AI Automation Agencies Worth a Shortlist, and Where We Fit

agentclaw publishes this page, and agentclaw is one of the seven shapes described below. We sell the build and operate retainer, so a reader who follows this post to its conclusion may end up hiring us. That is the conflict, stated plainly, at the top rather than in a footer.

Here is what we did about it. We do not rank named agencies, because we compete with all of them and a ranking we produce is worth nothing to you. We publish our own prices in full, the same numbers that sit on our pricing page, so you can check whether we are even in your range before you spend a call finding out. And we wrote a section on why you should skip us, with four real reasons in it, because a list that cannot say no to itself is an ad wearing a list's clothes.

What we could not verify: any claim about client counts, win rates or results at any other firm in this market. Nobody audits those. We have left them out rather than repeating them.

Because almost nothing after it changes the answer. 6sense's 2025 buyer experience report found that buyers put about 3.6 vendors on a shortlist on day one, then choose the winner from that same set 95 percent of the time. Four in five deals go to the vendor the buyer already favored before contacting a single seller.

Sit with that for a second. The demos, the discovery calls, the carefully argued proposals: mostly theater performed for a decision that got made in a browser tab. The decision happened upstream, in the twenty minutes you spent typing a query and opening six blue links.

So the useful question is not which agency is best. It is what went into your four slots, and whether the thing that put them there was worth anything.

On 15 August 2026 we opened every ranked list of AI automation agencies returned on page one, and checked four things on each: where the publisher placed itself, whether the page admitted it, whether real prices appeared, and when it was last touched.

Six of six put their own publisher at number one. Two of six disclosed it anywhere on the page. Two of six carried no prices at all, one of them offering relative dollar signs where the numbers should be. The two that did disclose were also the two with the most careful methodology sections, which is not a coincidence, and one of them went as far as naming which firms in its own ranking it had competed against for work.

We have anonymized the six. Naming them turns a measurement into a hit piece, and we would be pointing at competitors while claiming to be neutral. Re-run the search yourself and the pattern falls out in ten minutes.

The directories are a different problem and worth knowing about. DesignRush lists 1,424 companies in its AI automation category as of 14 August 2026, marks some entries Sponsored, and states that some placements may be paid. That is not a shortlist. That is a phone book with an advertising department attached.

Three reasons, and the first one is the honest one. We compete with every firm we could name, so any order we put them in is marketing. You already knew that about the other six lists. It is true of ours too.

Second, a name list rots. Two of the six we checked were last touched in 2025, and in this market that is a long time: teams get acquired, founders leave, a shop that did three good builds last year is now selling something else. The Reddit and LinkedIn threads people actually use as a shortlist rot faster still, and nothing on them is dated.

Third, and this is the part that should annoy you, the names were never the useful bit. Seven genuinely different business models are all selling under the words ai automation agency, and they differ from each other far more than any two firms inside the same shape do. Two shops running a build and operate retainer are near-identical products. A retainer and a staff augmentation contract are not the same purchase at all, and picking the wrong shape is a much more expensive mistake than picking the second-best firm inside the right one.

So work out the shape first. Then take four names inside it, and run the checks at the bottom of this page on all four.

Start here, because it is the cheapest thing that could possibly work and a lot of buyers skip it out of embarrassment. Zapier, Make and n8n sell the plumbing directly, and your operations person builds the workflow.

The prices are public and small. Zapier's Team plan starts at $69 a month billed annually for 2,000 tasks. n8n's Starter cloud plan is EUR 20 a month, and its community edition is free if you are willing to run the server yourself.

Pick this when the work is if-this-then-that: a form fires, a record gets created, a message gets posted. No reading, no judgment, no exceptions that need a person.

It stops working the first time a step needs a decision. Somebody has to read the email and work out whether it is a complaint or a renewal, and no amount of branching logic gets you there. That is the boundary, and it is worth being honest with yourself about which side of it you are on before you spend money on either option.

One person, hourly or per project, found through a marketplace or a referral. This is the cheapest route that involves hiring anyone, and for a single contained build it is often the correct one.

