What is an AI agency?

What an AI agency actually does

The label is used loosely, and that is most of the confusion. In practice, a competent AI agency does three things, in this order.

• Diagnoses before it builds

  Maps how the work happens today and puts a dollar figure on the parts that repeat. That number decides what gets built, and whether anything should be built at all.

• Builds into the tools you already run

  The system fits around the phone, inbox, booking tool and accounting software already in use. Nobody is asked to learn a new platform, which is usually why these projects fail.

• Measures what changed

  Hours saved, response times, revenue recovered. Without a measurement, the work is a subscription rather than an investment, and nobody can tell whether to do more of it.

What an AI agency is not

• Not a chatbot vendor

  A widget that answers questions on your website is a feature. It is not an operating change, and it rarely moves a number that matters.

• Not a reseller with a markup

  Some firms put their name on someone else's software and call it an implementation. Ask what they actually built, and what happens to it if they disappear.

• Not a prompt-writing course

  Teaching a team to write better prompts is useful coaching. It is not a system, because it still depends on a person remembering to do it.

• Not a build-and-vanish project

  Anything that only works while the builder is in the room will quietly stop working about a month after they leave.

Agency, freelancer, in-house hire or off-the-shelf tool

An AI agency is one of four ways to get this done, and it is not always the right one. The honest comparison:

Most owner-led businesses do not need a hire. They need one or two workflows fixed properly and then left alone.

What it costs, and the test that matters

Pricing usually splits into project work to build something, and a smaller ongoing amount to keep it running and improve it. Ranges vary enormously by scope, which is why a quoted price with no diagnosis behind it tells you very little.

The test worth applying is simpler than comparing quotes. The first thing built should pay for itself within a year, and you should be shown that arithmetic before you commit rather than after. If nobody can show you the sum, nobody has done the thinking.

How to tell a good AI agency from a bad one

These six questions separate the firms doing real work from the ones selling access to software you could buy yourself.

• Do they put a number on the problem first?

  If nobody has estimated what the problem costs you today, nobody can honestly say whether the fix is worth buying.

• Do you own what they build?

  The code, accounts, data and documentation should be yours, and another provider should be able to take over. If that question is awkward to answer, you are renting, not buying.

• Do they work with your tools, or insist on theirs?

  An agency that needs you on its own platform has a retention strategy. That is not the same thing as a recommendation.

• Will they tell you no?

  A willingness to say a project is not worth doing is the cheapest honesty signal you will ever get, and the rarest.

• Who actually does the work?

  Ask whether the person on the call is the person building it. Handovers are where context and accountability go missing.

• What happens if you leave?

  Ask it on the first call. The answer tells you how the whole relationship has been designed.

When you do not need an AI agency

Plenty of businesses are better served by doing nothing yet. A good agency will say so.

• The process is not settled

  Automating a workflow nobody has agreed on multiplies the confusion and makes it harder to see.

• The volume is genuinely small

  If something happens twice a month, a person doing it by hand is cheaper than a system somebody has to maintain.

• Ordinary software already solves it

  Many problems are answered by tools that already exist for a modest monthly fee. That answer should be free advice, not a project.

• It is a people problem

  If work falls over because nobody owns it, no amount of software will fix the ownership question underneath.

Four questions worth asking on a first call

• What would you not build for us?

  Tests whether they are diagnosing or selling.

• What does this cost to run once it is live?

  Build cost is the smaller half of the decision for anything that runs continuously.

• Show me something you built that is still running a year later.

  Plenty of demos work once. Fewer survive contact with a real business over twelve months.

• Who maintains it, and what happens if we part ways?

  The answer should be boring. If it is not, that is the finding.