What is AI consulting, actually?

In short: AI consulting is paid help deciding where AI can take real work off a team, then building and running the systems that do it inside the software the business already uses. The same phrase covers an $800 questionnaire run through a language model and an eighteen-month transformation program, so the useful question is what you are actually buying, and that has four honest answers.

Key takeaways

• The published definitions (advising on, designing, and implementing AI systems) are accurate, and they describe an $800 questionnaire and a six-figure program equally well. They are written from the consultant's side of the table.
• Almost none of the hours go on AI. In our engagements they go on watching how the operation actually runs, integration code into the software already in use, prompts tuned against real messages, and deciding where a person stays in the loop.
• Four different products are sold under the name: advice, tool selection and setup, a custom build, and operating what was built. They are priced, owned, and staffed differently, so a proposal should say which one it is.
• Nobody trains a model for a forty-person firm. The models are rented. The consultant's job is to land one inside a business that already exists.
• Sometimes the right answer is a product at a couple of hundred dollars a month, and a consultant worth hiring tells you that on the first call.

The definition you will find, and why it does not help

Search the phrase and page one is a glossary entry, a couple of career guides for people who want to become one, and IBM. The Hackett Group's glossary calls AI consulting a specialized advisory service that helps organizations plan, design, and implement artificial intelligence solutions. CIO's version is shorter: advising on, designing, and implementing AI solutions. Both are correct.

They are also true of two things that have nothing in common. An $800 "AI readiness report" produced by pasting your business description into a language model is advising on AI solutions. So is an eighteen-month program with a steering committee and a delivery team you meet in month six. The definition holds for both because it describes what the consultant does. From your side of the table, the only definition that matters is what exists in your operation after the invoice clears.

So here is the working definition we use, and it has an operator's bias built in. AI consulting is help with three decisions, followed by a job. The decisions: where AI can take real work off your team, what to build or buy to do it, and where a person has to stay in the loop. The job: making that thing run inside the software and the habits you already have, and keeping it running when something on the other end changes. Strategy firms tend to stop after the first decision, and software vendors would rather you arrived having already made all three. Most of what you are paying for lives in between.

What the hours actually go on

The honest way to define a job is to say what the person does all day. This is where the time goes on ours.

Watching the operation. The audit is ten to thirty hours of a senior engineer, and most of those hours are not technical. They are spent beside the leasing coordinator or the dispatcher while the work happens, because the process an owner describes in a meeting and the process the coordinator actually runs are two different processes, and only the second one can be automated. A plumbing company we work with in the GTA had no CRM and no software worth the name. Work orders arrived by email and stayed there, the schedule lived in the owner's head, and invoices were written by hand on a pad. The finding fit on one line: the bottleneck was paperwork, not plumbing. The first phase we scoped was an application that takes a work order in and an invoice out from his phone, and it will never price a job, because pricing stays his decision. That was the whole of phase one. The only model in it reads the work-order emails, and it is AI consulting all the same.

Integration code. Most of a build is plumbing between systems that were never meant to talk to each other. Leads arrive from a website form, a Zillow or Padmapper inbox, a personal Gmail thread the owner has used for years, and a phone that rings to a cell. The work is a parser that reads every one of those into one record, writes it into the CRM the team already uses, and decides which ones a human sees first. One of the highest-value things we have ever built is an email parser. It has no interface, and nobody will ever demo it at a conference. The full design is in automating lead intake.

Prompts tuned against real messages. The model is rented from a lab. What you pay a consultant for is the two weeks of running it against last month's actual tenant emails, maintenance calls, and lead forms until the exceptions stop surprising it. The handful of cases everyone at the office handles by habit and nobody mentions in the meeting are most of the build.

The human gate. Deciding where a person stays in the loop is the least glamorous decision in the engagement and the one with the most liability in it. On the voice agents we run for a Toronto property management and brokerage operation, the rules about what counts as an emergency, who gets flagged, and what the agent may never say came from the operator, and writing them down took longer than wiring the phone line. One June call was a tenant disputing a $675 towing charge. The agent's entire job on that call was to capture the complaint in full and get it to a person, which it did. In a 277-call sample that month, none were missed and 74 came in after hours.

The part after launch. Something on the other end always changes. The CRM ships a new API version, the carrier changes a webhook, the listing portal changes its email format. Operating the system means someone notices before your team does. During operate we commit to shipping something new every month, and either side can walk away on thirty days' notice, because a retainer that only survives on the strength of its contract is rent.

Four things sold under one name

When someone says they do AI consulting, they mean one of these, and occasionally all four. They are priced differently, owned differently, and done by different people, so a proposal should say which one it is.

Advice. A readiness assessment, a roadmap, a workshop. The deliverable is a document. It is worth paying for when the person writing it has watched your operation and will be the one building what it recommends, and worth very little when it was generated from a questionnaire. You should own the document outright and be free to take it anywhere. Ours is the audit: two weeks, a fixed fee agreed before it starts, with the number given on the free call. What sets it is in what AI consulting actually costs.

