How to choose an AI automation partner

This guide is for people buying AI automation for the first time, or buying it again after a project that did not stick. It does not assume you can judge the technical work, because most of what predicts a good outcome can be checked without doing so. For a side-by-side table of the four routes, see how to get AI automation built. Legal points below are general information, so have your lawyer read the contract.

The four routes in brief

These are generalisations. Individual firms and people vary, and a strong freelancer can beat a weak consultancy.

Route	Tends to suit	Watch for
Freelancer	One well-specified task on one platform	Cover for holidays and illness; accounts set up in their name
Agency	A defined project with a clear brief	Change requests when the brief was incomplete; handover quality varies
In-house hire	A long, steady backlog and someone who can manage an engineer	Months to recruit; one person rarely covers every platform; leaver risk
Consultancy	Work across several teams, where finding what to change is part of it	Higher cost for small tasks; needs leadership time; can over-scope

1. Write down the work and the result you want

Write one page before you speak to anyone. Name the process, who does it today, how many hours it takes a week and which systems it touches. Then write the number that should change, such as days to close the month or invoices keyed by hand.

If you cannot fill this in, say so to the candidates. Some firms offer a paid audit for exactly this situation, and you can also run an audit yourself. A vague brief produces quotes you cannot compare.

2. Decide which of the four routes fits

Two questions do most of the sorting. How long will the work last, and how well can you describe it today?

Work you can describe precisely and that will finish suits a freelancer or an agency. Work that will keep coming for years suits a hire, if you have someone to manage them. For scale, the US median wage for a software developer was $135,980 in May 2025, before recruitment, benefits and management time, and pay varies widely by country. A steady backlog can justify that, and a few projects a year usually cannot. Work that you cannot yet describe, or that crosses several teams, suits a consultancy or a paid audit from any route.

These routes also combine. Many companies have an outside firm build the first systems, then hire someone once the backlog is clear.

3. Ask who owns the accounts and code

Skipping this question causes a lot of trouble later, and it costs little to settle at the start.

• Whose name is on each account? Hosting, automation platforms, model providers, databases and code repositories should be in your company name, with billing you can see. The supplier gets access that you can remove.
• Who owns the work under the contract? Do not assume that paying transfers ownership. In the US, for example, work by an employee as part of their job belongs to the employer, but commissioned work counts as made for hire only in nine listed categories and with a signed agreement, and custom software is not named in that list. Rules differ by country, so have the contract assign code, prompts, workflows and documentation to you.
• What is reused? A supplier may have its own libraries or templates. That can be fair, as long as you get a written licence to keep using them if you part ways.
• Who handles personal data? Under the GDPR, a supplier handling personal data for you needs a contract covering that processing, and it needs your written authorisation before bringing in another processor. Ask for the agreement and the list of sub-processors.
• Could someone else take over? Ask what they would hand to a new supplier, and ask to see an example of a handover document.

4. Ask how the price is set

No pricing model is honest or dishonest on its own. Each one moves risk to a different party, so ask how this candidate’s works.

• Fixed price. You know the number, and the supplier carries an overrun. Ask how scope is defined, how a change is priced and who approves it.
• Hourly or daily. This fits unclear work, and you carry an overrun. Ask for a cap, regular reporting and the right to stop.
• Retainer or monthly. This fits ongoing work. Ask what is included, whether unused time rolls over and how much notice ends it.
• Share of savings. It sounds aligned, but ask who measures the savings and what happens when the numbers are disputed.
• Paid discovery or audit. This reduces the risk of building the wrong thing. Ask what you receive, whether you can take it to another firm and whether the fee is credited to the build.

Then ask about the costs that sit outside the fee: model usage, platform licences, hosting and support. A cheap build on a tool with a high per-user bill can cost more over three years. Ask for a figure at your expected volume, and for what happens to it if volume doubles. If a firm publishes prices, as we do on our pricing page, use them as one reference point and check what they include. A firm that does not publish prices can still be fair, so judge it by how clearly it explains a quote. Our guide to what AI automation costs breaks down the parts.

5. Ask what happens after go-live

Automation breaks in ordinary ways. A vendor changes an API, a password expires, a model starts answering differently or a form gains a new field. So ask how each candidate deals with that.

