How Should Business Analysts in Canada Write Agentic AI Requirements in 2026?

Most business analysts in Canada are still writing AI requirements the way they wrote requirements for a chatbot: define inputs, define outputs, ship it. That approach breaks the moment an AI agent is allowed to take actions on its own, such as approving a refund, rescheduling a shipment, or updating a CRM record without a human clicking 'confirm.' The direct answer: agentic AI requirements need a fourth dimension that traditional BA documents never asked for — authority. A requirement isn't complete until it states exactly what the agent is allowed to decide, what it must escalate, and how that decision can be undone.

What is the Concept

Agentic AI refers to systems that don't just generate text or predictions but take multi-step actions toward a goal — querying a database, calling an API, sending a communication, or triggering a workflow — with limited or no human approval at each step. For a business analyst, this shifts the job from 'document what the system should output' to 'document what the system is permitted to do, under what conditions, and what happens when it's wrong.'

Traditional BRDs (Business Requirements Documents) capture functional requirements, business rules, and acceptance criteria. Agentic AI requirements add three new categories: autonomy boundaries (what decisions the agent can make unsupervised), escalation triggers (what conditions force a human handoff), and rollback conditions (how an incorrect autonomous action gets reversed cleanly). Skipping any one of these three is the single most common reason agentic AI pilots stall in Canadian enterprises after the proof-of-concept stage.

Why It Matters in Canada (2025–2026 Context)

Canadian financial institutions and insurers — think RBC, TD, and Manulife-scale organizations, but also mid-market firms in Toronto, Calgary, and Vancouver — are under OSFI and provincial privacy expectations that make 'the AI just did it' an unacceptable answer during an audit. At the same time, Bill C-27's AI and Data Act discussions have pushed procurement teams to ask vendors and internal teams for documented decision boundaries before greenlighting any AI system that touches customer data or money. A BA who cannot produce a clear autonomy matrix is now a procurement blocker, not just a documentation gap.

The economics reinforce the urgency. A mid-sized Canadian logistics or insurance firm running a poorly scoped agentic pilot commonly burns CAD 40,000 to CAD 90,000 in rework once legal or compliance teams demand retroactive guardrails — money that a properly scoped requirements phase would have saved outright. Firms in Ottawa's public-sector-adjacent tech corridor and Alberta's energy services sector are already building this into RFP requirements, asking vendors directly: 'show us the escalation matrix,' not just 'show us the demo.'

How AI Is Changing This

Here is the contrarian view: AI is not making the business analyst role smaller — it's making it the single most important role in the AI project team, above the data scientist. A model can be retrained; a poorly scoped authority boundary can trigger a regulatory incident, a customer-facing error, or a contract breach that no amount of retraining fixes after the fact. BAs who treat agentic AI requirements as 'just another integration spec' are undervaluing the one skill AI cannot replace: judgment about where human oversight is non-negotiable.

The non-obvious idea most Canadian teams miss: agentic AI requirements should be written as a decision tree with explicit 'stop' nodes before a single line of code is written, not layered on afterward as a compliance patch. Writing the stop conditions first — before the happy path — produces cleaner, cheaper systems, because engineering teams build guardrails natively instead of bolting them on as exception handling six weeks before launch.

Real-World Examples

A Toronto-based fintech piloting an AI agent for small-business loan pre-qualification initially let the agent auto-decline applications below a credit threshold. The BA team's second-pass requirement introduced a mandatory human review for any decline involving a business under two years old — a segment where the model's training data was thin. That single escalation rule, documented in the requirements phase, prevented a wave of disputed declines and an estimated CAD 15,000 in compliance review costs per quarter.

A Calgary energy services firm building an agent to auto-schedule field technician dispatches required the BA team to define a rollback window: any dispatch made by the agent could be reversed by a supervisor within 30 minutes without penalty to the technician's schedule. Without that requirement, the pilot's first version had no clean way to undo a bad dispatch, forcing manual firefighting that nearly cancelled the project.

Practical Insights / Actions

Use a named framework to structure agentic AI requirements gathering: the AGATE Framework — Autonomy scope (what the agent decides alone), Guardrails (hard limits it cannot cross regardless of confidence score), Auditability (what gets logged for every autonomous action), Triggers (conditions that force human handoff), and Escalation path (who is notified, how fast, and through what channel). Every user story for an agentic feature should map to at least one AGATE category before it's considered ready for development.

The founder mistake to avoid: treating the agent's confidence score as a substitute for a business rule. A 92% confidence score is a model output, not a business decision — the BA's job is to define, in business terms, what confidence threshold is acceptable for which action, and that threshold should come from risk appetite discussions with finance and legal, not from the data science team alone. The hidden opportunity here is that Canadian firms that formalize this AGATE-style documentation early can reuse it across every future agentic project, turning what looks like compliance overhead into a reusable requirements template that shortens future project timelines by weeks.

Future Outlook

Expect Canadian regulators and enterprise procurement teams to formalize autonomy documentation as a standard deliverable by late 2026, similar to how privacy impact assessments became mandatory after PIPEDA enforcement tightened. Business analysts who build fluency in agentic requirements now — rather than waiting for a template to be handed down — will be the ones staffed on every AI initiative their organization runs over the next three years, simply because they're the only ones who can translate 'the agent should be smart' into a specification an engineering team and a compliance officer can both sign off on.

RP SoftTech works with Canadian mid-market and enterprise teams to translate agentic AI ambitions into implementable, audit-ready requirements — closing the gap between a promising AI demo and a system that survives procurement, compliance, and production traffic.

Conclusion

Agentic AI hasn't made the business analyst obsolete — it has made the BA the person who decides how much authority a machine is allowed to hold. The teams in Toronto, Calgary, Vancouver, and Ottawa that win with agentic AI in 2026 will be the ones whose BAs learned to write autonomy boundaries, escalation triggers, and rollback conditions as fluently as they once wrote functional specs.