What you are buying is throughput on a task you have already specified. What you are not buying is anyone to think about whether the task is the right one, and you are not buying cover. One person means one calendar, one holiday, and one inbox that goes quiet.

Pick this when the scope is small, you can write it down precisely, and you are comfortable being the project manager. Skip it when the thing has to keep running after they invoice you, because nobody is on the hook for month three.

A named, scoped offer at a fixed price. You approve a number in writing, they build the thing, they hand it over, the engagement ends. Some of the better ones publish the price on the site.

This is a good purchase and an underrated one. Fixed price puts the estimation risk on the seller, which is where it belongs, and a clean handover means you are not renting your own system back. On Clutch's AI agent development category, published minimum project sizes on the first page ran from $5,000 to $50,000 when we checked on 15 August 2026, so the spread inside this one shape is a factor of ten.

Where it goes wrong is month three. The build shop is gone, your ops lead has changed the spreadsheet the automation reads, and nobody notices for a fortnight. If the thing is load-bearing, budget for who watches it before you sign.

Fixed price for the build, then a monthly fee to keep running it, fix what drifts, and add the next workflow. The pitch is that automation is not a project with an end date, and in our experience that is simply true: models change, APIs change, and the people using the system change what they do with it.

This is our shape, so read the next paragraph with that in mind. The honest weakness of a retainer is that you can end up paying monthly for something that is finished. A shop selling this should be able to tell you what the next three months actually contain, and if the answer is vague, you are buying insurance rather than work. Ask what happens if you cancel. The answer should be that you keep everything.

Pick this when there is a queue of work rather than one job, when the workflow touches money or customers, and when nobody on your payroll is going to own it at 6am on a Tuesday.

A senior person, two or three days a week, sitting inside your org and deciding what gets built. Sold as a fractional AI engineer or a fractional chief AI officer depending on how senior the seat is. Usually a monthly retainer on a three to twelve month engagement.

The distinction from shape four matters and gets blurred constantly in sales calls. A retainer sells you throughput: systems shipped and kept alive. A fractional operator sells you judgment: which systems, in what order, and which ones to kill. If you have three departments all convinced their process is the urgent one, judgment is the scarce thing and you should buy that.

Skip it when you already know what you want built. You will be paying senior day rates for a decision you had already made, and a fractional chief AI officer is an expensive way to confirm your own plan.

Vetted engineers by the hour, working to your architecture, on your backlog, managed by you. Sold by the head, priced by the hour, and largely undifferentiated between suppliers, which is why the rates keep drifting down.

This only works when you already have the technical leadership to point them at something. You own the design, you own the review, you own the decision about what good looks like. Hand that to the supplier and you have quietly bought shape three at shape six prices.

Pick this when you have an engineering function that is simply short of hands. Skip it when your reason for looking is that nobody internally knows how to build this, because a body shop cannot fill that gap and will not tell you so.

Multi-year, multi-department, program managers, a statement of work with its own table of contents. The big names sit here and so do the mid-tier firms trying to look like them.

There are real reasons to buy this. If the change spans six business units, if procurement requires a supplier with a recognizable name and a pile of certifications, or if a regulator will eventually want to see governance documentation, a consultancy is doing something the other six shapes cannot.

There are also real reasons not to. Timelines run in quarters, the people who sold it are usually not the people who deliver it, and the deliverable is often a plan rather than a running system. If your problem is one workflow, this is the most expensive way to not fix it.

There is no trustworthy price list for this category. Every band you find in a cost guide traces back to an agency quoting itself, which makes it marketing with a decimal point. Rather than launder a confident-looking range, here is what a buyer can actually open in a browser and check today.

Two of the four sources below are platform vendors publishing their own pricing, one is a directory publishing minimum project sizes that firms declared themselves, and one is us. Ours are marked, and they are the same figures on our pricing page rather than a range invented for this post.

We are shape four. We build one workflow at a time, we hand you the accounts and the repo, and then we run it and expand it monthly. Starter builds are $1,500 to $2,500 fixed. A two-week production sprint that puts one full workflow into production is $5,000 fixed. Retainers start at $5,000 a month, and that is the floor for ongoing build-and-run work rather than a minimum you have to spend to talk to us. The starter build exists precisely so it does not have to be.