Tool selection and setup. You do not need anything custom. You need someone to pick the right off the shelf product, configure it properly, connect it to what you already run, and train the team on it. A lot of good AI consulting is exactly this, and a firm that only sells custom builds will never recommend it.

A custom build. Integration code, prompts, a small service or an application, tested against your data, with a fixed scope and an end date. This is the product most people picture when they hear the phrase, and it is the one where ownership matters most. Source code, prompts, credentials, and documentation should be yours at handoff. If any of it stays with the firm, you have rented a dependency.

Operating what was built. Monitoring, fixing what drifts, extending the system month by month. This is the product most firms leave out of the proposal, and it is the one that decides whether the system is still running a year later. The engagement structure we use, with the edges on each phase, is in the Audit, Build, Operate model.

What it is not

Some of the confusion around the term comes from things that sit next to it and borrow the name.

Model research is a different job. No one trains a model for a forty-person property manager or a plumbing company, and no one should. The models come from a handful of labs, get rented by the call, and are already good enough for nearly everything on a small operator's list. A consultant who talks about "our proprietary AI" is usually describing a prompt.

A chat widget on the website is a product. Buying a product only counts as consulting when someone with no stake in the sale decided whether you needed it.

Your managed service provider is not it either. They keep the laptops patched and the email running, which matters. AI consulting is about the work that happens on those laptops: the retyping, the after-hours calls that go to voicemail, the month-end report that eats a week.

A slide deck can be the residue of the first decision, and a good one is worth having. If the operation runs the same way six months later, though, what you bought was a document, whatever the invoice called it.

And in our case it is not a headcount project. We do not build systems whose purpose is to cut staff. The systems we build give a coordinator back the two hours a week she spends retyping leads so she can spend them on the part of the job that needs a person. That is a choice about what the firm is for, and it is on our trust page so you know it before the first call.

When you need one, and when you do not

You probably do when the same routine work crosses two systems and a phone line, when someone retypes information that already exists somewhere else, when nobody can tell you how many calls arrive after six, or when a process lives entirely in one person's head. Those are the conditions under which a build pays for itself and an off the shelf product does not fit.

You probably do not when one tool solves the problem, when there is no process yet to automate (a business that has not decided how it handles a lead cannot automate handling a lead), or when the whole operation is one person and the honest budget is a couple of hundred dollars a month of software. We say the last one on the discovery call, because selling a build to someone who needs a subscription costs us the referral we would otherwise get. The arithmetic for working it out yourself is in can a small business afford an AI consultant, and who we are wrong for is spelled out on our Toronto page.

Telling a real one from the outside

Two earlier pieces already carry the questions to ask: four for any firm in the boom, including us, and six for reading a Toronto quote. If you only have time for one, ask who will pick up the phone in month four when the software on the other end changes. A name means you are buying a system. A ticket queue means you are buying a demo, and the price on the proposal will make a lot more sense once you know which of the two it is.

Frequently asked questions

What does an AI consultant actually do?

Maps where AI can take real work off a team, decides what to build or buy, sets where a person stays in the loop, and then builds and runs the thing inside the software the business already uses. Counted in hours, most of the job is watching the operation, writing integration code, tuning prompts against real messages, and monitoring after launch. Very little of it is what people picture when they hear the word AI.

Is AI consulting the same as automation consulting or an AI agency?

Mostly the same work under different labels. Automation consulting predates the current models and usually means wiring tools together with Zapier or Make. AI consulting adds a language model where judgment used to require a person: reading an email, classifying a call, drafting a reply. An "AI agency" is often a marketing agency that added the word. The test is identical in every case: who does the work, and what do you own at the end.

Do I need an AI consultant, or can I just use ChatGPT?

For drafting, summarizing, and one-off questions, use ChatGPT and keep the money. A consultant earns the fee when the work has to happen automatically, every time, inside your systems, without someone pasting things into a chat window: a lead parsed at 11pm, a call answered on a Saturday, a work order written into the CRM. Where the line sits for a landlord is in ChatGPT for landlords.

How is AI consulting different from a large consultancy's AI practice?

Scale, and who does the work. A national firm sells strategy by the hour and hands the build to a delivery team you meet later, at engagement sizes that begin where a boutique's largest build ends. A small firm's engineers scope the system, ship it, and answer the phone when it drifts. Neither is wrong. They are for different sized problems, and a firm that says which one you have on the first call is being straight with you.

How much does AI consulting cost?

Published rates in Toronto run from about $150 an hour for an independent to past $1,000 an hour at the global firms, and project quotes for what sounds like the same job land anywhere between $5,000 and $500,000. Ours are quoted on the first call: a two-week audit at a fixed fee agreed before it starts, a build at a fixed fee quoted against that audit, and a monthly fee to operate it. What moves each number, and the second bill for software that follows it, is in what AI consulting actually costs.

What should I own when the engagement ends?

Everything that runs: source code, prompts, credentials, documentation, and the audit document itself, from day one of operation rather than at the end of a contract. If any of it stays with the firm, you have rented a dependency, and the monthly fee is the rent.