What does an AI consultant actually cost?

Start with the uncomfortable finding

I had this researched properly before writing it, and the first result was not a price. It was this: there is no independent research on what AI consulting costs. No academic study, no government dataset, no trade association rate survey.

Every range circulating online traces back to one of two places: consultancies publishing their own rate cards, or content-marketing posts with unsourced numbers that cite each other in a loop. Several of the "authoritative" pricing guides cite Gartner and McKinsey reports that contain no pricing data at all.

This page is written by a consultant too, so apply the same scepticism here. What I can do is separate what is documented from what is claimed, and label which is which. That distinction runs through everything below.

The number almost nobody mentions

The JPMorganChase Institute analysed de-identified transaction data from 4.6 million small businesses between 2019 and 2025. Not a survey. Actual money leaving actual business accounts.

The 2025 median small business spend on AI software was about $28 a month. The 25th percentile was around $20, the 75th percentile around $90. Entry costs have fallen, not risen: businesses adopting in 2019 started at around $50 a month, those adopting in 2024 at around $20.

Set that beside a published $10,000 readiness audit and you have the honest shape of this market. The typical small business is spending less than a phone bill on AI tools, while the consulting layer is priced in five figures.

Both numbers are real, and they are measuring different things. The JPMC figure is software subscriptions on business cards, not services. But the comparison matters, because if you are a small business owner feeling like the quotes you are getting are wildly out of proportion to what you currently spend on this, you are not imagining it.

The useful conclusion is not that consultants are overpriced. It is that you should exhaust the $28-a-month option before you buy the $15,000 one. Most people have not.

Two different jobs get called the same thing

"AI consultant" covers advising and building, and they are priced completely differently. Most pricing pages blend them, which is how the ranges end up so wide.

You can see both inside a single published document. Accenture Federal Services publishes a loaded labor rate card, which is an actual contractual document rather than a marketing claim:

The advisor costs roughly twice the person who builds the thing. That ratio holds across the market: advisory work runs $150 to $500 an hour, engineering labour runs $50 to $125 even in North America.

This matters commercially. If a quote is mostly advisory hours, you are buying opinions. If it is mostly build hours, you are buying a system. Ask which, and ask for the split.

What firms actually publish

Most consultancies will not publish a price. A few do, and those are worth more than any claimed range because you can go and check them.

For build work, the most granular current published breakdown puts an LLM-powered feature at $10,000 to $40,000 over three to eight weeks, a retrieval assistant over your own documents at $25,000 to $120,000, and a tool-using agent at $40,000 to $150,000. Those come from a development firm, so read them as that firm's prices rather than market data.

One line item that is routinely missing from quotes: ongoing maintenance typically runs 15 to 25 percent of the build cost per year. If nobody has mentioned that to you, ask.

What it should cost you

Pulling the above into what a small or mid-size business should actually expect to pay.

The missing middle

Put the numbers next to each other and something obvious falls out.

At one end, business-tier accounts cost roughly $20 to $90 a month and the median small business is spending about $28. At the other, the cheapest published professional help is a $7,500 audit or a $2,000 monthly retainer. Between a $28 subscription and a $2,000 retainer there is a seventy-fold jump and very little on offer.

That gap is not a pricing accident. Consultancies price for the engagements that are worth their overhead, which means small businesses either do it alone or do not do it. The U.S. Chamber of Commerce found 58% of small businesses now using generative AI, and the JPMorganChase data shows most of them spending less than a phone bill on it. An enormous number of businesses are in the "figuring it out alone" category, and that shows in the results they get.

If you are in that gap, the useful question is not "can I afford a consultant." It is "what would it be worth to have someone competent turn up once a month and keep us pointed at the right thing." For some businesses that is worth nothing, because they are already doing fine alone. For most, the honest answer is that a few hundred a month would change the trajectory more than another subscription will.

How long until it pays back

This is where the marketing and the evidence part company.

Deloitte surveyed 1,854 senior executives across 14 countries in October 2025. The finding: most respondents reported achieving satisfactory return on a typical AI use case within two to four years. Only six percent reported payback in under a year. Even among the most successful projects, just 13 percent saw returns within twelve months. For comparison, the same report notes that technology investments are typically expected to pay back in seven to twelve months.

