AI consulting for a small business: what it costs, and how to tell if you need it

What AI consulting costs a small business in 2026

Ranges are practitioner-reported: what actually got invoiced, compiled in our AI agent cost guide. Vendor marketing quotes trend higher. The timeline column is ours, taken from the scopes we run and quote, because no survey data on delivery time exists in this market and we would rather say that than dress an estimate up as research.

Two of those rows decide your total, and only one of them is on the proposal. Every bill has a build line and a run line, and the run line is the one that ends programmes.

What does an AI consultant for a small business actually do?

Three things, in this order: work out which of your processes an AI can actually do, build one of them, and get your people to use it. A good one spends the first engagement narrowing rather than expanding, because the failure mode at this company size is a rollout that touches everything and lands nowhere.

In practice the work looks unglamorous. Somebody sits with the person who does the job today and writes down what they actually do, which is usually different from what the SOP says. Somebody decides which of two contradicting documents is right and gives that area an owner. Somebody works out where the data lives and whether the systems involved have APIs that can be called without a screen-scraper held together with hope. Only after that does anyone write a prompt.

That preparation is the part people underestimate hardest, and we have the number for it. Not more than 25% of the clients who come to us arrive with their knowledge in a state you can point retrieval at. Three in four need remediation first. When we built company second brains, the real answers were almost never in the wiki: they lived in Slack and WhatsApp threads, in two or three senior people's heads, and in documents that flatly contradicted each other.

The deliverables worth paying for are boring and specific: a ranked list of candidate workflows with a reason for the ranking, a build-or-don't verdict on each, an architecture that names the systems and the model per step, acceptance criteria you can test, and a runbook your side can follow without the consultant in the room. If what arrives is a slide deck about the AI opportunity in your industry, you bought the wrong thing.

A consultant who is good at this will spend most of your first call trying to make the project smaller.

How much does an AI consultant cost, and why do quotes vary so much?

The table above is the short answer: $499 to $15,000 covers almost everything a company under 50 people should be buying. The longer answer is that buyers routinely report quotes from $3,000 to $32,000 against near-identical scope documents, and that spread is real rather than padding.

What the premium end is actually pricing: mistakes already made on somebody else's project, a written change process for work that falls outside scope, and a firm that still answers the phone in month four. The cheap end prices in none of that. Which is fine, right up until the moment something breaks and you discover the agent lives in a repository you cannot access, on API keys you do not own.

Hourly rates are a trap at this size. An hourly rate prices the consultant's time rather than your outcome, so the risk of the work running long sits entirely with you, and AI work runs long far more often than it runs short. For a sense of the market, LeewayHertz publishes a $50–$99 hourly band on its Clutch profile alongside a $10,000 project minimum, checked in August 2026. We publish no hourly rate at all. Fixed fee against a written scope, or do not sign.

One more reason quotes look strange to a 30-person company: most of the firms with the best marketing are not trying to sell to you. Published minimums on Clutch profiles in August 2026 ran from $1,000 (Axe Automation) through $10,000 (LeewayHertz) and $25,000 (Morningside AI) to $50,000 (Markovate). A minimum is how an agency declines your business without saying so.

If the price is not on the website, the number being hidden is usually the minimum, and the minimum is usually bigger than your whole project.

Is AI consulting worth it if you are under $5M in revenue?

It can be, and revenue is the wrong variable to decide it on. What decides it is whether one role in your company does enough repeatable digital work that clawing back five to ten hours a week out of it changes something. If it does, the maths works at almost any revenue. If it does not, a bigger budget will only buy you a more expensive disappointment.

Small companies have one real advantage here. You can decide on Tuesday and be running by the following month, because there is no procurement committee, no security review queue and no change-management workstream. The disadvantage is just as real: you have no slack. There is no analyst who can absorb three weeks of documenting a process, and no ops team to babysit an agent that starts behaving oddly in month three. That is the actual constraint at this size, not budget, and it is why we push so hard on one workflow with one named owner rather than anything company-wide.

Start with the cheapest thing that produces a decision. A written verdict on whether to build at all costs $500–$3,000 across the market and regularly talks the buyer out of a five-figure project. That is exactly why so few firms sell one.

Buy the decision before you buy the build. If the decision comes back "do not build", you just saved ten times what you spent.

AI consulting, AI automation consulting, AI implementation consulting: what is the difference?

Three labels, three different scopes, and the gaps between them are where small projects get lost. AI consulting decides what to do. AI automation consulting connects systems so a fixed process runs without a person. AI implementation consulting owns getting a specific thing into production and used.

