Can a consultant actually implement AI?

Some can. Plenty cannot, and the word "implementation" appears on the websites of both.

Three types of firm exist in this market and the differences matter more than anything on their homepages.

The advisory firm

Produces analysis, strategy, and a roadmap. Does not build. Frequently does not have anybody who has built.

Honest strength: no incentive to recommend a build, so the advice is structurally cleaner than ours.

Honest weakness: recommendations from people who have not implemented tend to be too tidy. They underestimate the reconciliation problem in your data, the exception rate, and the amount of work in escalation design, because those things are invisible until you are standing in them. You get a plan that is directionally right and unpriced in reality.

How to tell: ask what they built last, personally. Then ask who does the building if you proceed.

The build shop

Ships things. Knows what breaks because it has broken on them.

Strength: the work becomes real, and their estimates are grounded.

Weakness: they build what you ask for. If your specification is wrong, you get a well-made thing that solves a problem you never ranked, and the conclusion inside your company becomes that AI does not work here. Some build shops will diagnose properly. Many will accept your brief because accepting briefs is the business.

How to tell: give them a process that should not be automated and see whether they identify it. The working-session interview.

The firm that does both

Ours, and most of the useful ones at the small end of the market.

Strength: the diagnosis is informed by knowing what it costs to build, and nobody hands off between two suppliers who blame each other for the gap in the middle.

Weakness, stated plainly: an obvious conflict. A firm that builds has an incentive to recommend building, and specifically to recommend building the things it is good at.

How to tell whether the conflict is being managed: ask what they would recommend if they built nothing themselves. If the answer never includes buying something off the shelf, or fixing a process without technology, or waiting, the diagnosis is a sales instrument.

The gap that decides the outcome

Whichever type you choose, the same set of tasks sits between the plan and the working system, and neither contract usually mentions it.

Reconciling data across systems with different identifiers. Enumerating the exceptions the roadmap did not. Designing escalation and testing it. Getting credentials from platforms where only an external administrator can grant access. Defining what "working" means so somebody can sign it off.

Advisory firms call this implementation detail. Build shops call it discovery that should already exist. It lands on you in week three as a delay.

One question fixes it: who owns data cleanup, exception design, and access, and is that in writing? The scope boundary.

What implementation capability looks like

Not a portfolio. Four specifics.

They can show you something live and let you try to break it, and they are relaxed about you doing so.

They can describe their failures in detail, including the silent kind where a vendor renamed a field and output went subtly wrong for two weeks.

They raise escalation design unprompted, early, before you ask.

They quote run costs alongside build costs without being pushed, because they have paid those bills.

Any firm missing all four has not implemented anything, regardless of what the proposal says. How to verify.

The arrangement to prefer

Keep the two purchases separate with a genuine stopping point between them.

Buy the diagnosis, learn what working with these people is like, then decide about the build with actual evidence. Require the roadmap to be executable by a third party, which both preserves your option to shop the build and forces the advice to stand on its own merits.

That structure works regardless of which of the three types you engage, and it is the reason we sell it that way: fixed fee for the roadmap, build optional, roadmap yours either way. How it runs.

We build, so read our recommendations accordingly, and ask us the question in the third section. It is the one we would want a client to ask.