Deloitte sells AI consulting. A finding this unflattering from a firm with every reason to publish the opposite is more credible, not less.

So when a consultant promises payback in three to six months, they are promising something that six percent of the market achieves. That does not make it impossible. It makes it a claim worth interrogating, and worth writing into the contract if they are confident.

There is a genuine counterweight, and I will flag my interest in it because it happens to favour me: MIT's NANDA working paper found that external partnerships saw roughly twice the success rate of internal builds, and that mid-market companies moved from pilot to full implementation in about 90 days while enterprises took nine months or more. Being small is a structural advantage here. Just note who benefits from me telling you that.

Statistics you will see quoted that are wrong

If you research this topic for an afternoon you will meet all of these. Several are repeated by people selling to you.

• "MIT found 95% of AI pilots fail." The report says 95% of organizations are getting zero return. The pilot funnel is a separate finding entirely: 60% evaluated tools, 20% reached pilot, 5% reached production. The two got welded together in press coverage. It is also a version 0.1 preliminary working paper based on 153 survey responses collected at conferences, not a peer-reviewed MIT study, and its original URL is now dead.
• "AI returns $3.70 for every $1 spent." This is an IDC study commissioned and paid for by Microsoft, and the returns are self-reported by survey respondents rather than measured. It is also flatly contradicted by the Deloitte payback data above.
• "Gartner says 30% of GenAI projects were abandoned." That was a prediction made in July 2024 about the end of 2025, routinely reported in the past tense as a measured outcome. If you want a measured figure, S&P Global found 42% of businesses scrapping most of their AI initiatives in early 2025, up from 17% the year before.
• "96% of organizations say their data isn't ready for AI," attributed to Gartner. I could not find any Gartner release containing it. Gartner's actual published figure is 63% who do not have, or are unsure they have, the right data management practices for AI.
• "The average small business spends $2,400 a year on AI." Attributed vaguely to industry survey data and contradicted by the JPMorganChase transaction analysis, which puts the median nearer $336 a year. The circulating figure is roughly seven times the measured one.

None of this means AI does not work. It means the numbers in this market are unusually unreliable, and a consultant quoting them at you without a source is telling you something about their rigour.

What actually drives your quote

These are the drivers with real evidence behind them, not vendor opinion.

• The state of your data. This is the big one. Gartner projects that organizations will abandon 60% of AI projects not supported by AI-ready data, and found 63% unsure whether their data management is adequate. If your pricing, past work and templates are scattered and inconsistent, structuring them is a real project before the AI project starts, and it will be the largest line item.
• Integration into how you actually work. The most-cited cause of failure is brittle workflows and misalignment with daily operations, not model quality. Standalone tools are cheap. Things that fit into an existing process cost more and are the ones that survive.
• Unstructured material. Documents, photos, scanned paperwork. Expensive to make usable, and frequently discovered mid-project rather than scoped up front.
• Compliance obligations. Published audit tiers show the premium directly: narrow small-business scope at $2,000 to $8,000 against enterprise governance scope at $15,000 to $50,000 or more, for the same category of deliverable.
• Custom models. Gartner's 2026 assessment is that most organizations attempting custom model development will abandon it over cost, complexity and technical debt. For a small business this is almost always the wrong road.
• Running cost, which is not the build cost. Infrastructure, inference and maintenance continue after launch, and model choice alone can swing monthly inference cost by a factor of eighty.

Things that should not move the price much: which model is used, how many buzzwords the proposal contains, or how good the demo looked.

What a $45,000 quote includes

A client of mine was recently quoted $45,000 to $60,000 for a single proposal-generation system. That number is not automatically outrageous. It depends on what is inside it.

Defensible if it covers discovery and process mapping, structuring years of historical data, integrations into systems you already run, the build, permissions and access control, testing against real past work, deployment, training, documentation and a support period. For reference, Gartner puts genuinely transformational generative AI initiatives at $5 million to $20 million, so $50,000 for a scoped workflow is not in that territory at all.

Not defensible if what arrives is a large prompt, a configuration of someone else's tool, a demo that only runs on the consultant's laptop, or a strategy document restating your own discovery call.

The question is not "why does this cost $50,000." It is: how much of the data, integration, security and testing problem does this solve, and what do I own at the end? Ask for it broken into those lines. A good consultant has them ready. A bad one resists, and the resistance is your answer.

