • After launch: monitoring, fixes, provider updates and one improvement a month.
  • Leaving: access removed and your data returned or deleted.
  • Evidence: no third-party security certification is claimed.

Three of those terms in full: you are told within seventy-two hours of Mavrin Labs discovering, or reasonably suspecting, unauthorized access to your data. Access is removed within five business days after the work ends, and your data is returned or deleted within thirty days after the agreement ends.

The page on Mavrin Labs' data terms sets out these terms in full, and the three engagement stages show where each one sits. The Mavrin Labs AI consulting process, stage by stage adds what happens in each stage.

Sources behind this guide

Every source below was opened on one of the last-checked dates above, and the guide is reviewed quarterly. The table pairs each outside fact with its publisher.

Sources behind this guide
Fact Source
Using AI gives no business an exception from the law, and promised AI income drew cases FTC announcement of its sweep on AI claims
Put vendor security terms in the contract, and confirm the vendor follows them FTC cybersecurity guidance for small businesses
Give contractors no access to sensitive data they do not need FTC guide to starting with security
Check critical suppliers before you buy, and ask for evidence of their practices NIST quick-start guide to supply chain risk
Third-party software and data carry risks, and pre-trained models need monitoring NIST framework core functions
Ask whether multi-factor sign-in is on by default and whether logs come as standard CISA guide for software buyers
A person should check what AI tools write before it goes out SBA page on AI for small businesses
Companies that hold data for another business owe it breach notice and safe disposal Florida's law on the security of personal information
A commissioned work belongs to the buyer only in listed categories, with a signed agreement Copyright Office circular on works made for hire
Business use of AI, overall and among the smallest firms Census Bureau story on AI use by firm size
Reported internet crime losses in 2024 FBI 2024 internet crime report
A voluntary framework for AI risk, released in January 2023 NIST page on the framework

Corrections and changes to this guide, newest last:

  • Tuesday, November 17, 2026: the Copyright Office citation was added, the opening sources line was corrected to cover it, a paragraph pointing to demonstrations no longer published was removed, and the title was changed to the one the owner approved.
  • Wednesday, November 18, 2026: the team is described as a team in mixed roles, with no headcount. Later that Wednesday, the opening sentence on ownership was aligned with the ownership section, the question diagram and the dated account of buyer guidance were added, and the values to ask for were named in the question sheet.

Key takeaways

  • Ask every question before you sign, and keep the answers in writing.
  • Agree one task, a baseline and a review date, never a promised figure.
  • Know who does the work, where your data goes and who approves what is sent.
  • Get ownership and exit terms into the contract, with accounts in your name.
  • Treat pressure, vague data answers and silent contracts as reasons to walk away.

Two related pages help before you meet any firm. The AI automation readiness guide checks that the work itself is ready, and the guide to testing an agent before it reaches clients covers acceptance tests in depth.

The page on AI and your data sets out the terms behind a Mavrin Labs system, and workflow automation describes the service. To put these questions to Mavrin Labs, start a conversation.

Frequently asked questions

Should we pick the firm that promises the biggest result?

No firm can know the result before it has seen your work, so a big promise made early is a warning sign, not a strength. Pick the firm that agrees to measure one task against a baseline you record together, and to check the result on a date you both accept. You can hold a firm to that promise.

Is it rude to ask a firm for its answers in writing?

Asking a firm for written answers is not rude. Careful buyers do it, and a good firm expects it, because written answers become the scope, the data terms and the exit plan in your contract. If a firm will not put an answer in writing, treat that as the answer, and keep talking to other firms.

Do we need a lawyer to read the contract?

A lawyer is worth asking about the contract clauses on ownership, data, breach notice and leaving, because those decide what you can do when something goes wrong. This guide gives you the questions and what a good answer sounds like. Your counsel tells you how the answers fit the law where your business operates.

What if a firm cannot name clients we can call?

A firm may not be able to name its clients without their written permission, so a missing list of clients is not a warning sign on its own. Ask for work you can check another way, such as a sample of past work with private details removed, and judge the firm by how plainly it answers this guide's questions.

The Mavrin Labs mark

Who writes these guides

Client work at Mavrin Labs is delivered by a team in mixed roles, led by our founder. AI tools help with research and drafting, and a person at Mavrin Labs checks each fact against its named source before a guide is published.

Enquiry form

Start with one workflow

Tell Mavrin Labs what is stuck and which systems it touches. A person at the firm reads every enquiry and replies by email.

Start a conversation

Fields marked with an asterisk are the ones a reply needs.

So a reply can address you properly.

The address the reply goes to.

Optional. Give one only if you would rather talk than write; the reply goes by email either way.

Optional, and useful when it says what the work has to fit around.

The result you want, and when you would want it working. Two or three sentences are plenty.

The task as it actually happens, roughly how often it runs, and which systems it touches. Two links at most, please.

Optional.

What you send goes to the Mavrin Labs inbox and is used to answer you, and a phone number, if you give one, is used for the same thing and nothing else.

A person reads it and replies by email.

If the form does not send, write to info@mavrinlabs.com instead; it reaches the same inbox.

Which hours would you take back first?

Name the one workflow that eats the most of your week. Mavrin Labs reads it, and the reply comes from the people who would build the fix.

Start a conversation