• Who gets told when it fails, and how? Ask to see an alert from a live system. A workflow that stops quietly is worse than none, because people assume it still runs.
• How fast do they respond, and is that in writing? Ask about support hours, and what counts as urgent.
• What does support cost after the first period, and can you cancel?
• Who checks that people use it? Ask what happens if the team goes back to the spreadsheet, and whether anyone is responsible for adoption.
• How are AI outputs checked over time? Ask about test sets, what happens when a model is updated, and which actions need a person to approve them. Our guide to putting an agent into production lists what to expect.
• Can your own team make small changes? Ask for the documentation and training that allows it.

6. Check the proof and who does the work

Ask for one example close to your work, with four things: the number before, the number after, the time period and how it was measured. Then ask whether the system is still running and who looks after it now. A result measured over a week is a pilot, and a result still holding after a year tells you much more.

Ask to speak to a past client, without the firm on the call. Good questions to ask them are what went wrong, how it was handled and whether they would hire the firm again. Some firms cannot name clients because of confidentiality, which is a fair reason, but they should still be able to offer a reference call under the same terms.

Treat partner badges as a partial signal. Make’s programme requires partner training and prior automation experience, and Zapier’s has tiers with their own requirements. That shows a firm met a vendor’s criteria, and it is a reasonable filter, but you still need to ask about client results.

Finally, ask who will do the work. Meet the person who will build your system, and ask what they built last. Ask about subcontractors, about the backup if that person is ill or leaves, and how much of the work is written by AI tools and who reviews it. Using AI tools is normal now, so what matters is that a named person understands and answers for the result.

7. Score the answers and start small

Put each candidate in a column and each question in a row, and score the written answers before you think about who presented best. Weight ownership and support heavily, because those are hard to fix after signing. Price comes after those, and only once you are sure the quotes cover the same scope.

Then buy something small: an audit, one workflow or a two-week piece with its own price and a clear finish. You will learn how they communicate, whether they test and what their handover looks like, and you can stop if it goes badly. An Automation Audit is how we start, and other firms sell similar first steps. Ask whether you can take the findings elsewhere, wherever you buy one.

Warning signs

None of these proves a firm is bad, but each deserves a direct question.

• Accounts or code stay in the supplier’s name, or access is offered only on request.
• The firm guarantees a percentage saving before it has looked at your process.
• The proposal names a tool before anyone has asked how the work runs.
• Every example is a demo, or results come with no baseline or measurement method.
• The people in the sales meeting cannot say who will do the work.
• There is no answer on how failures are detected, or the answer is that nothing fails.
• Support is undefined, or cancelling means losing the system.
• Pricing cannot be explained in a few sentences, or running costs are left out.
• Prompts or logic are described as proprietary, so you could not move them.
• Personal data questions get vague answers, and no processing agreement is offered.
• The firm pushes for a large contract with no smaller first step.
• Nobody ever tells you what they would advise against.

When SBD is the wrong choice

We sell fixed-price builds to companies of 5 to 1,000 people, starting from an audit, so several buyers are better served elsewhere.

• You have one well-specified task on one platform. A good freelancer will usually be faster and cheaper, as long as the work is built in your accounts and comes with a written handover. Our audit is more process than this needs.
• Your budget is below our starting price. Builds start at $1,500 after a $500 audit. Below that, a freelancer or a person on your team using Zapier, Make or n8n is a sensible first step.
• You have years of steady work and someone to manage an engineer. Hire in-house. You will build knowledge inside the company that no outside firm can match.
• You need a full ERP or CRM rollout, such as Dynamics 365. Use a specialist implementation partner for that product, because we do not run large rollouts of that kind.
• You have more than 1,000 people, or procurement needs a large supplier. A large systems integrator or global consultancy is a better fit for that scale and paperwork.
• You want staff to learn to build it themselves. A training provider or a hire is a better match, although we can include training in a build.
• Nobody on your side can give it time. Any outside firm will struggle if no one can answer questions and make decisions. Wait until someone can, whichever route you choose.

Common mistakes

• Choosing on the demo. Demos use clean data, so ask to see the system handle a bad input.
• Comparing quotes that cover different scope. Send one brief and ask for the costs to be split the same way.
• Leaving ownership until the contract stage, when the supplier has already set up the accounts.
• Treating go-live as the end. Budget for support and for someone on your side to own the system.
• Starting with the biggest project. Begin where a failure is cheap and the result is easy to measure.