For a company under 50 people, buy strategy and implementation from the same firm. The handoff between a strategy vendor and a build vendor is where small projects go to die: the roadmap assumes systems the builder finds have no usable API, nobody owns the discrepancy, and four weeks disappear into arguing about whose scope it was. Larger companies can afford that separation as a governance control. You cannot.

There is a fourth label, agent consulting, which mostly means implementation consulting where the thing being built makes decisions rather than following a fixed path. Our agents vs workflows vs RPA guide walks that distinction properly, and the punchline is that most jobs people bring us want a workflow.

One firm, from verdict to production, or you will pay twice for the same misunderstanding.

What is an AI implementation consultant?

Someone who owns the stretch between the recommendation and the thing running in production, and who is still accountable after handover. The work is tool selection, connecting the systems, writing or configuring the agent, defining acceptance criteria you can test against, and getting your people to actually use it.

The distinction that matters commercially is where the engagement ends. A strategy consultant is finished when the report is delivered, so their incentive is a persuasive document. An implementation consultant is finished when the thing works and somebody on your side owns it, which is why every question worth asking one is about ownership, warranty and month four rather than about methodology.

Job titles are unreliable here, so ignore them and ask two things instead. What is the last thing you personally put into production, and what broke afterwards? Anyone who implements for a living answers both in specifics inside thirty seconds.

If the engagement ends when the document is delivered, you did not buy implementation. You bought a document.

The two-question test to run before you hire anyone

Pick one role. Ask two questions about it. Does this role have clearly determined KPIs, and does it have clearly determined SOPs? If both are yes and the work is mostly digital, an AI agent can very likely do that work better, or at least take a real part of it off the person doing it now. If either is no, there is no honest quote available yet.

This is the strongest buy signal we know of, and it beats headcount and budget as a predictor. The logic is not complicated. KPIs tell you what good output looks like, which is the only way to grade an agent. SOPs tell you what the steps are, which is the only thing an agent can follow. Take either away and the build becomes a discovery project wearing a build's price tag.

Run it yourself in about five minutes. Open the job description for one role. Write down the three numbers you would use to tell whether that person had a good month. Then find the document that tells a new hire how to do the main task, and read it — not skim it, read it. Now ask whether a competent stranger could do the job from those two artefacts alone. That is roughly the position the agent is in.

The most common result is that the SOP does not exist, or exists and is two years out of date. That is not a reason to stop. It is a reason to make the first invoice a cheap one for writing the thing down, because an SOP good enough to hand a new employee is an SOP an agent can run. If a consultant quotes you a build price before asking either question, the number came from a template rather than from your business.

No written SOP, no honest quote. Everything else on a first call is decoration.

If you would rather have this scored than judged by eye, our AI readiness scorecard runs the same check across documentation, data access and ownership, free and without an email gate.

What the bill looks like after the build is finished

Three lines, every month, and none of them appear on most proposals: model and API usage ($50–$500 for typical small-business workloads), infrastructure ($20–$200), and maintenance ($500–$1,500 with a boutique, $2,500–$8,000 with a full-service agency). Across eighteen months the run cost routinely exceeds the build cost, and almost nobody models it before signing.

Here is the version of that going badly, and it is ours. A 20–50-person client decided every employee should have a personal always-on agent with its own token budget. Spend reached roughly $3,000–$5,000 a month. They abandoned the whole programme inside about two months.

Two causes, both avoidable. The first was idle spend: always-on agents burning tokens on heartbeats, polling, memory refresh and cron loops while nobody was asking them anything. An agent that is awake is an agent that is billing. The second was the flat rollout — everyone got one, few used one, and the company paid for the gap between those two facts every single day.

What we do now instead: shared departmental agents before personal ones, a budget cap per role rather than per person, idle loops killed outright, and cheap tasks routed to cheap models. That last one is not a rounding error. On Anthropic's published pricing, checked 25 August 2026, Claude Haiku 4.5 costs $1 per million input tokens against Claude Opus 5 at $5 — five times the price for classification, extraction and routing steps that do not need the larger model. The full arithmetic is in our token cost guide.

So do not ask a consultant what it costs to run, which invites a comfortable average. Ask what a bad month looked like on their last build and what caused it. Ask whether there is a hard spend cap outside the framework, what happens when it is hit, and who on your side can press the kill switch without phoning them first. Somebody who has run agents in production has a story and a number. Somebody who has shipped demos will tell you it depends.

A proposal with one number on it is a proposal missing the number that ends programmes.

How long does it take?

A readiness audit takes two weeks. A first agent deployment takes two to four weeks from a signed scope, provided the SOP already exists. A company-wide rollout is a couple of months of back-and-forth, and anyone promising otherwise has not run one.