When you should pay more

• The system touches money or client commitments, where being wrong is expensive.
• You are regulated, and the documentation genuinely matters.
• Many people will depend on it, and silent failure would be worse than it never existing.
• The work includes cleaning up your data properly. That is what makes everything afterwards possible, and skimping there is a false economy.

When you should hire nobody

Sometimes the right spend is zero, and a consultant who never says so is selling rather than advising.

• You have not properly tried the $28-a-month option. Buy business accounts, give someone a few focused hours, and see how far that gets you. Do this before any paid engagement.
• The task happens monthly, not daily. Automating twelve occurrences a year rarely pays back, especially against a two-to-four year payback norm.
• The real problem is a process disagreement. If three people describe your quoting process differently, AI will automate the disagreement. Fix that first, for free.
• Your data genuinely is not there. No record of past work means nothing to learn from. Start recording, revisit in a year.
• Nobody internally wants it. A system nobody asked for becomes a system nobody opens.

How I price, and why it is this low

Since the page is about transparency, here is mine rather than a gesture at "it depends."

• The first call is free. Thirty minutes, no prep, no system access. You leave with the two or three places AI would pay for itself, and that list is yours whether or not you continue.
• A $750 Quickstart, one off, credited in full against your first month if you carry on. Ninety minutes on your real work with your team driving. If I cannot find five hours a week worth recovering, you do not pay for it.
• Then $1,000 a month, or $1,900 for a weekly session and a bigger team. Two sessions a month, email support for anyone on your team, and goals checked at thirty, sixty and ninety days.
• I am teaching, not taking over. The work happens on your keyboards. If your team still needs me in a year as much as they did in month one, I have failed.
• Month to month. No annual contract, cancel any time. If a month was not worth paying for, you should not be charged for the next one.
• You own everything we set up. Accounts in your name, your data, your systems. If we stop, nothing stops working.
• Builds are quoted separately, and only when something genuinely needs building. Most businesses get a long way before that point, and I will tell you when you are not there yet.

That sits below the published retainer floor on purpose. The reasoning is in the numbers above: if the median small business is spending $28 a month and the next rung up is $2,000, almost nobody takes that step. Most of them do not need a $2,000 engagement. They need someone competent turning up once a month who they can email in between.

The honest trade-off is scope. At $1,000 you get a few focused hours a month from one person who teaches, not a team who builds. If you need somebody embedded two days a week, that is a real need and I am the wrong answer to it.

The calculation to run first

Before any of this, work out what the problem costs you now. Six hours a week on proposals is about three hundred hours a year. Put your own number on an hour and you have the figure that decides whether $12,000 is expensive or obviously worth it.

If you have not done that calculation yet, the one week retyping audit is how to get the number without spending anything.

Then hold it against a two-to-four year payback rather than the three months you were promised. If it still works on those terms, it is a real project. If it only works on optimistic assumptions, that is useful to know before you sign rather than after.

Sources

Split by how much weight each deserves. The first group is independent or documentary. The second is published by firms with a commercial interest, including, to be clear, this page.

Independent or documentary:

• JPMorganChase Institute, Understanding the use of AI among small businesses, April 2026. Transaction data from 4.6 million firms. Source of the $28 a month median.
• Deloitte, AI ROI: the paradox of rising investment and elusive returns, October 2025. 1,854 senior executives. Source of the payback figures.
• Accenture Federal Services, published loaded labor rate card. A contractual document, source of the hourly figures.
• Gartner, Lack of AI-ready data puts AI projects at risk, February 2025.
• The Register, reporting Gartner's 2026 generative AI hype cycle, May 2026. Budget overruns and custom model warning.
• CIO Dive, reporting S&P Global Market Intelligence, March 2025. The measured 42% abandonment figure.
• U.S. Chamber of Commerce, Empowering Small Business, August 2025. Small business AI adoption.

Published by firms with a commercial interest:

• MLDeep and Aries Consulting Group, published audit fees.
• Iternal, published fractional AI leadership tiers.
• Uvik, project-type build ranges and the maintenance percentage.
• BCG, The Widening AI Value Gap, September 2025. Self-reported, as BCG itself notes.
• MIT NANDA, The GenAI Divide, July 2025. A preliminary working paper with contested methodology whose original link is dead. Treated here with those caveats attached.