Skip to main content

    AI agents for business: 2026 statistics, results and risks

    Most large Canadian firms say they use AI agents: but almost none can prove a return yet. Sourced 2026 figures on adoption, measured performance, failure rates, security and Canadian privacy law, plus our task scoring table.

    Written by Hamza Hai, Founder, Accountant & Project Manager

    Published 31 min read7,689 words25 primary sources citedHow we write our guides

    The short answer

    • KPMG Canada reported in May 2026 that 77 per cent of 306 Canadian executives at companies with $50 million or more in revenue use AI agents to assist with tasks, but only three per cent have achieved measurable returns on their AI investments.
    • IBM reported on 17 November 2026 that 42% of Canadian senior AI leaders surveyed say AI agents already initiate and execute actions autonomously, while 73% say governance, risk or compliance concerns delay AI-driven action.
    • In independent tests, the best AI agent completed 30% of realistic office tasks on its own (TheAgentCompany, 2025), and leading agents fell from about 58% success on single-turn business tasks to about 35% on multi-turn ones (CRMArena-Pro, 2025).
    • Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.
    • Canadian privacy law already applies to AI agents: the Privacy Commissioner says accountability rests with the organization, not the automated system, and Quebec's Law 25 gives people rights to an explanation and human review of decisions made exclusively by automated processing.

    AI agents for business are software that can take a goal, pick its own steps and act in your systems, not just answer questions. Searches for them have more than doubled in Canada this year, and the surveys are loud: KPMG reported in May 2026 that 77 per cent of large Canadian companies use agents in some form. The same survey found only three per cent can show measurable returns.

    This page collects the statistics on AI agents for business in 2026: what they are, who uses them, what they are used for, how well they actually perform in independent tests, why projects fail, the security and Canadian privacy rules that apply, and why the figures disagree. Every number has its source, date and sample, and we say which sources sell AI. We also add two tools of our own: a decision table that scores which tasks suit an agent today, and a readiness checklist for a 10 to 50 person Canadian business.

    If you want whole-economy AI numbers rather than agent numbers, start with our AI adoption in Canada statistics. If you want to know what an agent costs to build, our sibling guide on AI agent development cost covers that, and we link to it below.

    Get a free quote for your project

    What is an AI agent, and how is it different from a chatbot or RPA?

    An AI agent is software that uses a large language model to decide its own next steps and call tools, such as your email, CRM or accounting system, until a goal is met. A chatbot answers. An agent acts. That difference is what makes agents useful, and it is also what makes them riskier.

    Anthropic, which makes the Claude models, draws the line in a way we find practical. In its December 2024 engineering guide it describes workflows as "systems where LLMs and tools are orchestrated through predefined code paths" and agents as "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The same guide recommends "finding the simplest solution possible" and warns that the autonomy of agents "means higher costs, and the potential for compounding errors."

    Four terms get mixed up in vendor marketing. Here is what each one means in one sentence:

    • Chatbot: software that replies to a person's message, usually from a script or a knowledge base, and does not change anything in your systems on its own.
    • Robotic process automation (RPA): software that repeats a fixed sequence of clicks and keystrokes in other applications, exactly the same way every time.
    • Workflow automation: rules that move data between systems when a trigger fires ("when an invoice arrives, create a bill in QuickBooks"), with no judgment involved.
    • AI agent: a language model given a goal, a set of tools and permission to choose which tool to use next, so the path changes from one case to the next.
    • Agentic AI: the umbrella term for systems built from one or more agents, including multi-agent systems where agents pass work to each other.
    ChatbotRPAWorkflow automationAI agent
    Decides its own next stepNoNoNoYes
    Changes data in your systemsRarelyYes, by fixed scriptYes, by fixed rulesYes, by its own choice
    Handles messy or new inputsPartlyPoorlyPoorlyOften, not always
    Same input gives same outputUsuallyYesYesNot guaranteed
    Typical failureWrong answerBreaks when a screen changesBreaks on an unexpected caseWrong action taken with confidence
    Easy to auditYesYesYesOnly if logging is built in

    The last two rows matter most for a business. A rule-based automation fails loudly when it meets something new. An agent can fail quietly, by doing the wrong thing in a way that looks right. That is why the rest of this page keeps returning to two questions: how often agents get tasks right, and what it costs when they do not.

    Agent washing

    Not everything sold as an agent is one. Gartner calls the rebranding of assistants, RPA and chatbots as agentic "agent washing". In June 2025 Gartner estimated that only about 130 of the thousands of agentic AI vendors are real, as reported by W.Media. If a product cannot choose between tools on its own, it is a chatbot or an automation with a new label, which is not a bad thing, but it should be priced and judged as one.

    AI agents for business: the key statistics for 2026

    The headline numbers on AI agents for business are high on intent and low on proven results. Most large Canadian firms say they use or are testing agents, but very few can show a measured return yet. The table below lists every figure on this page with its source, date and sample, and says whether the source sells AI products or services.