Model API costs, tools and subscriptions are yours directly and never marked up. Cancel the retainer and you keep everything we built. The audit at the front is free and yours to keep whether you hire us or not.

Now the reasons to skip us, which is the part of a self-ranked list that is worth anything.

Your product is software and you have engineers. Then you should build this. You will build it better, you will own it properly, and you do not need a supplier in the middle of your own stack.

The work does not repeat. If it happens four times a year, automating it is a hobby. Do it by hand and spend the money on something that happens on Tuesdays.

You want one build and then silence. That is a legitimate purchase and a productized build shop is a cleaner fit for it than a retainer is. Do not buy a monthly relationship because a monthly relationship is what somebody sells.

Your problem is direction, not delivery. If you cannot say which workflow matters most, hire judgment first. That is shape five, and buying throughput before you have direction is how people end up with four automations nobody uses.

The instinct that says building it internally is lower risk has the arithmetic backwards. MIT's Project NANDA report "The GenAI Divide: State of AI in Business 2025", reported by Fortune, found that tools bought from specialized vendors and built through partnerships succeed about 67 percent of the time, while internal builds succeed one third as often. That study is also where the widely quoted line about 95 percent of pilots delivering no measurable profit and loss impact comes from, and it rests on 150 leader interviews, a survey of 350 employees, and 300 public deployments.

The other half of the picture is production. S&P Global Market Intelligence surveyed more than 1,000 organizations in North America and Europe and found the share abandoning most of their AI initiatives had jumped to 42 percent from 17 percent a year earlier, with 46 percent of proofs of concept scrapped before production.

Put those together and the shortlist question sharpens into something answerable. You are not looking for the best agency. You are looking for the one most likely to get a thing into production and keep it there, which is a much easier property to test for.

This is the part that outlives the list. Run it on all four names, in one sitting, before any of them get a call. Most of it you can do from their website, and the rest fits in the first ten minutes of a first conversation.

1. Find a price. Anywhere. A range, a minimum, a starting number. A firm that will not publish a floor has decided the number depends on what you look like you can pay.
2. Ask what broke. Ask for a workflow that failed in production and what they changed afterwards. Anybody who has shipped has one of these and enjoys telling it. A spotless record means nothing has been in production long enough to break.
3. Ask who holds the keys at the end. The repo, the API keys, the platform accounts, the billing. If the answer is that it lives in their workspace, you are renting.
4. Ask about model costs. What does it run per month, and do they mark it up? Both answers are informative, and the second one should be no.
5. Ask for a fixed price on the first build. Not the program. The first thing. A shop that cannot price one scoped workflow has not scoped enough of them.
6. Ask for a runbook from a past engagement with the client details stripped out. What each step does, what breaks it, what to do at 6am when it does. This one separates people who ship from people who demo.
7. Ask what they would tell you not to build. If everything you describe is a great fit for their services, you are talking to a salesperson.

Two of those seven are about ownership and two are about money, which is deliberate. Those are the questions that get quietly deferred to the contract, and the contract arrives after you have already chosen.

There is a longer version of this test for anyone who cannot read the code they are about to be handed. And once a proposal actually lands, run it past the separate list of red flags that should end the meeting rather than start a negotiation.

Give all four the same scoped brief, in writing, and ask for a fixed price on it. One workflow, described in a page: what comes in, what the rules are, what should happen to the exceptions, what the finished thing must be able to do.

You learn more from the four replies than from any amount of case study reading. One will requote your brief back at you. One will ask three questions you had not thought about, which is the good sign. One will refuse to price it without a paid discovery phase, which is a legitimate answer if the paid phase is small and produces a document you keep. And one will come back with a number that is either suspiciously round or suspiciously vague.

Then pick, and pick fast. The 6sense data says you are going to choose from those four anyway, so the value of a long evaluation is mostly the comfort of having done one. Spend the time on the brief instead. A precise brief makes a mediocre supplier useful, and a vague one makes a good supplier expensive.