22 of the 81 agencies in this directory that sell AI strategy work list nothing at all that builds software. They will assess your readiness, run the prioritisation workshop and hand over a roadmap, and then the roadmap needs somebody else. That is a legitimate business and it has been a legitimate business since long before AI, but it is not what most buyers think they are purchasing when a proposal says ai consulting and implementation. The two halves of that phrase are frequently sold by two different companies, and the seam between them is where AI programmes go to die.
The counts behind this are from the live directory on 9 October 2026 and anyone can check them on the category pages. Of 243 real listings, 81 carry the AI strategy consulting category. 167 carry at least one category that implies building something: AI integration, custom AI solutions or process automation. Only 59 carry both. So 22 firms advise without building, 108 build without advising, and fewer than a quarter of the market publicly claims to do the whole job.
The seam nobody owns
A strategy engagement ends with a document. An implementation engagement starts with a specification. Those are not the same artefact, and converting one into the other is real work that neither contract usually pays for. The roadmap says "deploy a document classification model to reduce manual review." The build team needs to know which documents, what accuracy threshold counts as done, who labels the training data, where the model runs, what happens to the 4 percent it gets wrong, and who is accountable when a human overrides it.
Somebody has to answer those questions. In a single-vendor engagement, the firm that wrote the roadmap answers them, and it answers them knowing it will have to live with the answers. In a two-vendor engagement, the advisory firm has gone home and the build firm is reading a document written by someone it never spoke to. It will either guess, or it will run its own discovery and charge for it, which means the buyer pays for the same decisions twice.
The second outcome is the common one. It is also the least visible, because it arrives labelled as the build firm's onboarding phase rather than as a failure of the handoff. A single contract for ai consulting and implementation does not remove that work, it just stops you paying for it twice.
Should one firm do both the consulting and the implementation?
Usually, for anything under about six months, yes. The cost of the seam is higher than the cost of the conflict of interest, and the conflict of interest is the standard argument for splitting them. That argument is not wrong: a firm that recommends work it will then be paid to perform has an obvious reason to recommend more of it. The question is which risk is larger in your specific case. For a first production system, the risk of a roadmap nobody can build is larger. For a multi-year programme with a procurement department and real money at stake, the independence is worth paying the seam cost for, and large organisations split them for exactly that reason.
The useful middle position is to split the decision but not the relationship: buy a short, genuinely independent assessment from one firm, then run a competitive process for the build in which the assessor is allowed to bid and is given no advantage. That keeps the advice honest and still lets the firm that understands the problem carry it forward.
Four ways the phrase gets used
Proposals offering ai consulting and implementation use the same words for arrangements that behave very differently once the work starts. These four cover most of what a buyer will be sent.
| What the proposal says | What it means in practice | Who converts advice into a spec | The thing to check |
|---|---|---|---|
| Strategy and implementation, one firm | Same team, or at least same company, from assessment to production | The firm, at its own cost | Whether the engineers who will build it attend the strategy sessions |
| Strategy with implementation support | Advisory firm oversees a build done by somebody else | Nobody, formally | Who signs off that the build matches the roadmap |
| Implementation partner | Build only, against a spec you supply | You | Whether you actually have a spec, or only a roadmap |
| End-to-end AI transformation | Anything from both of the above to neither | Unknown until you ask | Ask which named people write code |
The fourth row is the one to be careful with. "End-to-end" describes an ambition rather than a structure, and the test that separates the honest version from the brochure version is simple: ask for the names and titles of the people who will write the first production commit. A firm that builds can answer in a sentence. A firm that subcontracts will describe a process.
What the listings tell you about who does what
46 of the 243 real listings carry no category at all, which is worth stating plainly because it is a limit on what a directory can tell you. For the 197 that do, the category pattern is a reasonable first filter. A firm carrying only strategy is almost certainly advisory. A firm carrying integration, custom builds or automation and no strategy category is almost certainly a build shop that expects you to arrive with a plan. The 59 that carry both are the shortlist for a single-vendor engagement.
What the listings will not tell you is depth. 40 of the 243 publish an hourly rate and 42 publish a team size, so for most of the market you cannot see whether the implementation capability is 30 engineers or one contractor the principal calls when a project lands. The category-by-category count of this market shows how uneven the supply actually is, and the thinness is concentrated exactly where buyers assume it is not.
How do I tell an advisory firm from a build firm before the call?
Read the case studies for artefacts rather than outcomes. An advisory firm describes decisions: a prioritised portfolio, an operating model, a governance framework. A build firm describes systems: a pipeline, a latency figure, a model in production behind an API. Both will use the word "delivered." Only one will name a thing that runs. If the case studies describe only decisions, the firm is advisory regardless of what the services page says, and that is useful information rather than a criticism.
Where the money actually goes
Advisory work is priced on senior time and is front-loaded. Implementation is priced on engineering months and is where the bulk of a programme's cost sits, usually by a wide margin. A buyer who approves a strategy engagement on the basis that the number is manageable has approved roughly the smaller part of the eventual spend, and the roadmap it produces will have been written without a hard budget constraint on the part that costs more.
This is the practical reason to have the build side in the room early even when you are not buying from them yet. An engineer who knows what three months of data pipeline work costs will say so while the roadmap is still editable. The same information arriving four months later arrives as a renegotiation. The question of which engagement model you are signing matters here too, because a retainer absorbs this kind of surprise silently while a fixed-scope contract forces it into the open.
