QuarterlyMay 1, 20258 min read

HeitechSoft Quarterly: Spring 2025

February to April 2025

This quarter, businesses started asking us for AI that knows their own business. We built our own demo chatbot on the City of Kitchener’s public website, trained on more than 500 pages, that answers residents’ questions in the language they ask in. An Ontario concrete contractor then signed with us for an AI assistant that answers from its own documents, project history and past quotes. In the lab, we are sponsoring two Humber College student teams to fine-tune a Canadian AI model, giving ChatGPT and Gemini the same proposal to write, and testing AI clones of our own voices. We also cover the Greater KW Chamber of Commerce budgeting for AI training, our acceptance into NVIDIA Connect, and seven AI developments that matter to businesses of 20 to 200 people, from Shopify’s AI-first memo to Cohere’s privately run model.

Introduction

This quarter, businesses started asking us for AI that knows their own business. It began with a demo: in March we built a chatbot on the City of Kitchener's public website, our own project rather than the City's, and within three weeks it was answering residents' questions in the language they asked in. The moment that surprised me most was a question asked in Somali and answered in Somali. By April an Ontario concrete contractor had signed with us for an AI assistant that answers from its own documents and past quotes.

— Tom Heiber, CEO

From a city website demo to a contractor's first AI assistant

In March we built a demonstration chatbot on the City of Kitchener's public website. It is our own demo, not a City project. Within three weeks it was working: trained on more than 500 pages of the site, it answers the questions residents ask, in the language they ask in. A general contractor that saw it asked us for one for its own website.

Meanwhile an Ontario concrete contractor contacted us through our website. It had tried AI on its own and hit two limits: the AI did not remember from one day to the next, and there were privacy concerns. We proposed starting small and proving value before scaling, and in April it signed.

Work began with an Efficiency Edge review on site on April 28. One finding: a single person handles all of the company's quoting and estimating. The first step is a small pilot, an internal AI assistant that indexes the company's documents, project history and past quotes so staff can search them and ask questions.

For your business, that means the knowledge behind every quote no longer lives in one person's head, and you can see AI working on your own material before you commit.

Lab notes

What would it take to build a Canadian DeepSeek?

Humber College and we had talked about building a Canadian model along the lines of China's DeepSeek. Training one from scratch takes resources few organisations have, but fine-tuning an existing open-source model needs far less hardware. In April we agreed to sponsor two Humber capstone student teams, starting this summer, to develop a prototype Canadian AI language model built for privacy, security and localisation. Businesses with confidential documents want AI that stays under their control, and this is how we will learn what a private model really takes.

ChatGPT or Gemini? We gave both the same proposal to write

Before we standardise on an AI assistant, we test the candidates on our own work. This quarter ChatGPT and Gemini each drafted the same client proposal, and we ran prospect research through both. By the end of April most of our work had moved to Gemini, mainly because it works inside Google Workspace and Google Drive, where our documents and many of our clients' documents live. Quality was comparable, so where the documents already live was the deciding factor.

We cloned our own voices to test AI voice agents

Before we offer AI voice agents to clients, we want to know what they can really do. In April we made AI video avatars and AI clones of our own team's voices, recorded in a small studio at our office. A Romanian speaker who heard one clone speak Romanian said it was perfect, though an AI-generated voice still loses some detail. We now know where AI voices hold up before a client trusts one with its phone line.

Wins

The Greater KW Chamber of Commerce budgets for AI training

In April the Greater KW Chamber of Commerce asked how it could understand, train and implement AI in its own processes. We proposed an Efficiency Edge operational review and an AI knowledge chatbot as a quick win for members. We presented the plan to the Chamber's senior leadership team, which then included budget for AI training, with a follow-up planned for the fall. Training its own people first, before a larger project, is a sensible order for any organisation.

AI in front of local owners, from a lunch talk to a live demo

On February 6 we spoke about AI and Efficiency Edge at a local business networking lunch, asking the room one question: what is one way AI could give your company a competitive edge? On April 16 we exhibited at the Greater KW Chamber of Commerce Business Expo, where our City of Kitchener chatbot demo ran live and visitors put their own questions to it; one asked about the local crime rate. Seeing AI answer on the spot led to follow-up conversations about their own businesses.

