Do not buy an AI strategy deck. Buy a list of your own processes.

An AI assessment is worth paying for when the deliverable is a list of your own processes with hours and dollars against them. It is not worth paying for when the deliverable is a slide about the AI landscape, a maturity curve with your business plotted on it, and three recommendations that would apply to any company in Ontario.

The difference is not effort or price. Both documents take a similar amount of work to produce. The difference is whether anybody went and watched how your business actually runs.

The eight things that should be in it

Treat this as a checklist to hold a supplier to before you pay, whatever they charge.

1. A process inventory. Every repeated task in the business that a machine could plausibly do, gathered from watching the work rather than from a questionnaire. This is the part that cannot be done remotely from a template.
2. Hours per week against each one. Measured, not estimated. Owners routinely underestimate administrative work by roughly a factor of two, and overestimate how often the dramatic exceptions occur that they believe prevent automation.
3. A dollar figure per year. Hours multiplied by a loaded hourly cost, stated so you can argue with the assumption rather than accept a total.
4. The three worth doing first, ranked, with a reason. Not eleven opportunities. Three, in order, with the reason the fourth did not make it.
5. Named tools with real subscription costs in Canadian dollars. Not “a workflow platform”. The actual product, the actual plan, the actual monthly cost.
6. A build estimate for each. In the same ranges a supplier would quote you anyway. On this site those are published: $2,500 to $12,000 for workflow automation and $4,000 to $15,000 for an AI agent.
7. The ones to skip, and why. This section is the tell. An assessment with no negative recommendations was written to sell the next thing.
8. Where personal information enters each process. Because that decides which tools are eligible before anybody discusses features, for reasons set out in what PIPEDA means for your AI tools.

If items 2, 4 and 7 are missing, you have bought a summary of the news.

Why the process capture has to be watched

The single most common failure in this work is describing a process instead of observing one.

What people describe is the process as designed. What they do is the process as it survived: the workaround for the field that will not accept a hyphen, the second spreadsheet that exists because the report does not include one column, the folder somebody keeps because a customer once asked for something six years ago.

Those survivals are where the hours are. They never appear in a questionnaire because nobody thinks of them as part of the job. They appear within about ten minutes of sitting beside somebody while they work, or watching a recorded screen share of a real task rather than a demonstration of one.

This is also why the useful engagement is short. Four days of watching produces a better inventory than four weeks of workshops, because a workshop collects opinions about the process and observation collects the process.

The diagnosis problem, in one figure

Statistics Canada asked businesses not adopting AI why not. The leading reason, from 40.0%, was that AI is not relevant to what they do. Cybersecurity and privacy came third at 13.4%, cost fourth at 10.6%.

That top answer is almost always wrong, and it is wrong in a specific and forgivable way. The owner pictures a chatbot on the website, because that is what the marketing shows, and correctly concludes it would not move the needle. Meanwhile somebody in the back office is retyping the same order into two systems 40 times a week, which is 30 hours a month, and nobody has counted it because it has always been done that way.

Counting it is most of the value. Once the hours are on paper the decision usually makes itself, in either direction.

What the numbers should look like when you get them

A worked example, from the shape of business this comes up in most often. A 30 person Ontario distributor, order entry happening twice because the system that takes the order and the system that fulfils it do not talk.

• Time per order: 4 minutes of retyping, plus roughly 2 minutes a day of fixing what got typed wrong.
• Volume: 60 orders a day.
• Weekly total: about 21 hours.
• Loaded cost at $32 an hour: roughly $35,000 a year.
• Build estimate: $4,000 to $9,000, because three systems have to agree.
• Added subscriptions: $60 to $180 a month.
• Payback: under four months at the top of the build range.

Now the same business, a different candidate. Monthly reporting, rebuilt by hand from three exports.

• Time: 6 hours, once a month.
• Annual total: 72 hours, roughly $2,300.
• Build estimate: $3,000 to $5,000.
• Payback: over two years.

Both are real automations. Only one is worth doing this year, and no amount of enthusiasm about the second changes that. Presenting them side by side with the arithmetic visible is the whole job.

Questions to ask before you pay anybody

• What exactly will I have at the end, in the form of a document I can show somebody else?
• How many hours will you spend watching people work, and whose time do you need?
• Will build estimates be in the document, or will they follow in a separate proposal?
• What is the process if the recommendation is to do nothing?
• Does the fee come off the work if I proceed, and is that in writing?
• Who owns the accounts and subscriptions afterwards, mine or yours?

The last one is worth more than it looks. Model and platform costs belong on your own accounts, billed to your card, in your name. When they sit on a supplier’s account, changing supplier means rebuilding.

Where this site stands

The assessment here is $999, takes ten business days, and every item on the checklist above is in it, including the negative recommendations. The whole fee comes off any build you go ahead with, which means the diagnosis costs nothing if you also want the work done, and $999 if you take the document elsewhere.

Taking it elsewhere is a fine outcome and the document is written for it. An assessment that only works if I do the build would be a proposal wearing a different name, and you would be right to discount its conclusions.

There is a related question worth settling before you commission anything, which is whether a provincial program will pay for the planning work. The Ontario Centre of Innovation offers up to $15,000 toward a digital adoption plan for eligible for-profit SMEs with 1 to 499 employees, on a first come first served basis. Timing and eligibility are yours to check, and the detail as published is in the note on the DMAP grant.

The one sentence version

Pay for a list of your own processes with hours against them, ranked, with the ones to skip named. Do not pay for a description of artificial intelligence, however well designed.