Understanding Your AI Opportunity: Where Should Your Organization Actually Use AI?

Expert Series features guest perspectives from leaders across Nova Scotia’s tech community, sharing practical insights, lessons learned and real-world expertise through the Digital Nova Scotia platform.

In this edition, Marta Kryven, Assistant Professor at Dalhousie University and founder & chief scientist at Northlight Intelligence, explores how organizations can identify where AI will create measurable value — and where automation may introduce more cost and risk than it saves.

Throughout 2026, Meta actively experimented with replacing human engineering effort with AI. However, finding that more AI-generated code came with major increases in technical incidents and staff time required to fix them, in August 2026 it announced suspending AI-driven restructuring, acknowledging that the technology was not ready [1]. Canadian business has cautionary case studies of its own: in Moffat v. AirCanada 2024 [2], the company was held liable for its web chatbot misleading a customer about bereavement discounts. While the damages awarded to the customer by British Columbia’s Civil Resolution Tribunal were just about CAD800, the reputational costs of such incidents are harder to assess, and potentially more impactful. Many Canadian companies are feeling cautious about AI adoption, even as most Canadian businesses face much less dramatic problems. In 2026, Statistics Canada found that about 19% of Canadian businesses reported using AI to produce goods or deliver services [3]. Among those that did, a safe and tested use-case of data analytics was the most common application. It appears that the key question for most companies considering AI use is knowing how exactly it can benefit them. Many companies must introduce AI into existing legacy and interdependent workflows, where changing one process can affect others and create unexpected failures.Fortunately, there is a way to work around these issues to boost productivity by thinking about AI strategically. The key is asking “Where in the existing process can there be a measurable benefit from AI tech?” rather than simply asking “Could AI do this?” The answer to the second question is most likely affirmative. The answer to the first question determines whether a given AI deployment will make or lose money [4]. A practical way to make this decision is to separate potential AI use-cases into three categories:

1. Tasks with well-defined and predictable mappings between inputs and outputs,
2. Tasks with likely edge-cases that require oversight, and
3. Tasks where output can be ambiguous, and quality is hard to verify.

Tasks in the first category are good candidates for straightforward automation; the second can benefit from AI only if the cost of appropriate oversight is less than the revenue gained by human time-saving; and the third are better left human-led in the foreseeable future.Automate predictable, easily verifiable workRoutine data formatting and extraction, data transformations between systems, assembling recurring reports, formatting actionable meeting notes, summarizing vast amounts of spreadsheet information are examples of tasks where input-output mapping is predictable, and AI can be immensely helpful. Importantly, AI services have their own costs, and vendor-supplied data analytics tools are not guaranteed to work better than a prompt written in-house. The key to deciding whether automating each task is useful is measuring actual costs saved:Net value = employee cost without AI - employee cost with AI − ongoing AI cost.Fortunately, the learning curve to modern AI tools is undemanding compared to quirky enterprise software, and many employees may already be using such tools in private. Automation with oversightImagine using AI to summarize customer communications. People express themselves differently, for reasons that include, but go well beyond, cognitive styles, sentiment, social circumstances and cultural assumptions. Even two employees can sometimes disagree about classifying human-produced text into a narrow set of categories [5]. The common issue is that such data is open-ended, not fully specified by past examples: an automated system that makes no mistakes during evaluation may still fail on a real-world edge case, as happened in Moffat v. AirCanada. To prevent this, such cases call for a well-defined measurable oversight:Net value = employee cost without AI - employee cost with AI − ongoing AI cost - oversight costOversight can include an extra AI layer to determine when a case should be directed to a human, regular human oversight of a random selection of AI outputs, intermediate quality metrics, or any combination of the above. Its cost should also consider the expected cost of recovering from mistakes when they happen, which determines how rigorous the oversight process should be.Keep human-ledFor many creative tasks or novel strategic problems, it may initially seem like AI can help generate ideas. However, such first impressions are often misleading: an AI-generated creative output to a request is only surprising the first time it is produced. Yet, AI models tend to give the same response again and again to subsequent similar requests [6]. For the time being, humans have a comparative advantage over AI under novel or changing conditions, for instance, as documented in Tetlock's work on superforecasters [7]. The practical recommendation is to keep open-ended tasks human-led whenever decisions are financially consequential or creative quality matters. Aside from constrained response diversity, quality monitoring is a major issue: it can be hard for human reviewers to detect poor output to open-ended requirements, often leading people to accept superficially plausible content that lacks depth.The most useful AI strategy, then, is to identify where automation creates measurable value without introducing excess risk. This means starting with mapping existing workflows, measuring where employee time is spent, automating predictable tasks before moving to tasks that require oversight and quality monitoring. As those systems prove their value, organizations can strategically expand AI use rather than adopting technology for its own sake. [1] https://www.reuters.com/investigations/mark-zuckerberg-had-bold-plan-replace-meta-staff-with-ai-heres-how-it-imploded-2026-08-26/[2] https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html[3] https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm[4] https://finance.yahoo.com/news/mit-report-95-generative-ai-105412686.html[5] https://en.wikipedia.org/wiki/Cohen%27s_kappa[6] https://www.nature.com/articles/s41598-025-25157-3[7] https://journals.sagepub.com/doi/10.1177/17456916231185339