The governance side has standard references worth reading before either contract is signed. The NIST AI Risk Management Framework is organised around functions that map onto implementation decisions rather than strategy ones, and ISO/IEC 42001 covers AI management systems as something an organisation operates rather than something a consultant recommends. Both are useful precisely because they assume somebody will have to run the result.
What should be in the contract at the handoff?
Three things are routinely missing from a contract that covers both ai consulting and implementation, and all three cost money later rather than now.
- A definition of done a third party could test. That means a number and a method, not "improved accuracy." If an outside engineer could not run the test and get a yes or a no, the clause is decoration.
- Named ownership of the training data and the labelling. This is the single most common cause of an implementation stalling in week three, and it stalls because both parties assumed the other had it.
- A list of what the buyer owns at the end. The model weights, the pipeline code, the prompt templates, the evaluation suite. Name them individually, because "the deliverables" has been read both ways in front of lawyers.
The question of what you actually own when an engagement ends is far less settled in AI work than in conventional software, and a handoff between two vendors is the worst possible place to discover that.
The change management half that gets cut first
An implementation that works technically and is not used is the most expensive outcome available, and it is common enough that the long-tail queries reaching this site ask about it directly. Somebody has to retrain the people whose job the system changes, rewrite the procedure it replaces, and own the decision about what happens when the model and a human disagree. Advisory firms will scope this work. Build firms generally will not, because it is not engineering. It therefore falls into the same seam, and it is the first line cut when a programme runs over.
The MIT Sloan Management Review has published extensively on why organisational adoption rather than model quality decides whether these programmes return anything, and the Project Management Institute maintains the standard body of practice on the discipline itself. Neither is AI-specific, which is rather the point: the part of an AI programme most likely to fail is the part that is not about AI.
Do I need a separate change management vendor?
Rarely, for a single system. The work is real but it is small enough to belong to a named person inside your own organisation, and an internal owner has standing that an external one does not. For a programme touching several departments at once it becomes a genuine workstream, and at that scale the advisory firm is usually the right place to buy it, because it already understands the operating model it is changing.
How the shortlist should be built
Work backwards from which half you are short of, because that decides whether you are shopping for ai consulting and implementation together or for one of them alone.
- A clear view of the problem, no engineering capacity. You want the 108 firms that build and do not advise, and you want a written specification before the first call.
- Engineering capacity, no internal agreement on what to build. You want the advisory side. Buying a build now is premature and the build team will tell you so in month two.
- Neither. The 59 firms carrying both categories are the only group worth a first call, and a single vendor is the right structure for a first system.
Geography matters less than it used to but not zero. The directory covers 125 distinct cities, and for an implementation involving on-site data access or regulated systems, time zone and travel are real constraints. Browsing by city is a reasonable filter at that point: Denver, Chicago and Boston each hold a different mix of advisory and build capability. The cases where location genuinely matters are narrower than most buyers assume.
What is the single best question to ask on a first call?
Ask the firm to describe the last engagement where its own recommendation turned out to be wrong, and what it did about it. An advisory firm that has never been wrong has never been in production. A build firm with a good answer will describe a specific technical bet that failed and the cost of unwinding it, which tells you both that it ships and that it is honest about what shipping costs. A firm that cannot produce the story has either not done the work or will not discuss it, and both are answers.
What this market looks like underneath
The split is not a flaw in the market. Advice and delivery are genuinely different businesses with different cost structures, different margins and different hiring, and firms specialise for sound commercial reasons. What is a flaw is that the language has collapsed: "ai consulting and implementation" is written on the website of firms that do only one of the two, and buyers reasonably read it as a promise about scope.
So treat the phrase as a question rather than a description. The firms that do both will welcome it, because it is how they differentiate. The firms that do one will tell you which one, and that answer is worth more than the brochure it contradicts. The broader question of whether you need a consultancy at all is often the right place to start, and sometimes the answer is one experienced individual and your own engineers.
For the technical half, the shape of the work has its own literature. How AI software development differs from conventional delivery is the part that most surprises buyers who have commissioned software before, and the OECD AI Policy Observatory tracks the policy and regulatory conditions that increasingly shape what an implementation is allowed to do.
Two failure patterns worth recognising early
The first is the orphaned roadmap: a competent strategy document, approved, budgeted and then never built, because the build quote arrived at three times the figure the roadmap assumed and the programme was reopened as a planning exercise with no end date. The tell is a roadmap written without an engineer in the room.
The second is the specless build: a capable engineering firm six weeks into work against a brief that never defined done, discovering that the stakeholders disagree about what the system is for. The tell is a kickoff meeting in which the buyer's own people describe different outcomes. Neither pattern is caused by a bad vendor, and that is why buying a better vendor does not fix either one.
Starting from where you are
If the roadmap already exists and nothing has been built, the gap is a specification and a build partner, in that order. If there is no roadmap and a budget is approved, the gap is a short assessment with the build side present. If a vendor has sent a proposal using the phrase this article is about, the useful next step is to ask which of the four rows in the table above it describes, and to get the answer in writing before the first invoice.
You can tell us what you are trying to build and see the firms in this directory whose published categories match both halves of the job, or if you run an agency that genuinely does both, list it here so buyers filtering for that combination can find you. Browsing what a custom build actually consists of and what integration work involves is a reasonable way to sharpen the specification before any of it costs money.