    FigureSource and dateSampleWho published it
    77% of Canadian organizations already use AI agents to assist with tasksKPMG Canada, Global AI Pulse, May 2026 (fieldwork 19 Feb to 27 Mar 2026)306 Canadian executives, firms with $50M+ revenueConsultancy
    3% of Canadian organizations have achieved measurable returns on AI investmentsKPMG Canada, May 2026Same 306Consultancy
    27% had deployed agentic AI with active use casesKPMG Canada, April 2025 (fieldwork 28 Feb to 5 Mar 2025)252 Canadian businessesConsultancy
    42% of Canadian senior AI leaders say agents already initiate and execute actions autonomouslyIBM Institute for Business Value, 17 Nov 2026 (fieldwork Apr to Jun 2026)67 Canadian respondents in a global 1,667Technology vendor
    61% of CEOs say they are actively adopting AI agentsIBM CEO Study, 6 May 2025 (fieldwork Feb to Apr 2025)2,000 CEOs, 33 countriesTechnology vendor
    81% of leaders expect agents to be moderately or extensively integrated into AI strategy in 12 to 18 monthsMicrosoft Work Trend Index, 23 Apr 2025 (fieldwork 6 Feb to 24 Mar 2025)31,000 workers in 31 marketsTechnology vendor
    14% have begun implementing agents; 2% have fully scaled deploymentCapgemini Research Institute, 16 Jul 2025 (fieldwork Apr 2025)1,500 executives at firms with $1B+ revenueConsultancy
    About 51% of respondents use agents in productionLangChain, State of AI Agents (2024)1,300+ professionals, 60% in technologyAgent software vendor
    19% reported significant investment in agentic AIGartner poll, January 2025, reported June 20253,412 webinar attendeesResearch firm
    Over 40% of agentic AI projects will be cancelled by end of 2027Gartner prediction, 25 Jun 2025Analyst forecast, not a surveyResearch firm
    Best agent completed 30% of realistic office tasks autonomouslyTheAgentCompany benchmark, arXiv, revised 10 Sep 2025Simulated software companyAcademic researchers
    Leading agents about 58% single-turn, about 35% multi-turn success on business tasksCRMArena-Pro, arXiv, 24 May 202519 expert-validated sales, service and quoting tasksResearchers
    Agent hijacking success rose from 57% to 80% with 25 attempts per attackNIST CAISI technical blog, 17 Jan 2025AgentDojo test scenariosGovernment standards body
    13% of organizations reported breaches of AI models or applications; 97% of those lacked AI access controlsIBM Cost of a Data Breach, 30 Jul 2025600 breached organizationsTechnology vendor (Ponemon research)

    Read the right-hand column before you quote any of these. A survey of executives at large firms run by a company that sells AI consulting is still useful, but it is not the same as a benchmark that measures whether an agent actually finished the job. We separate the two throughout this page.

    How many Canadian businesses use AI agents?

    Among large Canadian companies, most now say they use AI agents in some form. KPMG Canada reported in May 2026 that 77 per cent of 306 Canadian executives surveyed are already using AI agents to assist with tasks such as knowledge sharing between departments. A year earlier, KPMG's April 2025 survey of 252 Canadian businesses found 27 per cent had deployed agentic AI with active use cases.

    KPMG Canada surveySpring 2025Spring 2026
    Sample252 businesses, fieldwork 28 Feb to 5 Mar 2025306 executives, fieldwork 19 Feb to 27 Mar 2026
    Deployed or using agents27% deployed with active use cases77% using agents to assist with tasks
    Experimenting or piloting35% experimenting with pilots; 29% exploring37% exploring or piloting
    Deploying at scaleNot asked in this form43% deploying, scaling or orchestrating agents across multiple workflows
    Multi-agent systemsNot asked20% developing or implementing
    Not exploring at all1%Not published

    Do not read the jump from 27% to 77% as a tripling in one year. The two questions are worded differently. "Deployed agentic AI and have active use cases" is a high bar. "Using agents to assist with tasks such as knowledge sharing" is a low one, and it would count a staff member using an agent feature inside an office suite. The 2026 breakdown is the better guide: 37% are still exploring or piloting, and 43% say they are deploying across multiple workflows.

    The newest Canadian figure: IBM, 17 November 2026

    IBM's Institute for Business Value published a study of Chief AI Officers and equivalent leaders two days before this page went live. Of the Canadian respondents, 42% said AI agents already initiate and execute actions autonomously across workflows, 73% said governance, risk or compliance considerations delay AI-driven action, and 72% named scaling successful pilots as the point where transformation stalls. Globally the figure for autonomous agent action was 48%. The Canadian sample was 67 people, so treat the Canadian percentages as directional, not precise.

    Who these surveys describe

    Every Canadian agent survey we found describes large organizations. KPMG's 2026 Canadian respondents all reported annual revenue of $50 million or more: 28 per cent at $50 million to $99 million and the rest above. KPMG's 2025 sample also started at $50 million in revenue. IBM surveyed senior AI leaders, a job title most small firms do not have. No survey we found measures AI agent use among Canadian businesses with 10 to 50 staff.

    Statistics Canada does not ask about agents as such. The closest measure in its business AI survey is the share of AI-using businesses that run virtual agents or chatbots, which we cover with the rest of its data in our AI adoption statistics page linked in the introduction. That page is the place for whole-economy numbers, including small firms. This page stays with agents.

    What this means for a smaller business

    If you run a 20-person firm, the 77% figure does not describe your competitors. It describes companies with at least $50 million in revenue, most of which have IT teams, data platforms and legal departments. The more useful signal is the 3%: even those large firms can rarely prove that agents paid off yet. That gives a smaller business time to pick one narrow task and do it properly.

    How many businesses use AI agents worldwide?

    Worldwide, surveys put the share of organizations using AI agents anywhere from 14% to 61%, depending on who is asked and how "using" is defined. Broad business AI use is far higher: the Stanford AI Index 2025 reported that 78% of organizations used AI in 2024, up from 55% the year before. Agents are a much smaller slice of that.

