Download PDF (141 KB) PDF version , print or share with your team.
Written by Mike Pearlstein, CISSP, CEO of Fusion Computing Limited. Helping Canadian businesses build and manage secure IT infrastructure since 2012 across Toronto, Hamilton, and Metro Vancouver.
A 40-person firm called me last quarter with a question that sounded simple. Their operations lead needed everything the company knew about one client before a renewal meeting the next morning. The history sat in four places. Three years of email, a SharePoint folder nobody had opened since the project closed, a shared drive with two competing versions of the contract, and the ticket queue in their PSA tool.
The answer existed. No single person held it, and pulling it together by hand would have eaten most of a day she did not have.
I am a CISSP and an MSP operator, not an AI vendor. What follows is how I am actually seeing Canadian SMBs use ChatGPT agents to pull dispersed knowledge into a usable answer, where it pays off first, and the security trap I watch clients walk into when they connect an agent to a file estate nobody has cleaned up in years.
Key Takeaways
- Knowledge workers spend roughly 1.8 hours a day searching and gathering information (McKinsey Global Institute). For a 40-person firm that is a standing tax on every week.
- A ChatGPT agent distills knowledge by reading across your systems and synthesizing an answer. That is a different job from automating a recurring task.
- The agent reaches across email, drives, and line-of-business tools that used to need custom integration, code, and APIs to connect.
- The bigger risk is permissions rather than the model. An agent inherits whatever a user can already reach, and most SMB drives are over-shared.
- Start with one bounded question, scope the access tightly, and keep a human reviewing the output before you trust it.
What does it mean to distill knowledge with a ChatGPT agent?
Distilling knowledge means pointing an agent at the places your data and documents live, asking one question, and getting back a synthesized answer drawn from all of them at once. A ChatGPT agent differs from a chatbot you paste text into, because it can reach into connected sources on its own. It also differs from workflow automation, which performs a repeatable task on a schedule rather than answering a question.
I draw that line carefully because the confusion costs people time. If your goal is to run the same manual task again and again, that is automation, and I wrote about automating recurring workflows with ChatGPT agents separately. Distillation is the other half.
You have a question whose answer is buried across systems, and you want the agent to do the gathering and the first-draft synthesis so a person can finish in minutes instead of hours.
The practical test I give clients is one sentence. If you can describe the output as “tell me what we know about X,” it is a distillation job, and an agent is a strong fit.
Why scattered information is quietly expensive for a 25 to 50 person firm
According to the McKinsey Global Institute (2012), knowledge workers spend about 1.8 hours every day, close to a fifth of the work week, searching and gathering information. IDC research (2023, via Box) puts 80 to 90 percent of business data in unstructured form, spread across documents, email, and chat rather than tidy databases.
I see what those numbers feel like on the ground. At a 40-person firm, a fifth of the week across the team is the equivalent of several full-time people doing nothing but hunting for things other people already wrote down. Nobody books that cost anywhere. It hides inside “I will get back to you” and inside meetings that exist only to re-share what someone already knew.
Not sure where your business knowledge actually lives? Talk to our team about mapping it →
The reason it persists is structural, not lazy. IDC found that more than half of enterprise leaders say their unstructured data mostly stays siloed, and less than half of information is shared between people or systems. The knowledge is there. The wiring to connect it was never built, because building it used to mean a project.
Where your business knowledge actually hides
According to Statistics Canada (2025), the most common business use of AI is text analytics at 35.7 percent, ahead of data analytics at 26.4 percent and virtual agents at 24.8 percent. Reading and making sense of written text is the job Canadian firms already trust AI with, and in a typical SMB that text is spread across at least six systems.
Business knowledge in a typical SMB lives in at least six places: email threads, SharePoint and Teams, shared drives, the PSA or ticketing system, spreadsheets, and the heads of two or three long-tenured staff. No single system is the source of truth, and the most valuable context is often the email reply nobody filed and the spreadsheet on one person’s desktop.
