Loading
The automations that pay back fastest are the ones where someone today reads something and retypes it somewhere else: product data, supplier documents, support replies and weekly reports. A well-chosen first automation costs CAD $10,000 to $30,000 to build and removes a recurring task within weeks. A chatbot on the homepage is usually the wrong place to start.
Updated · 4 min read
Look for read-and-retype work. Someone opens a supplier's spreadsheet and keys products into the store. Someone reads a purchase order PDF and enters the lines. Someone answers the same forty support questions by hand. Someone assembles the Monday report from three systems and a calculator. Those are the jobs where a model that can read messy input and produce structured output changes the economics.
Measure before choosing: hours a week, multiplied by the people doing it, multiplied by what their time is worth. Under two hours a week, leave it alone for now. Over ten, it is probably the first project.
Before language models, automation needed tidy input: a form, a CSV, an API. The new capability is handling the untidy kind, PDFs, emails, photos of labels, free-text notes, well enough to be useful. That is what makes the read-and-retype jobs automatable now when they were not three years ago.
What did not change is everything around the model. The automation still needs a reliable place to put its output, a rule for what to do when it is unsure, a review step for anything that reaches a customer or moves money, and logs so you can see what it did. The model is a fifth of the build. The plumbing and the checks are the rest, and they are what separate a demo from a tool the business trusts.
Catalog enrichment: descriptions, attributes, alt text and translation drafts generated from supplier data and your own rules, queued for a person to approve before anything publishes. Document extraction: supplier price lists, purchase orders and invoices turned into rows in a spreadsheet or in your inventory records, with the original attached for checking.
Support drafting: incoming messages classified and answered in draft from your own policies and past replies, with a person pressing send. Internal search: your manuals, policies, past jobs and product documents made answerable in plain language, so staff stop asking the one person who knows. All four have a clear before-and-after, keep a human on anything that matters, and reuse the same plumbing once the first one is built.
Homepage chatbots: low intent, high embarrassment risk, and they answer questions a good page would have answered. Anything that sends money or messages without a review step. Automations built on data nobody has cleaned, because the model will confidently extend the mess. Per-seat tools whose pricing scales faster than the value they return.
And the one that catches most teams: automating a process that nobody has written down. If two people do the job two ways, the first week of the project is agreeing on one way, and that week is where most of the value turns out to be.
Week one is spent with the person who does the job today, mapping the inputs, the exceptions and what good output looks like. Weeks two to four are the build, run against real documents from day one, with a review queue so every output is checked until the error rate is known. Then the measurement: hours saved against the baseline, errors against the human rate, and the decision to widen it or stop.
A first automation of this shape typically costs CAD $10,000 to $30,000 to build, with model usage at small-business volume running tens of dollars a month. We build them as custom software inside the Development suite, and the code, the prompts and the data stay yours.
No. Commercial models accessed through an API cover nearly every small-business case, and they improve without you doing anything. Your advantage is your data and your process, not a model.
Use API terms that exclude training on your data, keep personal information out of prompts wherever the task allows, log what the system did, and treat Ontario and Quebec privacy rules as design constraints from the start. Most of this is configuration rather than heroics.
In a small business it usually removes the part of a job people disliked and moves them to the part that needs judgment. Plan for the same people doing more, and for the one person who was the bottleneck to stop being one.
In Canada a focused internal tool typically costs CAD $15,000 to $40,000, a customer-facing web application $40,000 to $120,000, and a SaaS product you can charge for $100,000 to $250,000 before the first paying customer. Integrations, user roles and undecided requirements move the number; the count of screens does not.
RegulationLaw 25 applies to any business that collects personal information from people in Quebec, whether or not it has an office there. Since September 22, 2023 that means a published privacy policy, a named privacy officer, consent that is specific and separate, and tracking that stays off until the visitor turns it on. Administrative penalties reach the greater of $10 million or 2 percent of worldwide turnover.
Hiring and costA CAD $120,000 developer costs roughly $155,000 a year once payroll costs, recruiting, equipment, and ramp-up are counted, and buys about 160 usable hours a month whether you use them or not. Here is the honest comparison against a $5,000 monthly retainer, including the cases where hiring wins.
Nockta builds ecommerce, custom software, and Amazon operations for businesses across Ontario and Quebec. See where we work or the full service list.