How to choose an AI automation agency

How do you choose an AI automation agency? Judge them on five things: whether they quote a fixed price or bill by the hour, whether you see working software every week, who owns the code at the end, whether they will ever tell you not to automate something, and who runs it after handover. Agencies that answer all five well are uncommon. Ones that dodge two or more are telling you, in advance, how the project will go.

This is not a list of vendors. It is the criteria we would use if we were buying instead of selling — including the places where the honest answer rules us out.

What an AI automation agency actually does

The category is young enough that the label covers several different businesses.

Some agencies build software that runs a process end to end — reading documents, routing requests, moving data between systems that were never designed to talk to each other. Some are marketing agencies that use AI tools internally and sell the output. Some are resellers who configure a no-code platform on your behalf and charge a monthly fee for the privilege.

None of those is illegitimate, but they are not substitutes. If you need a supplier invoice pipeline and you hire the second kind, you will get a competent content strategy. Establish which one you are talking to in the first ten minutes.

The test: ask what they would actually deliver. A build agency describes software and where it runs. A reseller describes a platform. A marketing agency describes deliverables.

The five questions that predict whether the project succeeds

Do they quote a fixed price, or bill by the hour?

This is the single most informative question, because it reveals who carries the risk of a bad estimate.

A fixed quote against a written scope means the agency absorbs it. If they misjudge the work, that is their problem. Hourly billing means you absorb it, and it quietly rewards the slowest possible implementation — every complication becomes revenue.

There is a legitimate case for hourly on genuinely exploratory work. But “we can’t quote until we start” on a well-defined process usually means they have not estimated work like this before.

Will you see working software every week?

Ask how often you will see the thing running. Not a status update, not a slide — the software, doing the job, on your data.

Weekly is achievable. We ship a working demo every week and have not found a project where it was impossible. The reason it matters is not transparency theatre: it is that requirements are wrong until someone sees the system working. The tenth invoice always reveals something the kickoff call did not. Discovering that in week two costs a conversation; discovering it in month four costs the project.

If the answer is “we’ll show you at the end of the phase”, ask what happens if the phase is wrong.

Who owns the code at the end?

Get this in writing before anything is built.

You should own the code and the accounts outright, with nothing licensed back to you and no runtime that dies if the relationship does. Some agencies retain ownership and license you the result — which means the price you agreed is a subscription you did not know you were signing.

The practical test: ask what happens if you want to hire someone else to extend it next year. If the answer involves permission, negotiation, or a platform you cannot export from, you are renting.

Will they tell you not to automate something?

Ask for a process they turned down, and why.

An agency that has never declined work is either very new or is selling you whatever you ask for. Plenty of processes should not be automated yet — they run twice a month, the rules live in someone’s head, the data is in a system nobody can get an export from, or nobody will own the result once it exists. We wrote up the checklist we run before quoting, and roughly half of our first calls end with “not yet”.

That sounds like a strange thing for an agency to advertise. It is straightforwardly commercial: the fastest way to lose a client is to build something that automates a broken process, which then fails faster than the manual version did.

Who runs it after handover?

Automation is not a delivery, it is a system that keeps running while the world changes around it. Suppliers change their invoice format. A category gets renamed. Someone adds a field.

Ask two things: what happens when the AI gets something wrong, and who on your side will own the result.

The first answer tells you whether they have shipped anything real. A serious answer describes a review queue — the system handles what it is confident about and routes the rest to a person, with a threshold set against the cost of a mistake. A vague answer about accuracy percentages means they have not operated a system in production, where the interesting question is never the average case.

The second question is about you, not them. Software with no owner quietly dies within a year regardless of who built it.

Red flags

• Accuracy claims with no error handling. “99% accurate” is meaningless without an answer to what happens to the other 1% — and whether you find out.
• A demo on their data, never yours. Any pipeline looks good on the examples chosen to make it look good. Ask for a run on a sample of your real documents, mess included.
• No named engineer. If you cannot find out who will actually write the software, you are talking to a sales layer over a subcontractor.
• Scope that only ever grows. Compare the quote’s scope against the first call. If it has quietly doubled, that is a pricing strategy.
• Platform lock-in framed as a benefit. “Everything lives in our dashboard” means your process lives somewhere you cannot take with you.

Agency, freelancer, or in-house?

Most of the time the real comparison is agency versus doing nothing — because the process is painful but not painful enough to get onto an internal roadmap. That is the honest competitor: the spreadsheet stays, and someone keeps doing the work by hand.

What it should cost

Nobody publishes this, which is itself a signal.

The number varies with three things: how many times the process runs, how many exceptions it has, and how many systems have to be integrated. Exceptions dominate — a process that is 95% identical every time is cheap; one with forty special cases is mostly a specification problem, and that is where the cost goes.

What you should insist on regardless of the number:

• A written scope you could hand to a different agency and get a comparable quote
• A delivery date for a first working version, measured in weeks
• A clear statement of what is not included

An agency that will not commit to a first-version date is telling you it has not sized the work.

Where we fit, and where we don’t

Havoric is an AI automation and web development agency based in Ahmedabad, India, working with clients worldwide. We automate repetitive manual processes and build the web and mobile apps around them, on a fixed quote with a working demo every week, and you own the code.

We meet the five criteria above because we wrote them to describe how we already work, so treat that as disclosure rather than evidence.

We are a bad fit if: you want hourly billing; you need someone on site; you want a platform subscription rather than software you own; or your process is genuinely exploratory research where no one can say what “working” means yet. In those cases a freelancer, an in-house hire, or a product you buy off the shelf will serve you better than we will.

We are a reasonable fit if: you have a process someone does by hand every week, it follows rules you could write down, and you want the software that does it to be yours at the end. That is the AI automation work, and it usually comes with a dashboard or internal tool so the results land somewhere your team can see.

The five questions are worth asking whoever you talk to — us included. An agency that gets defensive about them has answered the question.