AI consultancy or AI agency: what you are actually buying

When a company starts looking for help putting AI into its operation, it searches for "AI agency" and "AI consultancy" — usually within the same afternoon. Both searches return providers who introduce themselves identically: same vocabulary, same client logos, same tone. Behind them, though, sit two different business models, with different incentives, that produce different outcomes for the company doing the hiring.

This article isn't about deciding which one is "better". It's about knowing what you're signing. Because the expensive mistake isn't picking the wrong model — it's hiring one model while expecting what the other does, finding out in month four, and having no contract to point at.

What separates an AI consultancy from an AI agency?

The difference isn't team size or technology. It's what the provider commits to deliver.

An agency sells deliverables. A working chatbot, an integration, a dashboard, a generated video, a campaign. Scope is closed before work starts, price is tied to that scope, and the project ends when the deliverable exists and is signed off. It's a production model: the more efficiently the agency builds that piece, the better it does. Its muscle is executing quickly things that resemble what it has built before.

A consultancy sells judgement and accountability for the outcome. The engagement doesn't start at "build me this", it starts at "is this the right thing to build?". Diagnosis, prioritisation, solution design, and then yes, execution — but with a commitment that the thing works inside your operation, not just in the demo. Scope gets adjusted when the diagnosis says it should be.

There's a third figure worth naming, because it muddies the market: the product integrator or partner, who sells implementation hours for one specific tool. It's neither consultancy nor agency; it's a channel. It may be exactly what you need if you've already chosen the tool. It's the worst possible place to ask "which tool do I need?".

If your provider is paid to deliver a piece, they will argue the piece is the solution. That's their incentive, not their bad faith.

What are you really contracting in each model?

On the proposal cover page both offer "AI applied to your business". In the fine print, they don't.

| | AI agency | AI consultancy | |---|---|---| | Starting point | A deliverable already defined by the client | A business problem still to be diagnosed | | What the contract fixes | Scope and price of the piece | Objective, indicators and phases | | Who decides what gets built | The client | Decided jointly, using diagnosis data | | Who you deal with | Salesperson plus builder | A partner or senior who owns the project | | End of project | Delivery and sign-off | Adoption measured in production | | Typical cost | Lower and predictable | Higher, with more variability up front | | Risk you carry | The piece doesn't solve the problem | Diagnosis stretches the start date | | Good fit when | You know exactly what you want | You don't know where to start, or there are several fronts |

Read it in both directions. The agency column isn't the bad column: if your problem is diagnosed, the process is stable and you just need someone to build the piece, paying consultancy fees to confirm what you already know is wasted money.

The trouble starts when the diagnosis was done by the client in a 45-minute board meeting and left that room as a requirements document. The agency will build exactly what you asked for, and will be right when it hands it over. What it won't build is what you needed.

When does an agency fit, and when do you need a consultancy?

Four real situations, with the honest answer:

• "We want a website chatbot that answers from our catalogue." Clear scope, data available, low risk. Agency. Ask for CRM integration and human handover, and don't pay consultancy fees for this. If you're not sure a chatbot is the answer at all, start with when chatbots pay off and when they don't.
• "We have eight AI ideas and budget for two." That's prioritisation, not production. Consultancy. The value is in discarding six with reasoning and putting numbers on the two that survive.
• "We ran a pilot last year and it stalled there." It's almost never a technical problem. It's process, internal ownership or data. Consultancy, with a diagnosis before anything else gets built — we cover the anatomy of that failure in why AI pilots fail.
• "We need to automate supplier invoice entry." Depends on whether your supplier master data is clean. If it is, agency. If you don't know, that "don't know" is precisely a consultancy's job.

The pattern: an agency pays off once the uncertainty is resolved. A consultancy pays off when the uncertainty is the main problem. Buying production with the uncertainty still open is the most common way to spend the year's budget on a technically correct piece that nobody uses.

Which signals reveal an agency presenting itself as a consultancy?