    SurveyWho was askedHeadline agent figure
    IBM CEO Study (May 2025)2,000 CEOs in 33 countries and 24 industries61% actively adopting AI agents and preparing to implement them at scale
    Microsoft Work Trend Index (April 2025)31,000 workers in 31 markets, including 9,037 leaders46% of leaders say their companies use agents to fully automate workflows or processes
    Capgemini Research Institute (July 2025)1,500 executives at organizations with more than $1 billion in revenue, 14 countries14% have begun implementation; nearly a quarter have launched pilots; 2% have fully scaled
    LangChain (2024)1,300+ professionals, 51% at companies with fewer than 100 peopleAbout 51% using agents in production; 63% among companies with 100 to 2,000 staff
    Gartner webinar poll (January 2025)3,412 webinar attendees19% significant investment, 42% conservative, 8% none, 31% wait and see or unsure
    IBM CAIO study (November 2026)1,667 senior AI leaders in 33 geographies48% say agents already initiate and execute actions across workflows

    The Microsoft and IBM figures come from vendors whose products include agent platforms, and Microsoft notes that its report itself was produced with help from Copilot and agents built in Copilot Studio. LangChain sells agent development tools, and 60% of its respondents worked in technology. Capgemini sells AI consulting. None of that makes the numbers false. It means the people answering were more likely than average to be building or buying agents.

    Where intent outruns delivery

    Every large survey shows the same gap. Capgemini found 93% of business leaders believe scaling AI agents in the next 12 months will give them a competitive edge, while 2% have fully scaled. IBM's 2025 CEO study found only 25% of AI initiatives had delivered expected ROI over the previous few years, and only 16% had scaled enterprise-wide. Microsoft found 67% of leaders but only 40% of employees were familiar or extremely familiar with agents.

    Gartner expects the share of software with agents built in to keep rising. In its June 2025 release Gartner predicted that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from none in 2024. For a small business this is the most likely route to using agents: through features that arrive inside software you already pay for, not through a project you start.

    Have an idea like this?

    Get a free, no-obligation quote for your project. It takes two minutes and there is no pressure.

    Free, no obligation. Fixed quote within 24 hours. Phone is optional.

    What can an AI agent do for my business?

    Today, AI agents mostly research, summarize, sort and draft, and then hand the result to a person. The most common uses in surveys are knowledge work and customer service, not fully automated decisions. LangChain's survey found the top uses were research and summarization (58%), personal productivity or assistance (53.5%) and customer service (45.8%).

    Capgemini reported that agents are expected to see the most extensive adoption in customer service, IT and sales first, spreading into operations, R&D and marketing over the next three years. It also found only 15% of business processes are expected to run at semi-autonomous to fully autonomous levels within a year, rising to 25% by 2028. In Capgemini's words, most agents today "function as assistants or copilots".

    Anthropic's Economic Index (September 2025) gives a usage-based view rather than a survey. It found that 77% of business uses of its API involved automation patterns, compared with about 50% for people using the Claude.ai app, and that "directive" conversations, where users hand over a complete task, rose from 27% to 39% between late 2024 and its third report. Businesses that connect models to their own systems are already delegating whole tasks more often than individuals chatting with a model.

    Examples by business type

    These are the kinds of tasks we see as realistic for agents at small and mid-sized Canadian firms, with a person approving anything that leaves the building or moves money. They are examples, not results from a study.

    Business typeTask an agent can take onWhere a person stays in the loop
    Accounting or bookkeeping firmRead incoming receipts and bank lines, propose categories, flag anything unusualAccountant approves postings and anything over a set amount
    Professional services (law, engineering, consulting)Search past files and draft a first summary or precedent listProfessional checks every citation and fact before use
    Distributor or wholesalerRead emailed purchase orders, match to SKUs and draft the order in the systemStaff confirm orders with mismatches or new customers
    Clinic or health practiceDraft appointment reminders and sort non-urgent inbox messagesNo clinical decisions; staff review anything with health content
    Trades and constructionTurn site notes and photos into a draft daily report or change-order listProject lead approves before it goes to the client
    Retail or e-commerceAnswer order-status questions from live order data and draft return approvalsRefunds above a set limit and complaints go to a person
    Property managementSort maintenance requests by urgency and draft work ordersManager approves contractor dispatch and spending

    Notice the pattern: the agent reads and drafts, and a person decides. This is where the measured results below say agents are reliable enough today. It also matches what Canadian leaders told KPMG they are doing to manage risk: 41 per cent use human-in-the-loop models to validate AI outputs, and 38 per cent restrict autonomous decision-making in high-risk uses such as hiring.

    Do AI agents actually work? Measured results versus vendor claims

    Independent tests show AI agents complete a minority of realistic multi-step business tasks on their own, and do well on short, well-defined ones. That is a long way from the "digital colleague" in vendor marketing, and it is why the gap between adoption and returns is so wide.

    TestWhat it measuredResult
    TheAgentCompany (academic benchmark, revised September 2025)Agents working in a simulated small software company: browsing, writing code, running programs, messaging co-workersThe most competitive agent completed 30% of tasks autonomously
    CRMArena-Pro (May 2025)19 sales, service and quoting tasks in a CRM, business-to-business and business-to-consumerAbout 58% success in single-turn tasks, falling to about 35% in multi-turn conversations; over 83% on workflow execution; near-zero confidentiality awareness
    METR (March 2025)How long a task (measured in expert human time) agents can finish with 50% reliabilityNear 100% success on tasks that take humans under 4 minutes; under 10% on tasks over about 4 hours; the task length has doubled about every 7 months for 6 years
    NIST CAISI (January 2025)How often an agent can be hijacked by instructions hidden in dataAttack success rose from 11% to 81% with new attacks; average 57% rising to 80% after 25 attempts

    Two findings stand out for business use. First, success drops sharply when a task needs several back-and-forth steps: CRMArena-Pro measured a fall from about 58% to about 35%. Second, the agents in that benchmark showed near-zero awareness of confidentiality unless specifically prompted, and the researchers found that prompting for confidentiality often hurt task performance. An agent that handles customer records needs controls outside the model, not just a polite instruction.

