Agentic AI: when to use it, when to avoid it and how much control to hand over

Handing an AI agent the follow-up on overdue invoices, the triage of fraud alerts or the preparation of a claims file means opening your systems to it and letting it act on your behalf. In a regulated industry, where every action can affect a client, a regulated file or the safety of a team, one question needs an answer before any deployment: how much control should the agent have? Three guideposts help you answer it, from choosing the use case to setting the autonomy of each agent.
What sets an AI agent apart from an AI model?
Agentic AI refers to AI systems that carry out tasks and interact with an organization’s systems, going beyond content generation. Adoption is moving quickly, with nine in ten companies already exploring it in some form, according to BSI, the UK’s national standards body.
An AI model produces an answer or a recommendation that you then decide whether to use, whereas an agent pursues a goal by carrying out actions in your systems itself, with the tools and access you have granted it. Within an organization, the agent is part of a broader setup: models that read and draft, written and versioned skills that describe how a task is done, limited access to authorized systems, orchestration that coordinates the steps, and people who set the rules and approve sensitive decisions. All of an agent’s governance therefore starts with its scope, meaning what it is allowed to do and which systems it can reach.
When is agentic AI the right choice?
An agent delivers the most value when four conditions come together: high-volume repetitive work, clear written rules, available data and errors that are easy to detect and correct. The more of these boxes your use case checks, the stronger it is, which is why this check is worth doing before you choose a technology.
- High-volume repetitive work. An accounting team can receive hundreds of supplier invoices every week, and that volume is what justifies investing in an agent rather than having your teams handle each document by hand.
- Clear written rules. A travel expense policy, with its caps, required receipts and eligible categories, gives the agent a solid foundation to rely on, while rules that exist only in the head of an experienced employee leave it with nothing to go on.
- Available data. When the history of past files sits in personal inboxes or on paper, the agent has no way to use it, however good the model may be.
- Detectable, reversible errors. An email filed in the wrong processing queue can be spotted and moved in seconds, whereas a wire transfer sent to the wrong account is much harder to recover.
How can you tell whether your rules are documented well enough?
A good test is to ask whether a new hire could apply the procedure using only the existing documents, without having to ask a colleague. If the answer is no, the agent will run into the same obstacle, and your first project will be to put those rules in writing, which benefits your teams as much as your future agents.
What volume justifies an agent?
The threshold comes from weighing the time saved against the cost of an error. Volume works both ways, since it makes the agent cost-effective but also amplifies its flaws: at a thousand files a day, a small systematic error quickly becomes a significant problem. That is why volume appears both among the conditions for success and among the criteria used to set an agent’s level of autonomy, as you will see further on.
AI agent or traditional automation: which should you choose?
When a process follows fixed, predictable rules, traditional automation or well-designed software does better than an agent, since it costs less, always produces the same result and is easier to verify. An agent makes more sense when the task meets the four conditions described above, and in practice, four situations call for a different answer.
- The process follows fixed rules. Calculating taxes, transferring a file every night or sending a payment reminder on a set date is a job for traditional automation, and adding an agent would introduce uncertainty with no gain at all.
- Human judgment or the relationship is at stake. Delivering bad news to a long-standing client, negotiating a contract renewal or supporting someone in distress are moments when trust is built or lost, and while an agent can prepare the meeting, a person stays in control.
- History and data are missing. With incomplete, scattered or poor-quality files, the agent produces results that look convincing but rest on nothing, so your best investment goes first into data collection and quality.
- The action is irreversible and its impact significant. Cancelling an insurance policy, ending someone’s employment or recalling a product from the market are acts whose errors cannot be undone, so a person should make the call.
How do the costs of the two approaches compare?
Costs point the same way, since giving an agent a task that a simple rule can handle means paying more for a less stable result: each run calls on a model whose cost varies with volume and whose answer can change slightly from one time to the next. For predictable tasks, coded rules and traditional automation therefore remain the most economical choice and the easiest to audit, and in short, when a single tool can solve your problem, conventional software is enough.
Can you combine both in the same process?
The two approaches work very well together: within the same process, the predictable steps can rely on coded rules while the agent gathers information, sorts files or writes a first draft. The choice is made step by step, when the solution is designed, based on what each step truly requires.
What does agentic AI look like in finance, insurance and manufacturing?
In regulated industries, the strongest use cases share one principle: the agent prepares the file, and the person decides. Preparation tasks, such as sorting, gathering information and checking that a file is complete, make a good starting point because they often meet the four conditions for success while leaving the decision with your teams.
- In finance, for fraud alerts. Agents sort the alerts, pull together the client’s history and document the file, so the analyst no longer has to search five different systems and can focus on the decision.
- In insurance, for claims. Agents classify incoming files, detect that a repair invoice or a police report is missing and alert the claims adjuster, so the file arrives complete, without back-and-forth with the policyholder.
- In manufacturing, for equipment anomalies. When a sensor flags abnormal vibration on a production line, agents suggest diagnoses, retrieve the maintenance history and bring it all together for the technician before they even reach the plant floor.
How do several agents work together?
To see this collaboration in action, take the fraud example again with a fictional scenario, a $4,800 transaction made abroad at 3 a.m. on the card of a client who has never travelled: this suspicious payment triggers the process, and three agents step in, each with its own role and access.
