- Four cost items, and two are systematically forgotten: implementation, recurrent use, maintenance, and adoption.
- The monthly usage cost varies greatly depending on the design. Two seemingly identical systems can have bills ten times different.
- An internal research assistant costs less than an integrated three-system automation. The difference comes from integration, not AI.
- Always request a twelve-month usage-cost estimate before signing. A provider who cannot produce one does not understand the operating economics of the system.
Why prices vary so much
This is not necessarily dishonesty. It is often that the three providers have included three different projects.
The first included a prototype. The second included a system in production, integrated, with training. The third included a complete platform with several use cases.
This guide gives you something to compare apples to apples.
Item 1. Implementation
Development, integration, testing, commissioning. This is the only position that all quotes present.
Which really makes this amount vary, in order of importance.
- The number of systems to integrate. This is the number one factor, by far. An assistant that reads documents from a shared folder costs a fraction of the price of a system that reads your ERP, writes to your CRM, and triggers emails. AI is the easy part. Plumbing is the real job.
- The requirement for reliability. An internal tool tolerated at 90% accuracy costs far less than a system that must reach 99% with automatic verification and logging.
- Data sensitivity. Controlled architecture adds real work.
- The quality of your initial data. Clean, structured documents speed everything up. Scanned PDFs of varying quality add a whole step.
- The extent of change management. A user or forty is not the same project.
Item 2. Recurrent use
The position most often missing from quotes, and the one that surprises in year two.
Every processed document and every question has a usage cost. Individually, it may be only a fraction of a cent. At your actual volume over twelve months, it becomes a figure worth budgeting for.
What makes this position explode is almost never the number of requests. This is the amount of context sent each time. A system that sends thirty pages to each question will cost dozens of times the price of a well-designed system that sends the relevant two paragraphs. To you, the two look the same. For the bill, no.
What to demand, word for word: at the monthly volume that I gave you, what will be the estimated usage cost, and on what assumption of average context. Have the answer written down.
Item 3. Maintenance
Models evolve, your systems change, quality must be monitored. An AI system is not software that you install and forget.
Concretely: monitor that the quality of responses does not drift, adjust when a provider changes a version, adapt when your documents or processes evolve, correct problematic cases reported by users.
It's modest compared to the implementation, but it's not zero. A provider who leads you to believe that zero is realistic is preparing a nasty surprise for you or intends to bill for fixes per piece.
Item 4. Training and adoption
The most forgotten position and the one that decides success.
It's not just vendor hours. This is your people's time: participating in the design, testing with real cases, learning the tool, adjusting their ways of doing things. This time exists whether it is budgeted or not.
I've seen technically flawless systems abandoned in three months because no one planned for that time. This is the most frustrating waste ever, because it was entirely avoidable.
Realistic ranges, by project type
These amounts come from real projects in Quebec SMEs. These are orders of magnitude to situate you, not fixed prices, and your situation can legitimately fall outside of these ranges.
An internal research assistant. Your procedures and documents become questionable in natural language. Low monthly user cost. This is often the best first project: visible gain, low risk, complexity contained.
Automated processing of incoming documents. Extraction, validation, routing. Mid-range, ascending according to the number of formats and systems affected. Usage cost proportional to the volume processed. This is the project with the best measurable return in the majority of cases.
Sorting and routing requests. The scope stays lighter when the integration is simple. The value appears quickly when someone on your team currently performs this sorting by hand every day.
An agent who acts in your systems. Top of the range and often beyond, because each connected tool adds development, testing and safeguards. The higher risk also requires more validation and logging.
It's no longer an AI project, it's a software project with an AI component. Budget it as such.
This is what turns all the previous ranges into a defensible figure for your case.
| Type of project | Normal position | Determining factor |
|---|---|---|
| Internal research | Bottom of range | Quality of documents |
| Document processing | Mid-range | Formats and integrations |
| Integrated agent | High or beyond | Tools, risks and safeguards |
What we forget to budget for
A list of actual projects, all positions.
- Data preparation. Cleaning, scanning, putting in order documents that have never been stored.
- Your employees' time during design and testing.
- The period of parallel operation, when the old and new processes run at the same time.
- Internal documentation.
- Adjustments for the first three months, when actual usage reveals unanticipated cases.
- Reviewing your internal policies, if the project affects personal information.
How to read an AI quote
Six checks that will avoid most bad surprises.
- The usage cost is calculated, over twelve months, at your actual volume. If absent, the quote is incomplete.
- The number of integrations is specified by name. “Integration with existing systems" is not a staff, it is a blank cheque.
- A target accuracy rate is stated, along with the measurement method. Otherwise, what will you judge the delivery?
- Are the human validation points defined? Their absence signals a demonstration design, not production.
- Training and adoption appear. Otherwise, this cost will land at your house, not budgeted.
- Is maintenance planned, and at what pace? Zero is not a credible answer.
The profitability calculation, in three lines
No need for a complicated spreadsheet. Take the task.
Hours per week spent on this task, times the busy hourly cost, times 50 weeks. This gives you the current annual cost.
Next, estimate the portion that automation actually absorbs. Not 100%. In real life, between half and three quarters for a good use case, because there remains validation and exceptions.
Compare the annual savings to the cost of implementation plus one year of use and maintenance. If the project does not pay off within a reasonable time frame, it is not the right first project, and I will tell you that.
Locate the potential of the project
An exploratory reading of the effort and the possible value, without any financial estimate.
The time involved justifies a structured validation of the use case.
What I see on the ground
Two price differences explain almost all the differences between bids.
The first is the definition of “finished”. Some providers deliver a prototype that works as a demonstration. Others deliver a system into production, with edge case management, logging, training and documentation. The price difference between the two is normal. The value gap too.
The second is integration. A quote that seems twice as expensive often includes two more systems. Compare scopes before comparing prices.
And one piece of advice worth its weight in gold: be wary of the cheapest quote that doesn't mention an accuracy rate, running costs or maintenance. These are not oversights. These are deferred invoices.
When the project is not justified
Let us be clear on cases where money is better placed elsewhere.
- Low volume. Twelve documents per month do not justify automation.
- Unstable process. Wait till he stabilizes.
- Messy starting data. Clean up first, the project will cost less later.
- No person responsible for adoption. The best system in the world will die.
- Image objective rather than profit. It never pays off.
Frequently asked questions
01What is a realistic starting budget?
Enough for a real, measured use case, with adoption support. The assessment will give you an honest threshold for your specific situation.
02Why can one proposal cost three times more than another?
Almost always because of the number of integrations and the definition of "finished". Compare litters, not totals.
03Will AI usage costs increase over time?
The unit price of models drops almost every year. Your volume generally increases. The two trends partly offset each other. A well-designed architecture also makes it possible to change models when better value for money appears.
04Is an AI assessment really necessary?
Not always. If your use case is simple and clear, go directly to the implant. assessment is justified when multiple leads exist, when data are sensitive, or when management needs a solid record before committing a budget.
05Can we start with a small AI project?
Not only can we, but I almost always recommend it. A process, a team, measurable gains in weeks. Most AI-driven transformation programs end up in a drawer.
Background guide prepared by GXN (Digital Governance), a division of MD79. Revision planned according to the rhythm documented for this cluster.
References to verify for your situation
These references support the external rules and frameworks cited in this guide. They do not replace legal or professional advice tailored to your organization.