AI agency or AI consultancy? The label is the least useful thing about them.

We call ourselves a consultancy, so this is not a neutral document. What follows is still the honest version, because the useful part is not the definition — it is the set of questions that separates a supplier who will deliver from one who will not, and those questions apply to us as much as to anyone.

What the two words traditionally mean

Stripped of marketing, they describe different commercial shapes.

The distinction that actually survives contact with reality is this: an agency is usually paid to build what you asked for; a consultancy should sometimes be paid to tell you not to build it. If a supplier has never ended an engagement at “this is not worth doing”, they are selling output regardless of what the website calls them.

Why the distinction stopped being useful

Something specific happened to this market between 2024 and 2026. The barrier to entry collapsed, and a large number of near-identical firms launched at the same time, taught the same playbook by the same wave of online courses. “Start your own AI automation agency” is still a product you can buy on Udemy and Gumroad this afternoon.

That is not a criticism of anyone learning a trade. It does mean the word “agency” in this category now covers an unusually wide range — from firms with a decade of delivery behind them to firms that finished a course last quarter — and the website will look much the same either way.

Two independent signals suggest the consequences are already showing.

Set against that, the more considered read of the market is that it is saturated for lookalikes and wide open for specialists. Firms with genuine sector depth report full pipelines. Generalists pitching identical automations to the same prospects do not. That is the real division in this market, and it does not follow the agency-versus-consultancy line at all.

The questions that actually sort it out

Ask these of any supplier, whatever they call themselves. Ask them of us.

1. What have you put into production, and can I use it?

Not a portfolio of mockups or a case study with no link. Something running, that you can open in a browser and interact with. A screenshot proves a design tool was opened; a live product proves the hard part was finished.

2. Who is actually doing the work, and what have they done before?

A named person with a traceable history, not “our team of experts”. If nobody is named anywhere on the site and no profile exists to check, that is information in itself.

3. Have you ever told a client not to proceed?

Ask for the circumstances. A supplier who has never concluded that a project was not worth doing has either been extraordinarily lucky or is not making that assessment at all.

4. What happens when the model is wrong?

Confidence thresholds, human review, logging, and defined behaviour when the model is unavailable. If these arrive as an afterthought rather than part of the quote, the answer is that nobody has run one of these in anger.

5. Does this need AI at all?

The most revealing question on the list. A great deal of what gets sold as AI is better served by rules and integration, which are cheaper, more predictable and easier to audit. A supplier who answers “yes” to every version of this question is selling, not advising.

6. What does it cost to run, not just to build?

Usage costs, model costs, and what happens to them at three times the volume. Firms who have not operated something at scale tend not to have thought about this.

7. What do I own at the end?

Code, data, accounts, credentials, and the ability to take it elsewhere. The answer should be immediate and unambiguous.

Which model actually suits you

Your situation	What you need
You know exactly what to build and have specified it	An agency or a development shop. You are buying delivery, and paying for advice you do not need is waste.
You have a defined, ongoing output requirement	An agency on retainer. Predictable scope suits a predictable commercial model.
You know something is wrong but not what to fix	A consultancy. Most of the value is in correctly identifying the constraint before anything is built.
The work touches several systems that must keep working	A consultancy with integration depth. The risk sits in the joins, not the features.
You are not sure AI is the answer	A consultancy — and specifically one prepared to tell you it is not.
You need it running and supported for years	Either, provided they can show something they still maintain. This is the question most suppliers fail.

Where we sit, and where we do not

Megabite is a consultancy. Engagements start with a free conversation about the constraint, and they do sometimes end there, because the decision order is buy, configure, integrate, automate, then build — AI enters only where the input varies in ways rules cannot capture.

On our own list of questions: the work is done by Thomas Barrie, twenty years across estate agency, residential property investment at HBOS, lead generation and a decade of software delivery, with a team behind the engagement. Three of the products on our work page are live and you can go and use them right now, which is a more useful signal than any case study.

Where we are not the right answer: if you have a fully specified build and simply need hands, a development shop will be cheaper. If you want the lowest possible price, we will not be it. And if you want a large simultaneous programme across many departments at once, that is a different kind of firm.

Common questions

What is the difference between an AI agency and an AI consultancy?

An agency is generally paid to deliver a defined output, often on retainer or by project, with the scope arriving largely specified by the client. A consultancy is paid partly to determine what should be built in the first place, including concluding that nothing should. In practice the labels overlap heavily, and a supplier's track record predicts the outcome far better than the word on their homepage.

Is an AI agency or a consultancy cheaper?

An agency is usually cheaper when you already know precisely what you want built, because you are not paying for the diagnostic work. A consultancy tends to be better value when the problem is not yet well defined, since the most expensive outcome in this field is building the wrong thing competently.

How do I know if an AI supplier is any good?

Ask for something in production you can use yourself, a named person with a traceable history, and an example of an engagement they advised against. Then ask what happens when the model is wrong, what it costs to run at three times the volume, and what you own if you leave.

Why are there suddenly so many AI agencies?

The barrier to entry fell sharply and a wave of online courses between 2024 and 2026 taught the same playbook to a large number of people at once, producing many near-identical firms. The market is widely described as saturated for generalists while remaining open for specialists with genuine sector depth.

Do I need AI at all?

Frequently not. A great deal of what is sold as AI is better handled by deterministic rules and systems integration, which cost less, behave predictably and are far easier to audit. AI earns its place where the input varies in ways rules cannot capture — interpreting a message, holding a conversation, or extracting data from inconsistent documents.