Start here · Chapter 2
What an AI consultant actually does
The job title has spread faster than any agreed definition of it. This chapter describes the work itself: the four stages a competent engagement moves through, the situations where paying for outside help genuinely pays, the situations where it does not, and what to look for in a proposal before you sign it.
Key takeaways
- 01An AI consultant's real product is a decision, not software. Most of the value is in choosing which task to automate and which to leave alone.
- 02The work splits into four stages: find the task, prove it with a small test, build it into the daily process, and hand it over so it runs without them.
- 03Hire outside help when work crosses systems you cannot connect, when personal information is involved, or when an earlier attempt failed and nobody can say why.
- 04Do it yourself when the task is contained, the tool is off the shelf, and someone on staff has the time to own it.
- 05The clearest warning sign is a proposal that names a technology before it names the task and the hours it currently costs you.
The product is a decision, not software
The most valuable thing an AI consultant produces is usually a decision about where not to spend money. A business arrives with six ideas. Four of them will not repay the effort, one is worth doing later, and one is worth doing now. Sorting those six into that order is the work. The building that follows is comparatively mechanical.
This is why an engagement that opens with a product demonstration should make you cautious. The demonstration proves the tool can perform a task under favourable conditions, which was never in question. What matters is whether that task is one of yours, how many hours it consumes, how often the output would be wrong, and who checks it. None of that is visible in a demo.
It also explains why the failure rate is so high even at well-resourced companies. MIT's Project NANDA found that 95 percent of generative AI pilots showed no measurable profit-and-loss impact across 300 public deployments. Those organisations had no shortage of technology. What was missing was the decision about which work should change, and the follow-through to change it.
The four stages of a competent engagement
Find the task
Interviews with the people doing the work, plus timing of how long things actually take. The output is a shortlist of candidate tasks ranked by hours consumed and how easy the result is to check. Most of the value of an engagement is created here, before any tool is chosen.
Prove it small
One task, one narrow test, run against real work rather than a demo dataset. The purpose is to find out where the output is wrong and how often, because that error rate decides whether the thing is usable and how much checking it will always need.
Build it into the day
Putting the working version where the work already happens, rather than in a separate app someone must remember to open. This is the stage most failed projects skip, and it is the difference between a successful demo and a system people actually use.
Hand it over
Documentation, training, and a named internal owner, so the system survives without the consultant. An engagement that ends with only the consultant able to operate what was built has produced a dependency rather than a capability.
Stage three is where most projects quietly die. A tool that works but lives in a separate browser tab competes with habit, and habit usually wins. Building the output into the system people already open every morning is unglamorous work that rarely appears in a proposal, and it is the single strongest predictor of whether anyone is still using the thing in six months.
Hire, or do it yourself?
| Situation | Better answer | Why |
|---|---|---|
| Choosing a writing or meeting-notes tool | Do it yourself | Reversible in 30 days. The trial costs less than the advice. |
| One contained task, staff member has time to own it | Do it yourself | They know the process and they are still here afterwards. |
| Work spans two systems that do not talk to each other | Get help | Integration is where unassisted attempts usually stall. |
| Customer personal information is involved | Get help | PIPEDA obligations apply and mistakes here are expensive. |
| An earlier attempt failed and nobody knows why | Get help | Repeating it without a diagnosis usually reproduces the result. |
| Nobody internally can name the task yet | Neither, yet | Start with the readiness questions. |
Warning signs in a proposal
- ·It names a technology before it names your task. A proposal that opens with a platform or model name has decided the answer before studying the problem.
- ·There is no number for how long the task takes today. Without a baseline nobody can prove the project worked, which conveniently means nobody can prove it did not.
- ·Success is described as adoption or engagement. Seats filled and logins recorded are activity. Hours saved and errors avoided are results.
- ·There is no stopping point. An engagement without a date where you decide to continue or stop tends to continue by default.
- ·Ownership after handover is vague. Ask what happens to the system if you stop working together. If the honest answer is that it stops, you are buying a dependency.
A note on how this site prices the work
While researching this chapter I looked for an independent survey of Canadian AI consulting rates and could not find one. Every published 2026 rate card located was written by a firm selling the service, which makes it marketing rather than research, so no rate benchmark is quoted here. For what this practice charges, the cost and ROI chapter states the figures directly and labels them as our own published rates. The one independent data point available is that Statistics Canada found 10.6 percent of businesses cited cost as a barrier limiting AI use in Q2 2026, behind cybersecurity and privacy concerns at 13.4 percent.
