Great Lakes IT: the AI capability behind a managed service provider

Great Lakes IT is a Canadian managed service provider. Its clients — small and medium businesses, professional services firms, nonprofits and associations, healthcare and community services, public sector bodies and Indigenous organisations — trust it with their infrastructure, their security and increasingly their questions about AI.

That last part is the problem every MSP now has. Clients ask what they should do about AI, and the honest answer for most providers is that they do not have the capability in house. Saying so repeatedly erodes the advisory relationship the whole business model rests on. Saying yes without the capability is worse.

An MSP cannot sell an experiment. Its clients are the least tolerant possible audience for AI that occasionally invents things: organisations handling health information, public funds, member records and community services, several of them with governance obligations that make "the model said so" an unacceptable answer.

The work also has to be repeatable. A one-off custom build for a single client is a project; an MSP needs a solution it can scope, price, deliver and support across many clients of different sizes without rebuilding it each time.

And it has to be deliverable by their team, not ours. Whatever we build has to be operable by people who are excellent at infrastructure and security and who are not machine learning engineers.

The arrangement is a partnership rather than a project. Great Lakes IT holds the client relationship, the account management and the support; we are the AI and engineering capability behind it. That is a common shape in this market and rarely acknowledged publicly — most MSPs offering AI services have somebody like us behind them.

The core solution is a governed knowledge assistant. Documents are ingested from approved sources only, made searchable semantically across policies, procedures and files, and answered against with responses grounded in — and cited back to — the source material. Access is scoped by department or programme, so the assistant respects the permission model the organisation already has rather than flattening it.

Around that sit human review workflows for sensitive or high-impact content, usage analytics that surface which questions are being asked, and a knowledge-gap report that shows where the organisation’s own documentation is missing, outdated or contradictory. In practice that last output is often the most valuable one: it tells a client something true about themselves regardless of whether they keep the assistant.

The wider solution set extends the same approach into business process automation, customer and employee portals, secure modern workplace, cloud modernisation and business continuity — so an AI conversation can turn into an engagement that fits how the MSP already sells.

We also built and now run their public platform: a Next.js application deployed to Cloudflare Workers through the OpenNext adapter, fully prerendered, with structured content files so their team edits services, solutions, industries and case studies without touching code.

Grounding and citation are not optional features. Every answer points back to the source document. For a nonprofit board or a public-sector manager, an assistant that cannot show its work is not usable, no matter how fluent it is. This constraint shaped the retrieval design from the start rather than being added after a procurement question.

Human review on the sensitive paths. Content that carries real consequence routes through a person before it lands. This is slower and it is the reason the system is allowed near regulated work at all.

Productised, not bespoke. Building one repeatable solution with configuration rather than a bespoke system per client is what makes the economics work for an MSP. It also means the security review, the governance material and the training happen once.

A real engineering fix, worth naming. Deploying Next.js to Cloudflare Workers through OpenNext, we hit a failure where the default cache configuration made every request a miss, which caused Next.js to schedule a background revalidation that never settles in the Workers runtime — requests hung for roughly thirty-four seconds and were then cancelled, with no response ever sent. The fix was to configure the incremental cache against the read-only static-assets cache, since every page is prerendered. It is a small detail and an honest illustration of the work: the difference between a deployment that appears to succeed and one that actually serves traffic is usually a problem like this one.

The partnership is ongoing. Great Lakes IT can now hold an AI conversation with a client and follow it with something deliverable, supported and governed, without hiring an AI team or handing the relationship to a third party the client has never met.

For companies considering the same shape: if you are an MSP, an agency or a consultancy whose clients have started asking about AI, this is a well-trodden arrangement and we are comfortable operating inside it. We are equally comfortable being invisible in it.