Everyone in the market calls themselves a consultancy now. These five signals separate the label from the actual model, and all of them show up in the first meeting:

• They propose a solution before seeing any of your data. If you already know what they're going to build after the first call, there was no diagnosis. There was a catalogue.
• The proposal has no indicators. "Implementation of a conversational assistant" is not an objective. "Cut average tier-one response time by X, measured against ticket history" is.
• The person selling isn't the person building. Ask directly: which of the people in this room will still be on the project in month three? Then ask for it in writing.
• They don't ask about your systems. Every serious project lives or dies in the integration: ERP, CRM, permissions, who can see what. A provider who doesn't ask about that will hand you an island.
• The price doesn't move with what the diagnosis finds. If the quote is identical before and after looking at your data, the data never mattered.

If you want the full interrogation, with the exact questions for that meeting, it's in how to choose an AI consultancy.

What does each model cost, and where is the hidden cost?

Comparing hourly rates is misleading, because the two models hide the cost in different places.

In the agency model, the rate is lower and the initial budget tighter. The hidden cost sits in scope changes: anything not in the specification is a change order, and AI projects produce plenty of them, because how the system actually behaves with real data isn't known until it's tested. It also sits in maintenance: a piece that's delivered and then left alone degrades, and in AI it degrades faster than people expect.

In the consultancy model, the rate is higher and the start slower, because there's a diagnosis phase that produces nothing visible. The hidden cost is the never-ending diagnosis: beautiful reports, not one line of code, and five months in you still have nothing in production. That risk is real, and it's covered by something very simple: put a date and a deliverable on the diagnosis phase in the contract. Weeks, not months. And make it end in prioritised decisions, not a slide deck.

Our practical recommendation, whoever the provider is: contract in two tranches. A short, fixed-price first tranche to diagnose and prioritise. A build tranche that is only signed with the result of the first one on the table. That turns uncertainty into an informed decision instead of a change order. It's exactly the logic behind the AI audit.

How do you write the brief so the label stops mattering?

The provider's label matters far less when your brief is well written. Five elements, none of which require you to be technical:

1. The problem, not the solution. "Finance spends X hours a week reconciling delivery notes" instead of "we want an AI agent". 2. The metric and its baseline. If you don't know today's number, you won't be able to prove tomorrow's improvement. Getting that baseline is work, and it's the first piece of work. 3. The systems involved and who grants access. By name. Half the delays in these projects are people waiting for credentials. 4. Who owns it internally. Someone in your house who decides and answers for it. Without that, no provider can rescue the project. 5. What happens after delivery. Who maintains it, how often it's reviewed, what happens when quality drops. Putting it in the initial contract costs one paragraph; negotiating it afterwards costs a whole renewal.

With that brief in hand, a good agency will tell you honestly that there are questions it can't answer, and a good consultancy will tell you which part is diagnosis and which part is build. Both answers are useful. The one that should set off alarms is the one that says yes to everything.

Frequently asked questions

Is an AI consultancy more expensive than an AI agency? By hourly rate, yes. By total cost of the problem solved, it depends on how much uncertainty you start with. With clear scope the agency is cheaper; with fuzzy scope the consultancy usually avoids the bigger expense, which is building the wrong thing well.

Can one firm do both? Yes, and plenty do it well. The test isn't what they call themselves, it's whether they'll contract the diagnosis separately and at a fixed price. A provider who only knows how to sell build work will always push to skip that phase.

What if I already hired an agency and the project has stalled? Stop building and run the diagnosis that was skipped, even with a different provider. Restarting execution without knowing why it stalled usually multiplies the invoice without changing the outcome.

Do I need an in-house technical team to work with a consultancy? No, but you do need an internal business owner with decision-making power and allocated time. It's the one role no external provider can cover for you.

The cheap way to find out which of the two models fits you is to spend a few weeks looking at your processes and your data before signing anything large. That's what the audit is: a bounded diagnosis, with a date and prioritised decisions at the end, not a report for the drawer. And if you'd rather start with a conversation that has no proposal behind it, let's talk — we'll tell you straight whether your case calls for consulting or simply for someone to build well the piece you already have clear.