Gartner has warned for years that a large share of machine learning projects never reach production, and the analysts who track this have not had to revise the warning much as the technology has changed. The reason is rarely the model. It is that the buyer and the firm they hired were working from different definitions of the words on the contract. "AI and ML consulting services" is a label attached to at least four distinct businesses, priced differently, staffed differently, and carrying completely different risk. Buying the wrong one is not a bad deal. It is the wrong purchase.
I have watched this go wrong from both sides of the table. A company decides it needs machine learning, runs a procurement process, compares three proposals that all say "AI and ML consulting," and picks the one with the best day rate. Six months later there is a well-documented model sitting in a notebook that nobody can deploy, because the firm that won the work sells research and the company needed engineering. Nobody lied. The category is just too loose to buy from safely.
The four businesses hiding behind one label
Advisory. Strategy work. A firm assesses where machine learning could plausibly earn its cost inside your operation, ranks the candidates, and hands you a roadmap. The deliverable is a document and a set of decisions. It is genuinely useful when a leadership team cannot agree on where to start, and it is expensive wallpaper when they already know and are looking for permission. Advisory firms bill by the engagement, typically six figures for anything with a recognisable brand on the cover, and they do not build.
Data engineering. The unglamorous majority of the work. Before a model does anything, someone has to find the data, move it, clean it, join it, and keep it flowing. Most companies who think they need a model actually need this first, and the honest firms say so in the first meeting. Billed by the hour or by the sprint, usually $100 to $250 an hour in the United States for competent teams.
Applied model development. Building and training the thing itself, whether that means fine-tuning a foundation model, building a forecasting system, or standing up a retrieval pipeline over your own documents. This is what buyers picture when they say "AI consulting," and it is the smallest slice of the actual work. Rates run higher, $150 to $400 an hour, and the best practitioners are booked.
MLOps and production ownership. Deployment, monitoring, retraining, drift detection, cost control, and the on-call rotation for when a model starts producing nonsense at two in the morning. This is where the Gartner failure statistic actually lives. A model that works in a notebook and a model that works in production are separated by a body of engineering work that almost nobody budgets for, and the firms that specialise in it are the ones worth keeping.
A firm that does all four is either very large or overselling. When you see a ten-person shop offering the full stack, ask which one they have done more than twice.
Reading a proposal for what it leaves out
Proposals in this market are written to survive comparison, which means they are vague in the places that matter. Three omissions tell you most of what you need to know.
The first is data access. If the proposal does not say, specifically, which systems the team will need credentials to and who inside your company will grant them, the timeline is fiction. The single most common cause of a stalled machine learning engagement is a team that cannot get to the data for eleven weeks, billing the whole time. Make the access plan a named section with named owners and dates.
The second is the definition of done. "Deliver a working model" is not a definition. A definition looks like: the model runs on this schedule, against this data source, its output lands in this table, this named person gets an alert when accuracy drops below this threshold, and the runbook for retraining is checked into this repository. Firms that sell engineering will write that without being asked. Firms that sell research will negotiate it, and their negotiation is the answer to your question about what they do.
The third is what happens to the code. Ownership of the model weights, the training pipeline, the feature definitions and the infrastructure configuration should be stated in one sentence. Ambiguity here is not an oversight. A pipeline you cannot run without the consultancy is an annuity for them and a liability for you, and it is worth paying more up front to avoid it. The points to cover in an AI agency contract are worth reading before you sign anything in this category.
What the price actually reflects
Day rates in this market span an order of magnitude, and the spread is not explained by quality alone. It is mostly explained by three things: where the team sits, whether you are buying named individuals or a bench, and how much of the risk the firm has agreed to carry.
A fixed-price engagement with a defined outcome costs more per hour than a time-and-materials arrangement, because the firm is pricing in the chance that it takes longer than they think. That premium is usually worth paying for a first project with a new partner, and usually not worth paying once you know each other. A staff-augmentation arrangement where you direct the work is cheapest per hour and most expensive in management attention, which is a real cost that never appears in the comparison spreadsheet.
Be suspicious of anyone who quotes a price before asking about your data. The state of your data is the largest single variable in the cost of a machine learning project, and a firm that prices without looking at it is either guessing or planning to raise the number later. Our breakdown of what an AI agency costs goes into the ranges by engagement type.
When the honest answer is that you do not need this yet
McKinsey's annual State of AI survey has consistently found that the organisations reporting real financial impact from AI are a minority of those deploying it, and research from MIT Sloan Management Review and Boston Consulting Group has reached a similar conclusion about the gap between adoption and returns. The pattern behind those findings is not mysterious. Companies that get value from machine learning tend to have already solved a boring problem first: they know what their data means, they can get it in one place, and somebody owns it.
If you cannot produce a clean, current, agreed-upon number for something as simple as how many active customers you have, a model is not your next purchase. Data engineering is. Any consultancy that tells you otherwise is selling you the part of the project they enjoy rather than the part you need, and the six months you spend finding that out is the most expensive six months in the engagement.
The same logic applies in the other direction. If you have a clean warehouse, a defined question, and someone internally who can maintain what gets built, you may not need a consultancy at all. You may need two good contractors and a deadline. The comparison between an agency and an in-house team comes down to how often you will need this capability again.
A short diligence list that separates the four
- Ask for a production reference, not a case study. A case study describes a project. A reference is a person at another company who will get on a call and tell you whether the thing is still running.
- Ask what they would not build. A firm with real operating experience has a list of projects they turn down. A firm without one will agree to anything.
- Ask who is on the call in month four. Senior people sell. Find out whether they also deliver, and get the names in the statement of work.
- Ask about the last engagement that went badly. The answer tells you more about the firm than every slide in the deck. Anyone who claims there has not been one has not done enough work.
- Ask them to explain your own data back to you. After a discovery session, a competent team can describe where your data lives and what it means more clearly than you can. If they cannot, they were not listening.
Where to start looking
The market for these services is fragmented and most of it does not advertise. Large systems integrators are visible because they buy visibility; the specialist firms that do the best production work are frequently three referrals deep and have no marketing function at all. That is why buyers who go to Google first and stop at the first page tend to end up with a vendor chosen by advertising budget.
Browsing by discipline is a better way in than browsing by brand. The AI strategy and consulting category and the machine learning and data science category in our directory separate firms by the work they actually do, which is the distinction this whole article is about. If you would rather describe the problem once and have the shortlist come to you, get matched with vetted AI agencies and skip the comparison spreadsheet.
Background reading worth the time: Gartner on the production gap in machine learning deployment, and the MIT Sloan Management Review research with Boston Consulting Group on why adoption and financial return have stayed so far apart.
The short version
Four services share one name. Advisory tells you what to do, data engineering makes it possible, applied development builds it, and MLOps keeps it alive. Most buyers think they are purchasing the third and actually need the second, and almost nobody budgets for the fourth. Work out which one you are buying before you compare prices, because comparing a research firm's day rate against an engineering firm's is not a comparison at all.