What startups should know about how AI providers handle their data, performance, and costs.
Across Canada, businesses are being encouraged to embrace AI, with no shortage of tools promising to make work faster, easier, and cheaper.
LSP says obstacles to AI adoption have less to do with AI itself and more with “fragmented systems, lack of integration, and governance maturity” within companies.
But before committing to something new, there are important questions to consider, said Chandrashekar Lalapet Srinivas Prasanna, known as LSP, managing director of Zoho Canada, the Canadian arm of the global software firm Zoho Corporation.
Where will company and customer data be processed? Will the tool work with the software your business already uses? Who will be allowed to access the information it draws on? And will that price still make sense when hundreds of employees rely on it every day?
For LSP, the answers get at what it really takes to put AI to work—and depend on both the SaaS provider and the business adopting the tool. The provider’s choices can determine where information is processed and how much the service costs as more people use it. The business, meanwhile, needs connected systems and clear rules governing its data.
So, what needs to happen inside the company, and what should founders expect from the tools they buy?
Inside the business
Many of the obstacles LSP sees have little to do with the AI itself. Instead, he points to “fragmented systems, lack of integration, and governance maturity” within companies.
Fragmented systems leave information spread across tools that don’t necessarily communicate, giving AI only part of the picture it needs to do its job, while weak governance blurs who can access data and for how long. Those breakdowns often surface once a tool moves from testing (a few prompts, a sandbox dataset, or maybe a demo workflow) to daily use.
For example, when a sales team pilots an AI assistant to summarize calls, it can work in trial mode. But in daily use, the assistant needs to pull customer data from the CRM, push notes into ticketing, respect role-based access, and comply with retention policies, said LSP.
Layering AI onto a setup that isn’t properly integrated can create problems once the tool begins handling real work, including “data leakage,” “shadow AI,” and “hallucination-driven errors.” That can mean exposing sensitive customer information, employees turning to unapproved tools, or confident but inaccurate answers making their way into company decisions.
He also warned of data “non-compliance” when information is handled in ways that breach regulatory requirements, and “high failure rates” when weak data and disconnected systems prevent AI projects from working reliably or expanding beyond a trial.
Zoho’s own approach offers an example of how a provider can reduce integration problems. The company offers a cloud suite of more than 60 business software tools, and builds its AI into the broader platform, rather than as a separate product. That enables the technology to work across connected applications, data, and permissions.
Choosing a provider
How an AI system is built determines whether it becomes a strategic advantage or “a recurring operational headache,” said LSP. AI built on third-party models or rented infrastructure can inherit costs and be passed on to customers. Architecture also dictates exposure: “A vertically integrated provider reduces the number of external dependencies, which reduces the attack surface and improves security and reliability.”
“Additionally, a vendor that doesn’t control its own compute means scaling can be subject to someone else’s capacity,” he said. “That’s why some AI tools slow down or degrade quality as usage grows.”
Zoho’s strategy has been to bring more of those pieces in-house. It develops its own applications and AI models, and operates the platforms and data centres that run them, giving it more control over costs and performance as usage grows.
LSP’s checklist for choosing a SaaS provider:
- Follow the data. Where will it be stored, can it leave Canada, and will it be used to train the provider’s models? Look for answers in contracts, technical evidence, and certifications.
- Test the price at scale. Is AI included, charged per request, or tied to computing use? An inexpensive trial may become expensive as usage grows. Instead, look for vendors who own their tech stack or have long-term cost guarantees.
- Check who owns what. Who supplies the models, servers, and other technology? Outside dependencies can affect how much control the provider has over security, reliability, performance, and pricing.
- Confirm the fit. Will the tool connect to existing systems and follow access rules? It’s easier to address gaps before a trial expands.
The latest addition is Nathu La, an in-house server developed with Intel to support virtualization, high-performance computing, storage, and AI inference (the computing required each time an AI model responds to a request). It draws on principles from the Open Compute Project, with parts that are easier to maintain or replace, and cooling intended to reduce energy use. According to the company, Nathu La performs comparably to similar servers, while using 12 to 18 percent less power and costing 20 to 30 percent less to own and operate.
Those savings, in turn, can reduce AI inference costs—which can be relatively minor when a handful of employees are testing a feature but add up quickly once the same tool is adopted across a company.
Zoho also works to control the cost by matching the model to the job, using smaller models for specific tasks rather than turning to a much larger model for every request. The company builds AI “with usefulness in mind,” said LSP.
When choosing a provider, LSP advises customers to ask if pricing is tied to tokens, API calls, or compute spikes. “If the answer is ‘it depends,’ expect pricing challenges,” he said.
Where will your data live?
In 2023, Zoho opened data centres in Toronto and Montréal, giving customers the option to keep their data storage and processing in the country. LSP said the company offers “region-locked processing” and “explicit residency guarantees,” backed by contracts and certifications. He added that Zoho’s general AI models are “not trained on consumer data and do not retain customer information.”
Those details matter because selling software in Canada doesn’t necessarily mean the information entered will stay here. LSP recommends potential buyers request a data-flow diagram showing where their data will be stored, processed, cached, and backed up. If any component sits outside the claimed region, “residency is not guaranteed,” he said.
LSP believes that as more companies adopt AI, they will expect clearer answers about who controls the tools, where their information is kept, and how the technology fits into their existing work.
“Canadian organizations will want trusted, sovereign, workflow-embedded AI,” he said.
PRESENTED BY
Your life’s work, powered by our life’s work. Learn more
Feature image courtesy Unsplash. Photo by Adi Goldstein.