Buy an AI product when the task is common to many organizations and the product meets your data rules; build your own tool on a bought model when the workflow is specific to you, needs deep links into your systems or must keep data where products cannot. Many Canadian organizations end up with both. Remolda's AI Opportunity Audit makes this call per use case in two weeks for $2,900 CAD plus applicable taxes.
Bring the AI uses you are weighing. Send a written request with the two or three AI uses you are weighing and the data each one touches, or call +1 (613) 867‑7767. We reply within one business day.
The three options
"Build" rarely means training a model. In practice the choice is between three options:
| Option | What you own |
|---|---|
| Buy a product: Microsoft 365 Copilot, ChatGPT Enterprise, Claude Team or a specialised SaaS tool | Settings, users and the content you put in |
| Build on a bought model: your own app or agent calling Azure OpenAI, Amazon Bedrock, OpenAI or Anthropic through an API | Prompts, connections to your systems, tests, data flow |
| Train or fine-tune your own model | The model itself, plus the team to maintain it |
The third option makes sense only with large proprietary data, a measured gap that prompts and retrieval cannot close, and engineers who stay. Our comparison of RAG and fine-tuning explains when it is worth it. For most organizations the real decision is between the first two, and the usual answer is a mix.
A six-question scorecard
Answer each question for one use case at a time, never for "AI in general".
| Question | Buy if | Build if |
|---|---|---|
| Is the task common? | Notes, email, summaries | Your own process and rules |
| Where does the data live? | Inside the product's suite | In your own systems |
| Do products meet your data rules? | Yes, in writing | No, or partly |
| How deep is the integration? | Read and draft | Write back, approve, trigger steps |
| How much use? | Few hundred users | High volume of cases |
| Who will own it? | An admin | A named product owner |
Four or more answers in one column usually decide it. A split result often means a bought product for staff now and a custom tool for one workflow later. The AI build-vs-buy evaluator lets you weigh your own criteria before a call.
Data rules come before the scorecard
Where data may be stored and processed can remove an option before cost or speed are discussed.
- PIPEDA. Principle 4.1.3 keeps your organization responsible for personal information it transfers to a provider for processing, and expects contractual protection.
- Quebec, Law 25. Section 17 requires a privacy impact assessment before personal information is communicated outside Quebec; section 3.3 requires one for a new information system. Our AI compliance review covers both.
- Federal Protected B. Subsection 4.4.3.14 of the Directive on Service and Digital asks departments to identify and evaluate computing facilities in Canada, or on Government of Canada premises abroad, as the principal delivery option for Protected B, Protected C and classified data.
- Health information. Rules are provincial; see AI and health privacy in Canada.
- Contracts. Client, funder and insurer agreements often carry their own residency clauses. Read them before shortlisting.
All three large clouds run Canadian regions: AWS Canada (Central) in Montréal and Canada West in Calgary, Azure Canada Central in Toronto and Canada East in Québec, Google Cloud in Montréal and Toronto. Language models differ by region:
| Option | Where prompts are processed |
|---|---|
| Azure OpenAI | Canada East standard deployments for some models |
| Amazon Bedrock (Canada Central) | Logs and stored data in Canada; inference may run in US or global regions |
| OpenAI | Storage at rest in Canada for eligible customers; processing outside Canada |
| Anthropic API | US or global |
This is the state our AI vendor selection page recorded in September 2026; recheck each model and region on the date you sign.
What each option really costs
Prices for products are easy to compare; the costs that surprise people sit around them.
| Buy a product | Build on a model |
|---|---|
| Licence per user per month, plus any required base plan | Build work: design, connections, tests, security review |
| Admin time: settings, permissions, offboarding | Model usage, billed per million tokens |
| Training staff to use it well | Hosting, logs and monitoring |
| Price changes at renewal | Retesting when the provider updates or retires a model |
| Workarounds where the product does not fit | A named owner who keeps it working |
In our experience, for many internal tools the model bill ends up smaller than the people cost around it. Our AI ROI guide lists current model prices and shows how to compare both columns over 12 to 36 months, and the business case service builds that comparison with your finance team.
Speed: who gets to value first
A bought product is usually in staff hands within weeks, once the settings and an AI use policy are in place. A custom tool takes longer, but less than many expect when it is built on a bought model: our AI Pilot Sprint puts one workflow into daily use in six weeks. The slowest path is choosing a product that later fails a privacy review, so check the data rules first.
Keep the exit open
Lock-in comes from content you cannot export, tools only one vendor can run and staff habits built around one product. Four habits keep the choice reversible:
- Own your assets. Prompts, test cases, source documents and integration code live in your repositories.
- Put a thin layer between your code and the model provider. Switching or routing tasks to another model then becomes a configuration change. Our API integration work builds that layer.
- Read the export and exit terms. Data export, conversation logs, deletion on exit and notice of price changes.
- Keep a test set. Twenty to fifty real cases with known good answers let you compare a new model or product in a day.
A sensible starting order
For a first year, we usually recommend: buy one general assistant for staff with the right settings and training; pick one workflow where the scorecard points to building and build it on a bought model with a baseline; review both after six months with real usage data. Founders building AI into their own product face a different version of this choice, covered in AI for startups and scale-ups.
How Remolda helps
- AI Opportunity Audit ($2,900, two weeks): your use cases ranked, with buy, build or wait for each, the data rules and a 12-month cost.
- AI vendor selection: two or three shortlisted products tested on 10–20 of your real tasks, included in the audit.
- AI Pilot Sprint ($9,800, six weeks): the first custom tool built on a bought model and in daily use; see custom AI development.
- AI Readiness Review ($490, one week): whether your data, people and rules are ready before you choose.
Prices are in CAD plus applicable taxes and listed on the pricing page.
Compare a product and a custom tool on your own use case. Send a written request with the use case, the number of users and the data involved. We reply within one business day.
Sources
- Justice Laws: Personal Information Protection and Electronic Documents Act, Schedule 1, principle 4.1.3
- LégisQuébec: Act respecting the protection of personal information in the private sector (P-39.1), ss. 3.3 and 17
- Treasury Board: Directive on Service and Digital, subsection 4.4.3.14
- Government of Canada cloud guardrails: Data location
- Microsoft Learn: Foundry models sold by Azure, region availability
- AWS: Amazon Bedrock cross-Region inference in Canada (Nov. 24, 2025)