    METR's work explains why agents look brilliant in a demo and disappointing in production. A demo is usually a short task. Real work is a chain of short tasks, and the chance of a mistake somewhere in the chain grows with its length. METR also notes that its trend line is rising fast, so these figures date quickly. We will revisit them when newer independent results are published.

    Returns: what companies report

    Self-reported returns are weak even among large firms. KPMG Canada reported in May 2026 that 70 per cent of Canadian organizations say AI is delivering meaningful business value, but only three per cent have achieved measurable returns on their AI investments, against eight per cent globally. IBM's November 2026 study found the strongest group of AI leaders said 63% of AI projects deliver measurable business results, compared with 43% for other leaders.

    Vendor claims: how to read them

    Vendor reports often lead with single-company stories. Microsoft's 2025 Work Trend Index, for example, describes a solo founder "on track to earn $2 million this year" and a five-person startup "boosting margins by 20%". These are anecdotes chosen by the publisher, not measured averages, and the report does not say how they were checked. Expectations are also not results: KPMG's 2025 survey found 63 per cent of respondents expected agentic AI to boost profitability by 5 to 15 per cent, which is a forecast by the buyer, not an outcome.

    • Ask whether a figure is a measurement, an expectation or an anecdote.
    • Ask who was surveyed: revenue size, job title, and how many were in Canada.
    • Ask whether the publisher sells the thing being measured.
    • Ask whether "agent" in the question would include a chatbot or a copilot feature.
    • Prefer benchmarks with published tasks and methods over survey percentages when judging whether an agent can do a job.

    How many AI agent projects fail, and why?

    Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. That forecast, from June 2025, is the most quoted failure figure on agents. It is a prediction, not a count of failed projects.

    The causes Gartner names line up with what surveys and benchmarks show:

    CauseEvidenceSource
    Accuracy below what the task needsPerformance quality was the top concern, more than twice as significant as cost and safety; 45.8% of small companies cited it against 22.4% for costLangChain, 2024
    Pilots that never scale72% of Canadian AI leaders name scaling successful pilots as where transformation stallsIBM, November 2026
    Weak governance95% of organizations had at least one agent-related incident in 12 months; 57% of infrastructure and operations leaders call their AI governance weakGartner survey of 298 leaders, June to July 2026, as reported by CASRAI on 16 November 2026
    Falling trustConfidence in fully autonomous agents dropped from 43% to 27% in a yearCapgemini, July 2025
    Infrastructure and data not ready80% lack mature AI infrastructure; fewer than one in five report high data readinessCapgemini, July 2025
    Skills and staff resistance31% of Canadian organizations report employee resistance to agents, against 16% globallyKPMG Canada, May 2026
    Cost growthAgentic systems "often trade latency and cost for better task performance"Anthropic, December 2024

    Gartner also warned in November 2026, according to the same CASRAI report, that 40% of infrastructure and operations organizations scaling AI will suffer a business-critical service disruption by 2028 without stronger governance, and that only 22% of those leaders have designated, agent-specific control planes today. We could not open Gartner's own release directly, so these figures are given as reported.

    The pattern behind the failures

    Most failed agent projects did not fail because the model was bad. They failed because nobody defined what "done" looked like, the agent was given access to more than it needed, or the cost of checking its work was never counted. Those three problems are cheap to fix at the start and expensive to fix after launch. The decision table and checklist further down are built to catch them.

    Have an idea like this?

    Get a free, no-obligation quote for your project. It takes two minutes and there is no pressure.

    Free, no obligation. Fixed quote within 24 hours. Phone is optional.

    What are the security risks of AI agents?

    The main security risk of an AI agent is prompt injection: instructions hidden in an email, web page or file that the agent reads and then obeys. Because an agent can take actions, a successful injection can send emails, move files or run code in your name. OWASP lists prompt injection first in its Top 10 for LLM applications for 2025.

    Prompt injection, defined in one sentence: an attack where text the AI reads as data contains instructions that the AI follows as if they came from you. Agent hijacking is the same attack aimed at an agent that has tools, so the result is an action rather than just a wrong answer.

    The US National Institute of Standards and Technology tested this in January 2025 through its Center for AI Standards and Innovation (CAISI). Using the AgentDojo framework, new attacks raised the success rate against a leading model from 11% for the strongest baseline attack to 81%. Across five injection tasks the average success rate was 57%, rising to 80% when each attack was tried 25 times. The test tasks included remote code execution and sending phishing emails to everyone the user had a meeting with.

    OWASP Top 10 for LLM Applications 2025 (selected)What it means for an agentControl to ask for
    LLM01 Prompt InjectionHidden instructions in what the agent reads take over what it doesTreat all outside content as untrusted; the agent cannot act on instructions found in data
    LLM02 Sensitive Information DisclosureThe agent reveals personal or confidential data in an answer or actionGive the agent only the records the task needs
    LLM06 Excessive AgencyThe agent has more permissions or tools than the task needsRead-only by default; human approval for writes, payments and outgoing messages
    LLM09 MisinformationThe agent states wrong facts with confidenceShow sources; check anything that goes to a customer
    LLM10 Unbounded ConsumptionLoops or abuse run up model costsSpending caps, rate limits and step limits per task

    OWASP also published a separate Top 10 for Agentic Applications for 2026, developed with more than 100 industry experts, researchers and practitioners, aimed at systems that "plan, act, and make decisions across complex workflows." Anyone building or buying an agent that touches customer data should ask the vendor how it addresses both lists.

