Does my business actually need an AI agency?

“AI agency” does not describe a single trade

The label is used by companies that have almost nothing in common. Some build custom software and work on their own models. Others configure tools that already exist — chat assistants, automations, transcription — and fit them to your processes. Others again are web or marketing shops that added a line to their offering. All three are entitled to call themselves an AI agency. The label therefore tells you nothing about what you will actually receive.

A simple test before the first meeting: ask what they would deliver if the words “artificial intelligence” were banned from the conversation. A supplier who knows their trade describes a deliverable in ordinary terms — a form that sorts your incoming requests, a tool that drafts your quotes from your own price list, automatic transcription of your field notes. One who can only answer in generalities is selling the category, not a solution to your problem.

Name the problem before you shop for the tool

Most disappointing engagements start with the tool: someone saw a demo and wants to try it. The reverse order works better. First identify a task that recurs, that you can describe, and whose cost in hours you can roughly estimate. “Answering the same fifteen questions by email”, “transcribing notes taken on site”, “preparing the same quote with three variants”. Those are problems. “Using AI” is not one.

If you cannot explain how the task is done today, step by step, no supplier will automate it correctly — they will automate their guess at your process. That description work is yours to do, and it is not wasted. It is also what lets you compare two offers on the same basis, instead of comparing two pitches.

What must remain yours when the engagement ends

Ask before signing, not at the end: will the accounts opened with third-party providers be in your company's name or the agency's? Will your data be exportable in a format you can open without them? Will the instructions, configurations and automations be documented in writing? Who holds the access keys, and who pays the subscriptions?

A second question usually settles the rest: if we stop working together next month, what keeps running and what stops? The answer tells you whether you are buying a capability that stays with you, or renting a dependency. Both can be defended. But the price does not read the same way in each case, and that distinction almost never surfaces on its own in a proposal.

Test it on your own files, in the language your customers write in

A demo is built to succeed. Ask instead to test on your own material: your real emails, your trade vocabulary, your customers' names, the way people actually write to you. In Quebec that often means both French and English in the same inbox, with local turns of phrase and the abbreviations your team uses. A tool that impresses in English on a generic case can be unusable on your files.

On data, three questions to have written into the contract rather than merely said in a meeting: where is the content processed and stored, is it used to improve the provider's models, and who at the agency can access it? If your business handles personal information about its customers, have a qualified person confirm your obligations before anything is sent to an outside tool.

A first engagement you can actually judge

Before starting, agree on what would count as success and on the date you will judge it. Something observable: response time to incoming requests drops below a set limit, quotes go out the same day, a given task stops eating Friday afternoons. Write it into the agreement.

Above all, accept that the answer may be no. A pilot with no failure criterion always works: something positive can always be reported, and the natural conclusion is a second invoice. A pilot that is allowed to fail teaches you something either way — either you have a measurable gain worth extending, or you have ruled out an avenue for an amount you decided in advance.

When you do not need an agency at all

There are situations where the right decision is to buy nothing. If the task recurs only a few times a year, implementation and upkeep will cost more than the task itself. If your bottleneck is elsewhere — not enough requests coming in, rather than too much admin to process — an internal tool will fix nothing.

And if nobody on the team has the time or the appetite to maintain what gets built, the tool will be abandoned within months, however good it is. In those cases the useful step is usually more modest: train one person on tools you already pay for, and watch what they get out of it for a quarter before hiring anyone.

What does an engagement with an AI agency cost?

The spread between suppliers is wide, and it depends on whether you are buying a configuration of existing tools or custom development. What matters more than the number is what the price includes: the accounts, the documentation, the upkeep, and what keeps working if you stop. Ask for a detailed estimate rather than a monthly figure.

Can I not just use ChatGPT myself?

Often yes, to begin with. An agency becomes useful when the tool has to connect to your own systems, run without someone driving it, or be maintained by someone other than the person who set it up. If none of those three applies, start on your own.

How long before I see a result?

It depends entirely on the engagement, and no general timeline means anything. What you can fix in advance is the date on which you will judge it, and the observable measure you will judge it by. Without those two, any outcome can be presented as a success.

Is AI going to replace my employees?

The useful question is more down to earth: which specific tasks do you want taken off their plate, and what would they do with the time. Engagements that go well usually remove recurring administrative work, not people.

How can I tell whether the agency really builds, or just resells someone else's tool?

Ask directly, and ask for the names of the third-party tools involved. Reselling is not a problem in itself — hiding it is. What you want to know is what you would still hold if the agency disappeared: the accounts, the configurations, the data, and the documentation.