AI Consultant, Fractional Head of AI, or AI Agency: Which One Does Your Firm Actually Need?

Somewhere in Ireland this week, a managing partner has four tabs open. One is an AI consultant. One is an automation agency. One is a fractional Chief AI Officer. One is a training provider. All four websites sound confident, and it's genuinely hard to tell from the outside what you'd actually be buying from each.

I should declare an interest before going further: I sell three of the things in this article. So I've tried to write the guide I wish my own prospects had read before they rang anyone, me included.

The market has produced a shelf of titles faster than anyone can define them, and they overlap enough that comparing them feels like comparing four brands of the same thing, and it isn't. The way out is to stop asking "which provider?" and ask a different question first:

What responsibility is currently unowned inside your firm?

Somebody deciding? Somebody building? Somebody accountable, continuously? Or does nobody yet know what the work even is? Each of those is a different purchase, and the rest of this article takes them one at a time.

When you need an AI consultant

A consultant is the right purchase when you have a defined question and you need independent judgement on it.

The questions sound like this. Where are the worthwhile AI opportunities in our firm? Is the automation someone's proposing actually worth building? Which of these three tools should we standardise on? What are we not seeing? What should we do first?

What you're buying is a better decision. It might arrive as recommendations, an assessment, or a scoped piece of work, and the engagement can end when the question is answered. That last part matters: a consulting engagement should have an actual ending, and you should be able to see it from the start.

The test of a good one is simple, and it's the test I'd want applied to me: can they recommend doing less? You should hear "leave that one alone" at least once, with the reason attached.

I've written up what an AI consultant should actually do for a firm like yours, and if your question is defined enough to fit in an hour, that's what the Orientation is for.

When you need an AI agency

An agency, or an implementation partner, is the right purchase when you already know what needs to be built, with reasonable confidence, and the missing responsibility is delivery.

That's a genuinely valuable thing to buy. Connect these two systems. Build the intake workflow. Automate the monthly reporting pack. Stand up a knowledge assistant on our own documents. If the problem is correctly scoped, a good agency will build it well, and faster than you would.

The catch sits one step earlier. An agency will build the thing you asked for, and whether you should have asked for it isn't the responsibility you've hired them to carry. Before the brief goes out, somebody in your firm should be able to answer: what problem is this solving, how often does it occur, who owns the process, what does a person still review, how will we know it's working, and who minds it after launch. If those answers don't exist, the build will still happen. It'll just be an answer to a question nobody checked.

I've watched firms pay properly for well-built things that solved rare problems. The build was fine; the mistake happened one step earlier, when nobody counted how often the problem actually came up. Agency builds commonly start around €15,000, and that's fair money for good work. It's also a lot to spend before anyone's counted. If you're not certain the thing deserves building, count the work before you brief anyone — it's a cheaper afternoon than the alternative.

When you need a fractional head of AI

This one is for a different situation altogether: not a question, not a project, but an accumulation.

There comes a point where the firm has enough AI in the air that the subject itself needs an owner. Three experiments running on individual laptops. Two vendors mid-pitch. A workflow someone automated last year that nobody's checked since. Ideas parked in partners' heads, a licence renewal coming up, and a low hum of "should we be doing more?" in every quarterly meeting. No single item justifies hiring anyone, but the pile of them does.

A consultant answers the question. A fractional head of AI owns it.

Owning it means the unglamorous continuity: deciding what deserves attention this quarter and what doesn't, sequencing the work, setting the guardrails, judging the vendors, watching what's already deployed, stopping the things that aren't earning their keep, reporting plainly to the partners, and building your own people up as it goes. The ownership is the product.

The practical outcome is smaller-sounding than the title. A new tool appears, a vendor makes a claim, an experiment starts drifting, and there's already somebody who knows the history and owns the next decision. The partners can say "we have somebody looking after that", and name them.

Done properly, it also means owning what enters the firm's AI environment, what stays, and what leaves. That's the discipline I run as the five-way sort, it's what The Seven Shifts™ exist to check, and it's the standard the Fleet is held to in my own practice: everything deployed keeps justifying its place, or it goes.

This is what I offer as fractional AI leadership: a head of AI, one day a week, for firms where the question has become real but the full-time job hasn't.

When you need a full-time head of AI

Eventually, for some firms, the question genuinely becomes a full-time job, and then a full-time person is simply the proportionate answer.

The signals are about volume and consequence rather than headcount. Several teams deploying AI at once. A real internal build programme. Client-facing AI where an error reaches the people who pay you. Governance that needs continuous attention rather than a monthly look. Vendor spend that's become a line item the board asks about. When the pile of responsibility fills a week, every week, splitting it across a fractional day stops being economical and starts being a bottleneck.

Most fifteen-to-seventy-five person professional services firms in Ireland aren't there. Some will be in a few years, and part of the fractional head's job is to say so when the time comes: the role done well builds the internal capability that eventually replaces it. But hiring full-time before the responsibility is full-time gives a good person half a job, and half a job quietly grows to fill the week.

When you need none of them yet

Sometimes the right amount of external AI help is none. What's missing at this stage is an understanding of your own work, and nobody outside the firm can invent that understanding for you. The signs are recognisable:

• Nobody can explain the problem in one sentence without the word "AI" in it.
• The use case everyone's excited about happens four times a year.
• The process is painful because it's badly designed, and automating it would deliver the same bad process faster.
• The data the idea depends on lives in five places and agrees in none of them.
• The genuinely valuable part of the work is human judgement, and everyone privately knows it.
• A tool the firm already pays for can already do the thing — it's just never been set up.

Hiring anyone at this stage buys you activity. What the firm needs first is a count: what work keeps coming back, what it takes out of the year, how much judgement it holds, and what should happen to it — automate it, assist it, fix the process, keep it human, or leave it alone for now. That's an afternoon with a whiteboard if you do it yourselves, or one facilitated day if you'd rather have someone in the room whose only job is keeping the count fair.

Sometimes the most useful advice is "don't build anything yet".

The short version

One caveat about the table: the categories overlap at the edges. Consultants sometimes build. Agencies sometimes advise well. A fractional head certainly builds. What separates them is the primary responsibility you're paying them to carry, and that's the thing to get in writing.

Pick the wrong tool and it costs you a licence. Hire the wrong kind of help and it costs you a year.

The titles matter less than they sound

If you have a defined question, buy judgement.

If you have a defined build, buy delivery.

If AI has become an ongoing management problem, give somebody ownership of it.

And if you can't yet describe the work that deserves the attention, don't hire anyone to automate it. Count the work first.

Where that leaves us, plainly: a defined question fits in an Orientation. Work that needs finding and weighing is a Census day. And if you already know the firm needs the question owned, that's the fractional arrangement, and it starts with a conversation that costs nothing.