    Breach data

    IBM's 2025 Cost of a Data Breach study, run by the Ponemon Institute across 600 organizations, found 13% reported breaches of AI models or applications, and 97% of those had no AI access controls in place. 60% of those AI-related incidents led to compromised data. One in five organizations reported a breach due to shadow AI, meaning AI tools staff use without approval, and high levels of shadow AI added an average of $670,000 (US) to breach costs. IBM Canada's 2026 report found 28% of Canadian organizations reported an AI-generated attack and put the average Canadian breach at CA$7.11 million.

    For a small business the lesson is simple. An approved agent with narrow permissions and a log is safer than staff pasting client data into whatever tool they found. Our privacy by design checklist covers the controls in plain terms.

    Do Canadian privacy laws apply to AI agents?

    Yes. An AI agent that handles personal information is covered by the same privacy law as any other software you use: PIPEDA federally, or the provincial private-sector law in Quebec, Alberta and British Columbia. Quebec already adds specific rules for decisions made exclusively by automated processing. This section is general information, not legal advice.

    PIPEDA and the Privacy Commissioner's AI principles

    The Office of the Privacy Commissioner of Canada published principles for generative AI on 7 December 2023. They state that generative AI tools "do not occupy a space outside of current legislative frameworks". Two lines apply directly to agents: organizations should "know that accountability for decisions rests with the organization, and not with any kind of automated system," and people affected by a significant decision should have an effective way to challenge it, including enough information to understand how it was reached and the chance to request human review.

    Quebec Law 25: automated decisions

    Quebec's private-sector privacy law, as amended by Law 25, already applies to decisions based exclusively on automated processing of personal information. As summarized by Raymond Chabot Grant Thornton, a business must inform the person no later than when it tells them of the decision; on request, it must give the personal information used and the reasons and principal factors behind the decision; the person can ask for incorrect information to be corrected; and they can submit observations and ask for the decision to be reviewed by a person. If an agent approves or rejects applications, sets prices for an individual or screens candidates without a human deciding, this rule is likely to apply to Quebec customers.

    Bill C-36: proposed, not law

    Bill C-36, the proposed Protecting Privacy and Consumer Data Act, was introduced in June 2026 to replace the private-sector part of PIPEDA. OpenParliament listed its status as second reading in the House as of 15 June 2026. It would add explanations of automated decisions on request and a route to human review. It is a proposal, its final wording may change and no in-force date has been set. We cover it in Bill C-36 explained.

    A Canadian ruling on chatbot liability

    Moffatt v. Air Canada, 2024 BCCRT 149, decided by the BC Civil Resolution Tribunal on 14 February 2024, is the Canadian case most often cited on this point. Air Canada's website chatbot told a customer he could apply for a bereavement fare after travel, which contradicted the airline's own policy page. The Tribunal held that Air Canada "remains responsible for all the information on its website," whether it comes from a static page or a chatbot, and ordered it to pay a total of $812.02 in damages, pre-judgment interest and fees, as summarized by the Canadian law firm DWW.

    The amount was small. The principle is not: your business answers for what your agent says and does. That is one more reason to keep a person between the agent and any promise to a customer. For advice on your own situation, speak to a privacy lawyer, and see the Privacy Commissioner of Canada or Quebec's Commission d'accès à l'information for official guidance.

    Why do AI agent statistics disagree so much?

    AI agent statistics disagree because each survey defines "agent" differently, asks a different kind of person, and samples a different size of company. A figure of 14% and a figure of 77% can both be accurate descriptions of different things.

    SourceFigureWhy it is high or low
    KPMG Canada 202677% using agentsBroad wording ("assist with tasks"); firms with $50M+ revenue
    IBM CEO Study 202561% actively adoptingCEOs answering about intent and preparation, not live use
    LangChain 2024About 51% in productionRespondents chose to answer an agent-tool vendor's survey; 60% work in technology
    IBM CAIO 2026 (Canada)42% say agents act autonomouslyAsked only senior AI leaders; 67 Canadian respondents
    KPMG Canada 202527% deployedStricter wording: deployed with active use cases
    Capgemini 202514% implementing, 2% fully scaledStrict stages; asked firms with $1B+ revenue
    • Definition: some surveys count any AI assistant feature as an agent. Others count only systems that take actions on their own. Gartner's "agent washing" warning applies to survey answers too.
    • Respondent: CEOs and AI leaders report strategy and intent. Engineers report what is running. Employees report what they use.
    • Company size: every agent survey we found sampled large companies. No Canadian source measures agent use at small firms.
    • Sample size: a national sub-sample of 67 people has a wide margin of error. Most vendor releases do not publish one.
    • Publisher: firms that sell AI products or consulting have a reason to publish high adoption numbers. We list who published each figure so you can weigh it.
    • Timing: agent capability changes fast. METR measured task length doubling about every 7 months, so a 2024 figure can describe a different technology from a 2026 one.

    If someone quotes you a single adoption figure to justify spending, ask for the question wording and the sample. If they cannot give you either, treat the number as marketing.

    Have an idea like this?

    Get a free, no-obligation quote for your project. It takes two minutes and there is no pressure.

    Free, no obligation. Fixed quote within 24 hours. Phone is optional.

    Which tasks suit an AI agent today? Our decision table

    A task suits an AI agent today when it repeats often, a mistake is cheap and easy to undo, and the data the agent needs sits in systems it can reach. We built the scoring below from the evidence on this page: benchmark success rates, the drop on multi-step tasks, and the security findings on agents with write access. It is our framework, not an industry standard.

    How the scoring works

    Score each task from 1 to 3 on four tests. Add the scores for a total out of 12.