The rollout shape is predictable enough that we plan against it. The founder and exec team go first, always, because a tool the leadership does not use is a tool nobody below them will touch. The first department is onboarded two to four weeks after that. Then the engagement settles into the unglamorous middle: building skills per department, choosing or building the connectors and APIs, training people, and restructuring the agent hierarchy as real usage reveals what actually matters.

At three months, roughly 30% of staff are active users. We publish that number deliberately, because it sounds like a failure and it is not. It is the honest target, and treating it as one changes how you scope the project — you stop paying for a company-wide launch and start paying for the two departments that will actually use the thing.

Adoption follows specificity, which is the single most useful thing we learned doing this. Nobody touched the generic assistant at one client until we built skills per department. The same tool, the same people, and usage only moved when the thing could do a job somebody recognised as theirs.

About 30% of staff actively using agents at three months is a win. Anyone forecasting full adoption in a month is forecasting from a brochure.

What a small-business engagement actually looked like

The best outcome we have produced for a single small business belongs to a 65-year-old lawyer in Minnesota, running a small practice, with no technical background whatsoever. Of every founder we have set up with a personal team of agents, he got the most out of it.

He runs three to five agents. One covers intake and client communications. One handles drafting and document review. One does the billing and admin backoffice. He was self-sufficient in about two weeks, and he saves five to ten hours a week.

Why he won is the part worth copying, and it has nothing to do with technology. He had deep domain expertise, so he could tell immediately when a draft was wrong. And he already knew how to manage a team of people: delegate a task, set expectations, review the work, send it back when it is not right. That is exactly the skill managing agents requires. Being technical is not the requirement. Knowing what good work looks like is.

One thing we are not going to do here is give you a payback figure. We have not measured one on an engagement of this shape, and the worked examples you will find on competing pages — a precise fee, a precise week count, a precise payback period, no client named — are almost always constructed. Five to ten hours a week for a partner-level lawyer is a number you can price yourself against your own billing rate, and it is the honest version of the same claim.

The best predictor of whether agents will work for you is whether you already know how to manage people, not whether you know how to code.

For what it is worth on fit: the verticals where we have three or more deployments are marketing and dev agencies, SaaS and startups, law and professional services, and independent local business owners — dentists and cleaning-company owners turn out to be heavy users, mostly on scheduling, crew and staff ops, reviews and marketing, and lead follow-up.

When you should not hire an AI consultant

If your entire AI budget for the year is under $2,000, do not hire one. Put it into an n8n, Make or Zapier subscription and give one motivated person a few Friday afternoons. You will get further than a small consulting engagement would take you, and we say this on discovery calls most weeks.

The reason is structural rather than generous. A small engagement is the worst version of this work: big enough to cost real money, too small to change how anything runs. Nobody wins that project, including us.

Four more situations where the answer is no, or not yet:

• The process is not repeatable yet. If every instance of the work is genuinely different, there is nothing to encode. Consulting cannot fix that; only running the process a few dozen more times and noticing the pattern can.
• The work is one fixed path with clean APIs on both ends. Build it yourself. Roughly $20–$50 a month in tooling and a few working days. Paying somebody to build that same no-code workflow is a $1,500–$3,500 job, worth it only if nobody internally will own it.
• Nobody will own the agent after handover. An agent without a named human manager degrades quietly. It does not crash; it just starts being wrong, and nobody notices until a customer does.
• You need SOC 2 or HIPAA evidence, procurement sign-off and an audit trail. Hire a larger firm than us. The firms publishing $25,000–$50,000 minimums with certifications on their directory profiles are the right call, and we are not.

Our build vs hire vs agency tool runs this decision in a few minutes and will happily tell you to do it yourself.

Ask any consultant you are considering to argue you out of hiring them, and see whether they can produce anything specific.

What percentage of AI projects fail, and what does failure look like?

Nobody credibly knows, and the confident numbers being quoted at you are weaker than they sound. The claim circulating hardest in 2026 is the MIT/NANDA figure that around 95% of enterprise AI pilots return nothing. We do not rely on it: the base is small, it has not been peer-reviewed, and it counts a pilot that never reached full production as a failure, which describes plenty of pilots that were never meant to. Gartner's widely repeated 30% and 40% abandonment figures are predictions, not measurements.

What we can tell you is the shape of the failures we get called in to fix, because rescue work is a surprising share of what we do. The model is almost never the problem. The three real causes, in the order we see them: no owner, no budget cap, and no written process underneath.