Accepted into NVIDIA Connect

In April NVIDIA accepted HeitechSoft into NVIDIA Connect, its program for software companies, which NVIDIA says is designed to help companies build their solutions faster with a range of free benefits. It supports our private-AI work, including the Canadian model project with Humber College.

By the numbers

500+
pages of the City of Kitchener's public website behind our multilingual AI chatbot demo. An assistant can only answer from what it has read, so your own website and documents are where a useful build starts.
2
Humber College capstone teams we are sponsoring to fine-tune a Canadian AI model. Student teams will show what a private, Canadian-built AI model really takes, which matters to any business with confidential documents.
2
hours a week each team member is asked to spend finding better AI tools for their job. Regular time to test new tools is how we decide which AI is worth recommending to a client.

AI for mid-market

Shopify's CEO makes everyday AI use a baseline expectation for every employee

A memo from Shopify's chief executive, widely circulated on April 7, makes reflexive AI use a baseline expectation, and teams must explain why AI cannot do the work before asking for more resources. A 20-200 person firm can borrow the test: before adding headcount or buying another system, ask which part of the work AI could do, and give staff approved tools and time to learn them.

Source: Digital Commerce 360

Canadian leaders want AI agents to add capacity, but only about a third use them

Microsoft's 2025 Work Trend Index found that most Canadian business leaders expect AI agents to add capacity in the next 12 to 18 months, yet only about a third of Canadian organisations use agents to automate workflows. Automating one repetitive workflow this year would put a mid-sized firm ahead of most; start where staff say they are stretched thinnest, and treat the figures as a vendor's survey.

Source: Microsoft Canada

Canada's Cohere releases a capable AI model that runs privately on two GPUs

Cohere released Command A, which it says matches or outperforms GPT-4o on enterprise tasks and can run privately on two data-centre GPUs, aimed at organisations such as finance and healthcare that need internal deployments. Two GPUs is still a real cost for a 20-200 person firm, so most will do better with a business AI plan with clear data terms, but firms with sensitive data now have a Canadian option to shortlist.

Source: SiliconANGLE

Google Workspace Flows brings AI automation of multi-step work into Gmail, Docs and Sheets

Google announced Workspace Flows, which automates multi-step work with AI that can research, analyse and generate content from Drive files, now in alpha, with more Gemini features for Docs, Sheets and Meet to follow. For firms on Google Workspace, routine work such as sorting a shared inbox is heading into tools you already pay for; list candidate processes now rather than build on alpha features.

Source: Google Workspace Blog

Claude connects to Gmail, Calendar and Google Docs

Anthropic connected Claude to Gmail, Google Calendar and Google Docs in beta for all paid plans; on Team and Enterprise plans an administrator must enable it first. AI assistants are moving into the email and documents your team already uses, so decide which tools may connect to company accounts: one approved connection can read a great deal of client correspondence.

Source: Claude (Anthropic)

ChatGPT's deep research reaches business plans: a cited research report in minutes

OpenAI opened deep research to ChatGPT Plus subscribers and extended it to Team, Enterprise and Edu plans: it gathers information from many online sources and uploaded files into a cited report, typically in 5 to 30 minutes. A competitor scan or a summary of new rules can now arrive as a cited first draft in under an hour; open the sources before acting, and keep company files on a business plan, not personal accounts.

Source: Dataconomy

ChatGPT can now make images with readable text, from menus to promotion graphics

OpenAI built image generation into GPT-4o in ChatGPT, saying it renders text accurately and follows detailed prompts, such as a full restaurant menu. A promotion graphic or trade show sign can now be drafted in-house; check every word and number, and never pass off generated images as photos of your real products, premises or people.

Source: Search Engine Journal

If you want to see what AI could do with your own documents and processes, book a 20-minute conversation at https://heitechsoft.com/contact.

Get the next issue by email

Each quarter we share what we built, what we learned and the AI news that matters to mid-sized businesses.