When I walk a new client through this, I ask them to name where a specific answer would live. The answer is almost never one place. It is “check with Dave, then look in the 2023 folder, then the email from the vendor.” That sprawl is the real target. The AI knowledge management playbook covers the governance side of fixing it over time. An agent is the tactical lever you can pull this month.
How a ChatGPT agent reads across those silos, and what changed
According to the Microsoft Work Trend Index (2025), 81 percent of leaders expect AI agents to be moderately or extensively integrated into their strategy within 12 to 18 months, while only 40 percent of employees are familiar with agents today. The reach that closes that gap once required custom integration, code, and APIs.
An agent reads across silos through connectors that link it to your systems under a specific user’s access, then synthesizes one answer from many sources. The shift worth naming is this: the cross-system reach that used to require custom integration, code, and APIs is now something an agent can do without a developer in the loop. That is the part that is genuinely new in 2026.
I want to be precise, because this is where Fusion Computing spends a lot of its current client time. We are helping businesses use ChatGPT agents for three things.
Information retrieval across their systems. Ongoing manual tasks. And specific jobs that reach across ecosystems, the kind of work that in the past meant paying someone to build and maintain an API integration. The agent now bridges those systems directly, which collapses both the cost and the wait.
The pattern I teach is “ask once, synthesize many.” Instead of opening five apps, you ask the agent one question and let it assemble the picture. It is closer to Microsoft 365 Copilot reaching across your tenant than to a plain chatbot, and the two are worth comparing directly before you commit.
Three distillation jobs worth setting up first
The best first jobs are bounded, repeatable, and currently done by hand. Three fit almost every SMB: a “what do we know about this client” briefing assembled from email, files, and tickets; a quarterly operations or finance digest pulled from the spreadsheets and reports scattered across the team; and onboarding knowledge extraction, where the agent drafts the answers a new hire would otherwise interrupt three people to get.
Worried an agent could surface files it should not? Book a permissions and access review →
“Fusion gave us a CISSP-led security review in three weeks flat. We’d been quoted twelve weeks by two larger MSPs. They found a domain-admin gap our previous provider missed for two years.”
Operations Director, 85-employee Toronto law firm (client name on file)
I always start a client on the client-briefing job, because the value is obvious within a day.
Take the renewal meeting I opened with. The operations lead asked the agent for everything the firm knew about that account, and it returned a draft summary with the contract history, the open tickets, and the last three email threads in the time it took her to get coffee. She still read it and fixed two things. That is the point.
The quarterly digest is the second job I recommend, because it turns a painful manual chore into a review-and-approve step. The onboarding job pays off slowly but compounds, since every answer the agent can assemble is an interruption a senior person no longer fields.
Where this bites you: permissions, oversharing, and Canadian compliance
According to IDC (2024), the potential cost of a breach is nearly double for organizations with more fragmented, sprawling data. That matters here because an agent borrows the access of the user it acts for. Whatever that person can already reach, the agent can reach too, which in most SMBs is far more than it should be.
This is the part vendors skip, and it is where I spend the most breath. If your shared drive grants broad access by default, an agent acting for one staff member can now surface HR records, a former team’s client files, or a finance export sitting in a “temp” folder, all in one tidy answer.
The same oversharing problem shows up with Copilot, which I covered in Microsoft 365 Copilot oversharing. The agent widens nothing on its own. The exposure was already sitting in the over-shared drive, and the agent simply makes reaching it effortless.
For regulated Canadian SMBs the stakes rise. PHIPA for Ontario health information, PIPEDA federally, and Quebec’s Law 25 all govern who may see personal data and where it may travel. Data residency is a real question, since some AI providers store and process outside Canada.
Fusion Computing keeps a written data-residency map for regulated clients and reviews any new tool against it before deployment, not after. If your work touches client health, legal, or financial records, that review is not optional, and our cybersecurity services exist for exactly this gap.
