AI operations: from experiments to real workflows

Many Canadian organizations now have a collection of AI experiments: a chatbot pilot, a summarization tool one team loves, a proof of concept that impressed the leadership group. Far fewer have AI embedded in the way work actually gets done, with clear ownership, predictable quality and measurable results. The distance between those two states is the work of AI operations.

What we mean by AI operations

The term AIOps is widely used for applying AI to IT infrastructure monitoring, such as detecting anomalies in logs or predicting outages. That is a legitimate discipline, and IT operations is one place where our approach applies. But we mean something broader: the design and ongoing operation of business workflows in which people, AI systems, data, existing software and governance rules each have a defined role.

In that sense, AI operations applies to finance, legal, HR, sales, support and procurement as much as to IT. It is where trained people and well-informed AI systems come together.

Why pilots stall

Pilots rarely fail because the model was not capable enough. They stall for organizational reasons that are predictable and fixable:

• No owner. The pilot belonged to an innovation team or an enthusiastic individual. Nobody in the business is accountable for its results or its upkeep.
• No workflow change. AI was added alongside the existing process rather than into it. People now have one more tool to open, and no reason to change habits.
• No information layer. The system worked on a curated demo dataset, then struggled with the scattered, conflicting and permission-restricted information of real operations.
• No evaluation. Nobody defined what good looks like, so nobody can show it is working, and one memorable mistake ends the project.
• Unclear risk. Legal, privacy or security questions were raised late and never resolved, so the pilot sits in limbo.

How to pick the first workflow

The best candidates are not necessarily the most exciting. Look for work that scores well on most of these criteria:

Redesigning the workflow, step by step

1. Map the current work. Document how the workflow really runs today, including informal workarounds, handoffs, the systems involved and where time goes. The written procedure is often not the real one.
2. Decide what AI does and what people do. Assign each step deliberately. AI might gather information, draft, classify or check. People might decide, approve, handle exceptions and own relationships.
3. Set human review and escalation points. Define where a person must review before work moves on, what triggers escalation, and who receives it. Higher-stakes steps need tighter review.
4. Integrate with existing systems. AI that requires copying and pasting between tools rarely survives. Connect it to where the work lives, with permissions that mirror what the human role can access.
5. Build in measurement. Decide in advance what you will track: quality, cycle time, rework, escalation rates, user feedback. Log enough to investigate problems.
6. Write the operating rules. Who owns the workflow, how changes are approved, how errors are reported and corrected, and when the design is reviewed. Policy should show up as actual permissions, logging and escalation paths, not just a document.

Redesign is not a one-time event. Once a workflow is live, it needs operating like any other part of the business. Models are updated by their providers, source information changes, and people find new edge cases. A workflow that performed well at launch can drift quietly. Schedule regular reviews of quality measures and escalations, re-run evaluations after any change to the model, the instructions or the underlying data, and treat user feedback as the early warning system it is. The owner should be able to answer, at any point, how the workflow is performing and what changed recently.

Finally, bring the people doing the work into the design. They know where the real exceptions are, which steps carry hidden judgment and what would make the new process easier than the old one. Workflows designed without them tend to be quietly worked around.

Illustrative patterns by function

These are common patterns, not case studies. The right design depends on your systems, data and risk tolerance.

• Finance. AI drafts variance commentary from reconciled figures and flags unusual items; analysts review, adjust and approve before anything is reported.
• Legal. AI compares incoming contracts against standard positions and highlights deviations; lawyers decide what to negotiate.
• HR. AI answers employee policy questions with citations to current policy and routes sensitive matters to a person. Decisions about individuals stay with people.
• Sales and revenue operations. AI prepares account briefs from CRM history and public information, and drafts follow-ups that representatives edit and send.
• Support. AI classifies and summarizes incoming tickets, suggests responses from the knowledge base, and escalates by defined rules.
• Procurement. AI extracts key terms from supplier documents and checks them against policy, flagging gaps for a buyer to resolve.
• IT operations. AI triages service requests, gathers diagnostic context and proposes next steps; technicians approve changes to systems.

Agentic workflows and levels of autonomy

Agentic workflows use AI systems that plan and carry out several steps, such as looking up records, drafting a response and updating a system. They can take on more of a workflow than a single prompt can, but autonomy should be matched to risk. A useful way to think about it:

1. Assist. AI suggests; a person does the work.
2. Draft. AI prepares the work; a person reviews and completes it.
3. Act with approval. AI carries out the steps; a person approves before anything takes effect.
4. Act and report. AI completes routine cases within defined limits and logs them for review; exceptions escalate.

Most workflows should start at the first or second level and move up only as evaluation shows consistent reliability. Decisions with significant effects on people deserve particular care. In Quebec, Law 25 includes transparency obligations for decisions made exclusively by automated processing, and federal institutions work under the Directive on Automated Decision-Making. Confirm the requirements that apply to you with counsel, and see our responsible AI page for how oversight can be designed in.

How Trained helps

Trained’s AI operations work takes promising experiments, or a workflow that simply needs to improve, and redesigns it so people, AI, data and systems work together with clear ownership, review and measurement. We stay involved through evaluation and improvement. If you have pilots that have not become part of the work, start a conversation.