HeitechSoft - AI Business Solutions
All issues
Quarterly7 min read

HeitechSoft Quarterly: Fall 2024

August to October 2024

In our first quarterly issue, Tom Heiber shares what 18 months of building and training our own AI models taught us, and why we paused our private AI agent once capable models reached the market. Also inside: Efficiency Edge, coaching business owners at the Chamber’s Expert Exchange, an AI-ready technology roadmap for a growing franchise, and five AI developments that matter to mid-sized Canadian businesses.

Introduction

When ChatGPT made its splash, I saw how significant the technology was. After years of building IoT vehicle-interface hardware and a low-latency IoT network, we pivoted HeitechSoft to AI, and by mid-2023 our goal was set: become experts in how AI fits into a modern business, and build a private AI assistant a small business can afford. In the first half of 2024 I spoke about that work at the Greater KW Chamber's Operations Peer2Peer group, on a Humber College panel about AI and the future of work, and in two AI talks at a local business networking lunch.

This quarter was about the lab turning into practical help. We built and trained our own models so you don't have to, and we are starting to turn those lessons into services. Capable models kept appearing, so we paused building our own agent, and that changed which path we now recommend.

— Tom Heiber, CEO

Building our own AI models taught us what it really takes

Since spring 2023 our R&D team has built and trained AI models itself. We built a private, self-hosted AI agent that answers from a company's own documents. A proof of concept for it was built with a team from Humber College's AI and Machine Learning graduate program in early 2024. We also published two fine-tuned models on Hugging Face, and we ran AI models on ordinary off-the-shelf hardware instead of scarce data-centre GPUs.

This quarter the team kept training, fine-tuning and testing models, including the open-source and private models now reaching the market. As those capable models appeared, we paused building our own agent. The point was to learn the process, the costs, the effort and the results to expect, and we now know them first-hand.

What that gives you: we know what we can build, and where to find a ready-made model so you don't have to build one. That saves time and money, and it lets us advise which path makes sense: build, fine-tune or use an existing model.

Lab notes

Efficiency Edge: finding where AI pays off

We started shaping Efficiency Edge, a service where we review how a business runs and explore where AI pilot projects fit into its processes so that they pay off. In October we tested the idea with business owners and advisors and published a work-in-progress page. That page, including its code, copy and savings calculator, was created with generative AI.

The aim is a starting point grounded in your own processes: pilots placed where the work shows a return, rather than AI for its own sake.

We test AI on our own business first, including AI on the phone

Before we recommend an AI tool, we use it ourselves. We use ChatGPT and Perplexity for research and brainstorming, AI writing and image tools for marketing, a custom GPT assistant for fact-finding and pulling data from spreadsheets, and Microsoft Teams AI to summarise meetings and list action items. We also tested AI that places phone calls for market research and sales, and AI that answers sales and support calls, as experiments. We know where these tools fall short before you spend money on them.

Wins

Coaching business owners on AI at the Greater KW Chamber's Expert Exchange

The Greater Kitchener Waterloo Chamber of Commerce chose us as one of the tech experts for the first session of its new Expert Exchange series, held at Communitech on October 9. In small-group rounds we coached small business owners on tech trends, AI and productivity tools. A week later one owner told us they had already put some of our AI suggestions into practice, and another conversation from the session became a technology consultation.

Investors researching AI keep asking for our view

An expert research network that serves investors came back to us for our view on the generative AI industry. It is the latest of several requests since late 2023, covering AI-powered search, chatbots, conversational AI and large language models. The independent, hands-on judgement investors ask for is the same judgement clients get when choosing which AI tools to buy.

Choosing this year's systems with next year's AI in mind

A growing delivery company preparing to franchise asked us to help plan a technology overhaul covering dispatch software, the choice between Google and Microsoft, phones, infrastructure for new franchise locations and KPI tracking. We made AI readiness a deciding factor in the platform choice. As AI tools mature, the company will be able to automate a large share of its administrative work, so our recommendation tied the platform choice to how readily AI can be added later.

Humber College's AI program invited us back

After our private-AI capstone project earlier this year, Humber College's AI and Machine Learning graduate program invited us back as an industry sponsor for its Winter 2025 capstone projects. This time it can include teams from other programs as well. Graduate students get real business problems to solve, and we gain fresh hands and current research for our R&D.

By the numbers

  • 18
    months of hands-on AI model R&D, since May 2023. Long enough to learn the costs, limits and effort of AI models by doing the work ourselves, not just reading about it.
  • 2
    fine-tuned AI models we published on Hugging Face. Published models are a public record of our hands-on work that you can check for yourself.
  • 5
    AI talks, panels and expert sessions in 2024. Each talk puts practical AI decisions in front of owners and managers who are weighing them for their own teams.

AI for mid-market

One in ten Canadian businesses plans to use AI in the next year

Statistics Canada reported that 10.6% of Canadian businesses plan to use AI to produce goods or deliver services over the next 12 months, rising to 12.6% of those with 20 to 99 employees. Most firms of your size have no plans yet, so there is room to move early. The most common planned uses, text analytics and data analytics, can start from documents and spreadsheets you already have.

Source: Statistics Canada

Microsoft brings autonomous AI agents to Copilot Studio and Dynamics 365

Microsoft said building autonomous agents in Copilot Studio moves to public preview in November, alongside ten ready-made agents for sales, service, finance and supply chain in Dynamics 365. For a business of 20 to 200 people, agents that handle routine background steps, such as qualifying sales leads or chasing supplier updates, are moving from custom development into standard tools. Start with a well-documented, low-risk process and keep a person approving the output.

Source: Official Microsoft Blog

Claude can now operate a computer screen, in beta

Anthropic released an upgraded Claude 3.5 Sonnet with a public beta called 'computer use', which lets developers have Claude look at a screen, move the cursor, click and type. Anthropic called it experimental and at times error-prone. Many mid-sized firms run older software with no way to connect to it, so this points to AI that works those screens the way a person does, best suited to supervised trials such as copying data between two systems.

Source: Anthropic

Google NotebookLM turns documents into a podcast-style audio briefing

Google added Audio Overview to NotebookLM, which turns uploaded documents, slides and charts into a downloadable spoken discussion between two AI hosts, and Google called it experimental and English-only. For a mid-sized firm, manuals, product specs and training material can become a short audio briefing for staff on the road or on the shop floor. Review each one before sharing, because it reflects only the sources you upload and can get details wrong.

Source: Google

If any of these stories sounds like your business, book a 20-minute conversation with me at https://heitechsoft.com/contact and we will look at where AI could pay off for you.

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.

Subscribe

Remember which ad led you here?

With your permission, we save the ad source in this browser for up to 30 days and attach it if you send an enquiry. This helps us evaluate our advertising. This choice applies to this optional feature; it does not change other analytics settings.

Read our privacy policy