The token blowup above is the budget-cap version. The knowledge version is the 25% number: three in four companies that come to us cannot point retrieval at their own documentation, and a project built on that foundation produces answers that are confidently wrong, which is worse than no answer at all. The ownership version is the quietest — an agent nobody manages does not fail loudly, it drifts, and you find out in month four.

None of those are AI problems. They are project problems that AI happens to make expensive.

Ask a consultant which of their own builds failed and what it cost. Anyone who has shipped more than a handful has one, and the ones who claim otherwise are the worrying answer.

How do you choose between AI consultants?

Start from your constraint rather than from a list of firms. Regulated and audit-heavy means a bigger firm with certifications. A fixed repeatable path means no firm at all. Systems without clean APIs, judgment in the loop, permissions split by department, or a previous build that already failed — that is where a boutique earns its fee.

Then run the same five questions past everyone you shortlist. What will you refuse to build for me? What does this cost to run each month, as a line separate from the build? Who owns the code, the credentials and the prompts on the day we finish? What happens when it breaks at 2am in month four, and who is contractually obliged to fix it? And what number would you call a failure at three months? Our full red-flag and green-flag guide has the answers that should worry you, including the three tests we fail ourselves.

Watch what they ask you, too. A consultant who arrives already knowing your problems — deck named after your industry, case study for a company nothing like yours — is selling a product. If nobody has asked you a question you found slightly uncomfortable, you have not been scoped.

And check the things you can verify without them: published minimums and hourly bands on directory profiles, real reviews from companies your size, whether the partner badges they display exist in the vendor's own public directory. A badge on a homepage is a picture. A listing in n8n's partner directory is a fact.

Shortlist on what you can verify at the source, then decide on who tried hardest to make your project smaller.

Our guarantee, and exactly what it does not cover

Every agent we build carries a 90-day warranty from the day we hand it over. Defects in what we built get fixed at no charge, and we fix first rather than negotiate. The fix has a hard cap: if an accepted defect is still broken 15 business days after you report it, you get that agent's build fee back.

The warranty is not tied to a retainer. Ongoing maintenance stays optional and priced separately, and the guarantee stands whether or not you buy it. We are spelling that out because tying a warranty to a monthly fee is the standard way this industry converts a promise into a subscription.

What it does not cover, in plain terms: your process changing, model-provider price or behaviour changes, the token and API spend on your own accounts, edits your team makes after handover, and answers that come out wrong because two of your source documents disagree with each other. That last exclusion is the one that matters most in practice, and it is why we would rather sell you the $499 readiness audit first and find out.

A warranty with a stated cap and a published exclusion list is a commitment. "We stand behind our work" is a sentence.

Where we are the wrong choice

Three honest gaps. We have no reviews on Clutch or any other third-party directory, so every claim on this page is published by us and there is nowhere independent to check it, which is precisely the thing this guide tells you to be suspicious of. Our published case studies are software product builds for technology companies rather than internal agent rollouts at 20–50-person firms, so they prove we can build rather than proving we have done your exact job. And we are eight people, with the founder writing these guides and running the discovery calls, so the bus-factor question is fair.

Our only real answer to the last one is the ownership rule we apply to every build: your repository, your API keys, prompts in files you can open, and a runbook your side tests before the final payment clears.

Frequently asked questions

Sources

• Cognio Labs, how much does an AI agent cost — the market ranges in the price table, compiled from practitioner-reported invoices and firms' own published directory minimums.
• Anthropic, model pricing (checked 25 August 2026) — the per-token spread between Haiku 4.5 and Opus 5 quoted in the run-cost section.
• Clutch profiles for Axe Automation, LeewayHertz, Morningside AI and Markovate, read August 2026 — the published project minimums and the $50–$99 hourly band quoted above. Our comparison of AI agent development companies has the full table, including the columns where we lose.
• The MIT/NANDA "95% of pilots" figure and Gartner's 30%/40% abandonment predictions are cited here as claims in circulation, not as evidence we rely on. Our reasons are set out in the methodology section of the token cost guide.
• Cognio Labs deployment notes, August 2026 — the lawyer engagement, the token blowup, the rollout shape and three-month adoption figure, and the "not more than 25% arrive knowledge-ready" observation are our own client observations, anonymised and published with permission.

Related reading

• How much does an AI agent cost — every price band in the table above, with the sourcing for each.
• Questions to ask before you hire an AI agency — red flags, green flags, and the three tests we fail ourselves.
• AI readiness scorecard — scores documentation, data access and ownership in a few minutes. Free, no email gate.
• Build vs hire vs agency — the same decision as a tool, including the answer where you do it yourself.
• AI consulting — our engagements, what each one delivers, and what they cost.
• AI agent development — how we scope, build and hand over, including what we refuse to build.