What good looks like, and how to start without opening a hole
According to BDC (2026), only about 30 percent of Canadian SMEs used AI in 2025, yet those that did were 24 percent more productive than those that did not. Good adoption starts narrow: one bounded question, least-privilege access, and a person reviewing every output before the team relies on it.
Want to pilot one distillation job safely? Get in touch about a scoped, least-privilege start →
In practice that means a scoped pilot. You pick one bounded question, connect the agent under a single account with least-privilege access, keep a human reviewing every output, and refuse to connect your most sensitive repositories on day one. The firms that get value fastest are the ones that resist connecting everything and prove the model on one job first.
I run it in that order on purpose. First I clean up the permissions on whatever the agent will read, because the field notes above are not rare cases. Then I scope the connection to the smallest set of sources that answers the question. Then I insist the output stays a draft a person signs off on, until the team trusts it.
What I do not connect first is HR, payroll, and anything under PHIPA. Those wait until the access model is proven.
This is also where a steady advisory hand earns its keep. Deciding which sources to connect, in what order, against which compliance map, is a virtual CIO conversation as much as a technical one. Fusion Computing runs that planning with clients so the pilot creates value without creating a new liability. If you want a second set of eyes before you connect a single folder, get in touch.
Fusion Computing helps Canadian businesses across Toronto and the GTA, Hamilton, and Metro Vancouver with managed IT, cybersecurity, and Microsoft 365.
Why Canadian firms bring this work to Fusion Computing
CISSP-led, a Microsoft Solutions Partner and a CompTIA Managed Services Trustmark holder, securing IT for Canadian SMBs across Toronto, Hamilton, and Metro Vancouver since 2012.
Where to start this week
Pick one question your team answers by hand every week and point an agent at it, under a single account with least-privilege access. Clean the permissions first, keep a person reviewing the output, and leave your most sensitive records disconnected until the model proves itself. That is how a 25 to 50 person firm gets value from ChatGPT agents without opening a hole.
Frequently Asked Questions
What is the difference between a ChatGPT agent and a regular chatbot?
A regular chatbot only knows what you paste into it. A ChatGPT agent connects to your systems, retrieves information across them, and synthesizes an answer on its own. For business knowledge, that reach is the difference between re-typing context every time and asking one question that draws on email, files, and tickets together.
Is it safe to let an AI agent read our company files?
It is safe only if your file permissions are clean first. An agent inherits the access of the user it acts for, so an over-shared drive becomes a larger exposure. Tighten permissions, scope the agent to the fewest sources needed, and keep a person reviewing output before you trust it.
How is distilling knowledge different from automating a workflow?
Distillation answers a question by reading across systems. Automation performs a repeatable task on a schedule. They use the same tool for different jobs. If you want the same action run again and again, that is automation. If you want to be told what you already know about something, that is distillation.
Do we need a developer to connect an agent to our systems?
Usually not anymore. The cross-system reach that once required custom integration, code, and APIs is now something an agent handles through built-in connectors. That is the main change in 2026. A developer may still help with complex or custom sources, but most common business systems connect without a build project.
What Canadian privacy laws apply when an agent reads our data?
PHIPA, PIPEDA, and Quebec’s Law 25 all govern who may access personal data and where it can be stored. Some AI providers process data outside Canada, which raises residency questions. Keep a written data-residency map, review each tool against it before deployment, and never connect health, legal, or financial records until the access model is proven.
Where should a 25 to 50 person firm start?
Start with one bounded, repeatable question someone currently answers by hand, such as assembling everything you know about a client before a meeting. Connect the agent under a single least-privilege account, keep a person reviewing the output, and expand only once that first job earns trust.
How much does it cost to run ChatGPT agents for a small business?
Cost depends on the plan and how many staff use it. A business ChatGPT subscription runs on a per-user monthly fee, far below the price of building custom integrations. The larger cost is the setup work to clean permissions and scope connections safely, which is where planning with an IT partner pays off.
