How to buy from an AI automation agency without wasting your money

A professional services firm asked me to look at their business. We went through it properly and came out with thirteen things worth doing.

I sent them the list with no prices on it.

I asked them to tell me which ones actually mattered and in what order. Prices came afterwards, one project at a time, so they could buy the first one and see whether it worked before committing to anything else.

I should be straight with you: that is a sales method. It is how I sell. Saying so out loud is more useful than pretending it is pure altruism.

But it exists because of something I kept watching happen.

The number that should worry an agency more than it worries you

S&P Global research, reported by CIO Dive, found that 42 per cent of organisations abandoned most of their AI initiatives during 2025, up from 17 per cent the year before.

Most of them.

That survey asked enterprise IT decision makers, not small businesses, so it describes how big projects fail rather than your situation directly. I still think it is the most useful number in this whole area, because the failure it describes is one I watch small firms walk into on a smaller budget.

Nobody abandons a project that took a fortnight and worked. What gets abandoned is the big one. The programme with a discovery phase and a roadmap and a number at the bottom with a comma in it, that was going to change how everything worked, and eleven weeks later nobody can remember what problem it was solving.

An AI programme sold as one large number gets deferred. The same work sold as one project at a time gets started.

What an agency is for, and what it is not for

The British Chambers of Commerce, with the University of Essex, found that 54 per cent of firms in their network use AI, but only about one in ten use anything bespoke. Their own authors are careful to say that describes their responding sample rather than the whole country, and it is worth repeating that.

But the split is the interesting bit. Nine out of ten are getting whatever they get from tools they simply bought.

Which means most of what a good automation agency does is not building software. It is finding the two or three places where a bought tool, used properly by someone who has been shown how, removes a job nobody enjoyed.

The build comes later, if at all. Usually when you have something specific enough that no off the shelf tool covers it.

If the first conversation goes straight to a custom platform, you are being sold the expensive version of a problem you have not defined yet.

What to ask before you sign anything

Three questions, and you can ask all of them on a first call.

What is the smallest version of this that would tell us whether it works. If there is no small version, the project is too big to start.

What are you recommending we do not automate. Anyone who says nothing has not looked properly.

What happens to this when you are not here. A workflow only one person understands is a workflow with an expiry date on it, and if that person is the agency, that is worse rather than better.

What you could do on Monday

Write down every job in the business that happens on a schedule and produces a document. The monthly report, the weekly stock check, the invoice run, the quote pack.

Not the hard ones. The regular ones.

Then put a number next to each: roughly how long it takes, and roughly how often. Twenty minutes, weekly. Half a day, monthly.

You now have a shortlist ranked by hours a year, which is a better brief than most agencies get given. Take it to whoever you talk to and make them tell you which one they would start with, and why.

The honest limit

An agency cannot fix a business that has not decided what it wants.

The thirteen project list worked because that firm knew their own priorities once they saw the options laid out. Some businesses genuinely do not, and no amount of discovery from an outsider substitutes for the owner making a decision about what matters this year.

If that decision has not been made, the automation will get built, it will work, and it will quietly stop being used by March. Not because it was wrong. Because it was never anybody's priority in the first place.

That is the failure mode nobody puts in a case study.