    Test1 point2 points3 points
    Repeatability: how alike are the cases?Every case is differentSimilar, with frequent exceptionsSame shape most of the time
    Error cost: what does a mistake cost?Money, safety, legal or a client relationshipSome rework or an annoyed customerMinutes of staff time
    Data access: can the agent reach what it needs?Paper, memory or many disconnected toolsTwo or three systems, some manual exportOne or two systems with an API
    Reversibility: can you undo the result?No (payment sent, email delivered, record deleted)With effortYes, it is a draft or a reversible change

    Rule of thumb from the totals: 10 to 12, a good candidate for an agent with spot checks; 7 to 9, use an agent to draft and keep a person approving every result; 6 or below, use plain automation or leave it with people for now. Any task that scores 1 on error cost stays with a person to approve, whatever its total.

    TaskRepeatabilityError costData accessReversibilityTotalOur verdict
    Sort and tag incoming support emails333312Agent, with weekly spot checks
    Draft replies to common customer questions323311Agent drafts, person sends
    Summarize meeting notes or long documents333312Agent, reader checks key facts
    Match purchase orders to SKUs and draft orders322310Agent drafts, staff confirm mismatches
    Categorize expenses and bank lines323210Agent proposes, bookkeeper approves
    Research a prospect and draft an outreach note22239Agent drafts, person edits and sends
    Book, move or cancel appointments22228Agent with confirmation step
    Answer questions about a contract or policy21238Agent finds clauses, person answers
    Approve refunds or credits21317Person approves; rule-based automation for small fixed amounts
    Screen job applicants21227Not for an agent alone; legal risk under automated-decision rules
    Pay supplier invoices31217Rule-based workflow with human sign-off, not an agent
    Give tax, legal or clinical advice to clients11114Not for an agent

    A worked example: is an email-drafting agent worth the review time?

    The figures below are example inputs, not measured results. Swap in your own. Suppose a firm gets 400 routine customer emails a month and each takes a staff member 6 minutes to read and answer by hand. That is 2,400 minutes, or 40 hours a month.

    1. 1With an agent drafting every reply, assume a person still spends 1.5 minutes reviewing each draft: 400 x 1.5 = 600 minutes, or 10 hours.
    2. 2Assume 35% of drafts need a real fix. That rate is in line with the roughly 35% multi-turn success measured in CRMArena-Pro, turned around as a cautious error rate. Each fix takes 4 more minutes: 140 x 4 = 560 minutes, or about 9.3 hours.
    3. 3Total staff time with the agent: about 19.3 hours a month, against 40 hours without it. The saving is about 20.7 hours a month.
    4. 4At a loaded staff cost of $35 an hour (example figure), that is about $725 a month, or about $8,700 a year, before the agent's running and maintenance costs.
    5. 5Now change one input. If drafts are so unreliable that review takes 4 minutes each instead of 1.5, review alone is 1,600 minutes (26.7 hours), and with fixes the total passes 36 hours. The saving almost disappears.

    The lesson is that review time decides whether an agent pays off, more than the model does. Measure how long checking takes in a two-week trial before you commit. For what a build itself costs, see our separate guide to AI agent development cost; you can test the hours-saved side in the free build vs buy ROI calculator.

    Which AI agents are best for business?

    There is no single best AI agent for business. The right choice depends on where your data lives, because an agent is only as useful as the systems it can safely reach. For most small firms the best first agent is one built into software they already use, configured with narrow permissions.

    OptionGood fit whenWatch for
    Agent features inside software you already use (office suite, CRM, help desk, accounting)The task lives mostly inside that one productDefault permissions that reach more data than the task needs; per-seat or per-action pricing
    General assistant with connectorsStaff need help with research, drafting and summaries across toolsShadow AI: staff connecting accounts without approval
    No-code agent buildersA simple, low-risk task across two or three toolsHard to test and log properly; logic spread across settings screens
    Custom agent built on your systemsThe task crosses several systems, needs your business rules, or handles sensitive dataBuild and maintenance cost; needs clear success measures before work starts

    Questions to ask any vendor before you buy:

    1. 1What exactly can the agent do without a person approving it? Can we turn each action off?
    2. 2Which of our data can it see, and can we limit it to the records a task needs?
    3. 3Where is our data processed and stored, and is any of it used to train models?
    4. 4Is every action logged with the input that led to it, and can we export the log?
    5. 5How does it handle instructions hidden in emails, files or web pages it reads?
    6. 6What does it cost per user, per action or per task, and is there a spending cap?
    7. 7What is its measured accuracy on a task like ours, and can we trial it on our own data first?

    If a task crosses several systems or needs your own rules, a custom build may fit better than a product. Our AI development service page explains how we scope that, starting with a fixed written quote after a scoping call.

    AI agent readiness checklist for a 10 to 50 person Canadian business

    A small business is ready for an AI agent when it can name one task, measure it, give the agent access to only what it needs, and check its work. If you cannot tick most of the items below, fix those first. They pay off with or without an agent.

    Task and measurement

    • We have named one task, not a department, that we want help with.
    • We know how many times a month it happens and how long each one takes today.
    • We have written down what a correct result looks like, with five to ten real examples.
    • The task scores 7 or more in the decision table above, and does not score 1 on error cost without a person approving.

    Data and systems

    • The data the task needs sits in one or two systems, not in personal inboxes or paper files.
    • Those systems have an API or a supported connector, so the agent does not need a staff password.
    • We can give the agent its own account with read-only access by default.
    • We know which personal information the task involves and whether any customers are in Quebec.