Frequently asked questions
What does an AI consultant actually do day to day?+
An AI consultant spends most of their time on the business rather than the technology: interviewing the people who do a task, timing how long it takes, mapping where the work stalls, and deciding whether a machine can help. The building phase is usually the shortest part. A consultant who starts by demonstrating a tool, rather than by asking what your week looks like, has skipped the part that determines whether the project works.
What is the difference between an AI consultant and an AI developer?+
An AI consultant decides what should be built and whether it should be built at all. An AI developer builds it. On small projects one person often does both, which is efficient, but you should still know which hat is being worn at any moment. The risk in combining them is that someone paid to build has an incentive to conclude that building is necessary.
When is hiring an AI consultant worth the money?+
It is worth it in three situations: the work crosses systems you cannot connect yourself, personal information is involved and the privacy obligations are unclear, or a previous attempt failed and nobody can explain why. Outside those cases a capable staff member with a few hours a week usually gets further, because they already understand the process and they stay after the project ends.
When should I not hire an AI consultant?+
Do not hire one to choose a subscription for you. If the task is picking between two well-known writing or meeting-notes tools, buy the cheaper one and try it for a month. Consulting fees make sense when a decision is expensive to reverse. Choosing a per-seat tool you can cancel in 30 days is not that decision, and paying professional rates to make it wastes budget you will want later.
How do I judge whether an AI consultant is any good?+
Ask them to describe a project that did not work and what they changed afterwards. Someone with real delivery experience answers immediately and specifically, because most practitioners have had projects stall. Also ask what they would do if the answer turned out to be that you do not need AI for this task. A consultant who cannot describe that outcome is selling a product rather than advice.
What should an AI consulting proposal contain?+
It should name the specific task, state how many hours that task consumes now, describe what will be different afterwards, and set a date when you decide to continue or stop. It should also say who owns the system once the engagement ends. A proposal built around a technology name, a platform partnership, or a fixed number of workshops is describing activity rather than a result.
Do I need a data scientist or a machine learning engineer?+
Most businesses under 50 people do not. Those roles matter when you are training a model on your own data, which is rare at this size and expensive to do well. The common need is connecting existing tools and redesigning a process around them, which is integration and operations work. Hiring a research specialist for an integration problem is a costly mismatch.
How long does a typical small business AI engagement last?+
A first engagement aimed at one workflow generally runs a few weeks rather than a few months, because the point is to reach a keep-or-stop decision quickly. Longer engagements are appropriate once something is working and you are extending it. Be cautious about an open-ended retainer at the start, before either side knows whether the first workflow succeeds.
Who owns what the consultant builds?+
Settle this in writing before work starts. You want ownership of the configuration, the prompts, the documentation, and any data produced, plus the ability to keep running the system if the relationship ends. Some vendors build on a platform that only they can maintain, which converts a project into an indefinite dependency. Ask directly what happens to the system if you stop working together.
Can an AI consultant guarantee results?+
Nobody can honestly guarantee a business outcome from an AI project, and MIT's finding that 95 percent of generative AI pilots showed no measurable profit impact is the reason to distrust anyone who does. What a consultant can commit to is a defined scope, a measurement method agreed before starting, and a decision point where you stop if the numbers do not move.
What questions should I ask in a first conversation?+
Ask which task they would start with and why that one. Ask what they need from your team and how many hours it will cost your staff. Ask how success will be measured and who measures it. Ask what would make them recommend stopping. The answers reveal whether they are thinking about your operations or about their own delivery schedule.
Is an AI consultant the same as a fractional CMO?+
No. A fractional CMO leads the marketing function: positioning, demand, team, and budget. An AI consultant works on how tasks get done, which may or may not sit inside marketing. The two overlap when the AI work is aimed at marketing operations specifically. If your problem is that marketing lacks direction, the AI question is premature and the leadership question comes first.
Sources and references
- Statistics Canada: Analysis on artificial intelligence use by businesses in Canada, Q2 2026 (published 11 June 2026)
- Bank of Canada: Canadian businesses' use of AI: What the evidence shows (August 2026)
- MIT Project NANDA: The GenAI Divide: State of AI in Business 2025 (PDF)
- Office of the Privacy Commissioner of Canada: Principles for responsible, trustworthy and privacy-protective generative AI technologies
- Office of the Privacy Commissioner of Canada: PIPEDA in brief
- Parliament of Canada: Bill C-27, Digital Charter Implementation Act, 2022 (died on the Order Paper, 6 January 2025)
Figures verified against the primary sources on 3 September 2026. The Statistics Canada barrier figures are from the Q2 2026 release published 11 June 2026.