    Control and accountability

    • A named person owns the agent's results and reviews a sample every week.
    • Every action the agent takes is logged with what it saw and what it did.
    • Anything that sends money, sends a message to a customer or deletes a record needs a person to approve it.
    • There is a spending cap and an off switch, and someone knows how to use both.
    • Our privacy policy and customer notices describe the automated processing honestly.
    • Staff know which AI tools are approved and that client data does not go into unapproved ones.

    Money and timing

    • We have estimated the hours saved after review time, not before.
    • We have checked whether public funding applies, in Canada AI funding for business.
    • We have budgeted for upkeep, not just the first version, and agreed when we will decide to stop if it does not work.

    A six-step pilot

    1. 1Pick the highest-scoring task from the decision table and write its success measure, for example "80% of drafts sent with no more than a one-line edit".
    2. 2Collect 30 to 50 real past cases, with personal details removed if you can, to test against.
    3. 3Run the agent in draft-only mode for two weeks. A person approves every output.
    4. 4Log review time and fix rate every day. Compare them with the manual baseline.
    5. 5Decide at the end of week two: widen, adjust or stop. Write the reason down.
    6. 6Only after a clean month, consider letting low-risk actions run without approval, with weekly spot checks.

    If your answer to most of the checklist is no, the right first project is often plain automation or getting your data into one place. Our guide to business process automation covers that route, and is AI worth it for a small business helps with the go or no-go decision.

    Have an idea like this?

    Get a free, no-obligation quote for your project. It takes two minutes and there is no pressure.

    Free, no obligation. Fixed quote within 24 hours. Phone is optional.

    How we compiled this

    Every figure on this page comes from a source we opened on 19 November 2026, written as the source gives it, with the period it covers and the sample where one was published. The full list is under Sources at the end of the page.

    • Canadian surveys: KPMG Canada's agentic AI releases of April 2025 and May 2026, and IBM Canada's release of 17 November 2026 on its Chief AI Officer study.
    • Global surveys: the IBM CEO Study (May 2025), Microsoft Work Trend Index (April 2025), Capgemini Research Institute (July 2025) and LangChain (2024).
    • Independent measurements: TheAgentCompany and CRMArena-Pro benchmark papers on arXiv, METR's task-length research, and NIST CAISI's agent hijacking evaluation.
    • Security and privacy: OWASP's 2025 LLM Top 10 and 2026 agentic Top 10, IBM's breach reports, the Privacy Commissioner of Canada's generative AI principles, Raymond Chabot Grant Thornton's summary of Law 25 on automated decisions, OpenParliament's Bill C-36 page and DWW's summary of Moffatt v. Air Canada.
    • Gartner: Gartner's site did not load for us. The June 2025 cancellation prediction and January 2025 poll are as reported by W.Media, and the November 2026 incident survey is as reported by CASRAI. We say so wherever we use them.
    • Left out: McKinsey's State of AI and PwC's AI agent survey, because we could not load the original pages on the day we checked; figures that blogs repeat without naming the origin; and any statistic we could not trace to the publisher.
    • Search demand: Canadian searches for "ai agents" rose to 14,800 a month in October 2026 from 6,600 in spring, and "what are ai agents" to 1,600 from 390, according to Google Ads data pulled through DataForSEO.

    We will update this page when KPMG, IBM, Statistics Canada or Gartner publish new agent figures, and when Bill C-36 moves past second reading.

    What to do with these numbers

    Start small, measure honestly, and keep a person in charge of anything that cannot be undone. The evidence says agents work best today on frequent, low-risk, reversible tasks, and that review time decides whether they pay off.

    1. 1Score your five most time-consuming tasks with the decision table.
    2. 2Run the readiness checklist on the top scorer.
    3. 3Put the hours saved after review into the ROI calculator.
    4. 4Read the cost guide before asking anyone for a price, so you know what drives it.
    5. 5Pilot in draft-only mode for two weeks and decide on the numbers.

    If you would like a second opinion, Four Cents will look at one process with you and tell you in writing whether an agent, plain automation or no change is the better answer. Book a free consultation. We are rated 5.0 on Google (24 reviews).

    Frequently asked questions

    What is an AI agent in simple terms?

    An AI agent is software that uses a large language model to decide its own next steps and use tools, such as email, a CRM or an accounting system, until a goal is met. A chatbot only replies. An agent takes actions, which is why it needs tighter permissions and logging.

    What can an AI agent do for my business?

    Today agents are most reliable at research, summarizing, sorting and drafting, with a person approving the result. LangChain's 2024 survey found the top uses were research and summarization (58%), personal productivity (53.5%) and customer service (45.8%). Tasks that move money or make legal decisions should keep a person in charge.

    Which AI agents are best for business?

    There is no single best agent. For most small firms, the best first agent is a feature inside software they already use, set to narrow permissions. A custom agent fits better when a task crosses several systems or needs your own business rules. Judge any option on permissions, logging, data location and measured accuracy on your own data.

    What is the 30% rule in AI?

    There is no official 30% rule. The phrase is used informally online, most often to mean AI handles about 70% of routine work while people keep about 30% for judgment and oversight. No standards body or regulator defines it, so treat it as a slogan, not a guideline. The decision table on this page is a more specific way to decide what to hand to an agent.

    How many Canadian businesses use AI agents?

    Among large firms, most. KPMG Canada reported in May 2026 that 77 per cent of 306 executives at companies with $50 million or more in revenue use AI agents to assist with tasks, up from 27 per cent who had deployed agentic AI in its 2025 survey, though the questions were worded differently. No survey measures agent use at small Canadian businesses.

    Why do so many AI agent projects fail?

    Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Surveys also point to accuracy below what the task needs, pilots that never scale and weak governance.

    Are AI agents safe to connect to company systems?

    Only with controls. NIST's January 2025 tests found hidden instructions could hijack agents in up to 80% of attempts when attacks were repeated. Give agents read-only access by default, require human approval for payments, outgoing messages and deletions, and log every action.

    Does PIPEDA apply to AI agents?

    Yes. Any agent that collects, uses or discloses personal information is covered by PIPEDA or the applicable provincial law. Quebec's Law 25 adds rights to be informed, get an explanation and request human review of decisions made exclusively by automated processing. This is general information, not legal advice.

    How much does an AI agent cost to build?

    It depends on the systems it must connect to, the controls it needs and how much testing the task requires. We cover the cost drivers and our published price ranges in a separate guide on AI agent development cost, rather than repeating them here.

    Sources

    Each figure and date above comes from one of these. The retrieved date is the day we last opened the source and confirmed it still says what we say it says.

    1. Canadian leaders expect agentic AI to reshape workforce (May 2026) KPMG Canada. Retrieved 2026-11-19.
    2. The agentic shift: AI's next phase (May 2026) KPMG Canada. Retrieved 2026-11-19.
    3. Canadian organizations turning to AI agents (April 2025) KPMG Canada. Retrieved 2026-11-19.
    4. New IBM study: 75% of Canadian AI leaders surveyed say their organizations risk falling behind competitors in AI transformation (17 November 2026) IBM Canada. Retrieved 2026-11-19.
    5. IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles (6 May 2025) IBM. Retrieved 2026-11-19.
    6. IBM Report: 13% of organizations reported breaches of AI models or applications (30 July 2025) IBM. Retrieved 2026-11-19.
    7. IBM Report: Canada's Data Breach Costs Hit Record High (29 July 2026) IBM Canada. Retrieved 2026-11-19.
    8. 2025 Work Trend Index: The Year the Frontier Firm Is Born (23 April 2025) Microsoft. Retrieved 2026-11-19.
    9. Rise of agentic AI: trust and human-AI collaboration (16 July 2025) Capgemini Research Institute. Retrieved 2026-11-19.
    10. State of AI Agents (2024 survey) LangChain. Retrieved 2026-11-19.
    11. Over 40 percent of agentic AI projects could face the axe by end of 2027: Gartner (June 2025) W.Media, reporting Gartner. Retrieved 2026-11-19.
    12. Gartner warns on runtime governance for AI agents (16 November 2026) CASRAI, reporting Gartner. Retrieved 2026-11-19.
    13. TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (v3, 10 September 2025) arXiv. Retrieved 2026-11-19.
    14. CRMArena-Pro: holistic assessment of LLM agents across diverse business scenarios (24 May 2025) arXiv. Retrieved 2026-11-19.
    15. Measuring AI Ability to Complete Long Tasks (19 March 2025) METR. Retrieved 2026-11-19.
    16. Technical blog: strengthening AI agent hijacking evaluations (17 January 2025) NIST. Retrieved 2026-11-19.
    17. OWASP Top 10 for LLM Applications 2025 OWASP Gen AI Security Project. Retrieved 2026-11-19.
    18. OWASP Top 10 for Agentic Applications for 2026 OWASP Gen AI Security Project. Retrieved 2026-11-19.
    19. Building effective agents (19 December 2024) Anthropic. Retrieved 2026-11-19.
    20. Anthropic Economic Index report, September 2025 Anthropic. Retrieved 2026-11-19.
    21. The 2025 AI Index Report Stanford HAI. Retrieved 2026-11-19.
    22. Principles for responsible, trustworthy and privacy-protective generative AI technologies (7 December 2023) Office of the Privacy Commissioner of Canada. Retrieved 2026-11-19.
    23. Law 25: the issue of automated decisions Raymond Chabot Grant Thornton. Retrieved 2026-11-19.
    24. Bill C-36 (45th Parliament, 1st session) OpenParliament.ca. Retrieved 2026-11-19.
    25. BC tribunal finds Air Canada liable for inaccurate advice given by website chatbot DWW Law. Retrieved 2026-11-19.

    Keep reading

    Cost & Planning

    AI Agent Development Cost: 2026 Price Guide

    A custom AI agent costs $8,000 to $30,000 CAD to build and, in our worked examples, about $14 to $553 CAD a month to run. Here is the arithmetic from published token prices, a build vs platform comparison, the hidden costs and how to get a fixed quote.

    Research & Data

    AI Adoption in Canada: Business Statistics 2026

    About one in five Canadian businesses now uses AI to make or deliver what it sells. Here are the sourced numbers on who, for what, what stops them and what it costs when data goes wrong.

    Cost & Planning

    Is AI Worth It for a Small Business in Canada?

    AI is worth it when it removes hours from a task you can measure, using data you already hold. Here is how to tell, what to count, what it costs to wait, and what to check before you hire anyone.

    Compliance

    Canada AI Funding for Business: 2026 Deadlines

    Three things changed in 2026: a national AI strategy, a $500 million BDC loan program, and $500 million more for regional AI funding with deadlines in November and January. Who qualifies, and what to do first.

    Guides

    business process automation

    Automation is not about replacing people. It is about removing the repetitive steps that steal their time. Here is how to spot, plan, and measure it.

    Compliance

    Bill C-36: Canada's New Privacy Bill Explained

    Canada's third attempt to replace PIPEDA was tabled on 15 June 2026. It is still a bill, but it would add regulator fines, deletion and portability rights, and explanations of automated decisions, most of which land in your software.

    Our services and prices

    Ready to get a number for your idea?

    Tell us what you want to build. We reply within two hours with a clear, fixed-scope quote. It is free and there is no pressure.

    Free, no obligation. Fixed quote within 24 hours. Phone is optional.