- The detection agent analyzes the transaction and can only access the transaction feed.
- The investigation agent gathers the client’s history, such as spending habits and travel notices, and has access to the client file.
- The compliance agent checks that the applicable rules are followed and has access to policies and regulations.
At the centre, the orchestrator coordinates the work and tracks steps, errors and approvals, so if the investigation agent cannot find the file, the orchestrator detects it and flags it. At the end of the chain, the analyst receives a complete file and decides whether or not to block the card.
You will notice that each agent has different access, and that choice is deliberate. By limiting each one to the data it needs, you reduce the risk of leaks and unintended use, a principle that belongs to the Zero Trust approach and becomes essential when you move to production.
How much control should you give an AI agent?
An agent’s autonomy is set on a four-level scale, and the right level depends on the risk of each task, which lets you place the dial differently from one step to the next:
- The agent recommends, and a person decides and acts.
- The agent prepares and asks for approval before acting.
- The agent acts, then a person reviews its work.
- The agent acts on its own, within a pre-authorized, monitored scope.
This dial reflects a trade-off: at the first level, you stay in control, but the process is slower and depends on your teams’ availability, while at the fourth level, you gain speed and volume, provided you trust the agent and the guardrails around it.
What changes most from one end of the scale to the other is the person’s role, since at the first level they do the work with the agent’s help, whereas at the fourth they supervise what the agent does. “These are two different skill sets, and your teams need to be prepared for them,” says Gabriel Mongeon.
What criteria should you use to set the level of autonomy?
The level of autonomy should stay proportional to the risk, and six criteria help you assess it:
- Reversibility of the task. A misfiled document is easy to refile, whereas a payment that has already gone out is much harder to cancel.
- Impact. Measure the real consequences of an agent error for your clients, your operations and your organization.
- Volume. The more often a task repeats, the more a systematic error multiplies, as discussed earlier.
- The agent’s proven reliability. An agent that got it right on 500 reviewed files deserves more trust than one deployed last week.
- Human and ethical impacts. Refusing a loan or insurance coverage directly affects someone’s life.
- Legal and contractual requirements that apply to your industry and to your agreements with clients.
Can several levels coexist within the same process?
Yes, and it is an effective way to spread the risk. Going back to the fraud example, the detection agent can flag a suspicious transaction on its own, while the decision to freeze the account stays with a person. In the same way, an agent can be allowed to prepare 100 refunds without being able to send them until a person approves them, so each step gets the level that matches its own risk.
How do you move an agent up the scale?
You can start at the lower levels and move up over time as the agent proves its reliability on reviewed files. The movement works in both directions, since you can also step down a level if the context changes, for example after a model update or a change to your rules. To make these decisions on solid ground, you need a record of what the agent has done, hence the importance of an action log, which lets you measure its reliability based on facts.
Which decisions should stay with humans?
Irreversible decisions that directly affect a person, their rights or their safety should stay in human hands, whatever level of autonomy the agent has been given. Each industry has its own rules, but the logic is the same from one sector to another: identify these sensitive decisions, then set your agents’ autonomy accordingly, based on your risk tolerance.
- In insurance, denying a claim. The agent builds the file, but a person is the one who says no to the policyholder.
- In finance, freezing an account. The agent can flag a suspicious transaction, and the analyst decides what happens next.
- In defence, granting access to controlled data. The agent should never be the one that opens the door.
- In manufacturing, authorizing equipment to return to service after an incident. Worker safety is at stake, and that decision belongs to a qualified person.
These examples are illustrations, since every organization has its own sensitive decisions, which are worth mapping out before you deploy a first agent.
Which frameworks apply to your industry?
The regulatory landscape is moving quickly, and financial institutions face a firm deadline in May 2027, while other sectors rely on established frameworks:
- Federally regulated financial institutions. OSFI Guideline E-23 on model risk management takes effect on May 1, 2027 for all federally regulated financial institutions, and its definition of a model explicitly includes AI and machine learning methods.
- Insurers and financial institutions also operating in Quebec. Quebec’s Autorité des marchés financiers published the final version of its guideline on the use of artificial intelligence in April 2026, which takes effect on May 1, 2027 and applies to authorized insurers, financial services cooperatives, trust companies and authorized deposit institutions, among others.
- Ontario’s public sector and its suppliers. In January 2026, Ontario’s Information and Privacy Commissioner and the Ontario Human Rights Commission released joint principles for the responsible use of AI, and although these principles are not legally binding and mainly target the broader public sector, they call for a human in the loop and for organizations to be able to explain how an AI system works, expectations that may be passed on to the companies serving these institutions.
- Defence. The Controlled Goods Program governs access to sensitive goods and technical data.
- Manufacturing. ISO 9001, industry standards and Ontario’s Occupational Health and Safety Act set the boundaries for decisions that affect quality and safety.
- All sectors. If you serve clients in Quebec, Law 25 requires you to inform a person of a decision based exclusively on automated processing and allow them to have it reviewed, and the EU AI Act applies as soon as an AI system or its outputs are used in the European Union.
How do you document accountability?
Whatever the agent’s level of autonomy, a person remains accountable for its decisions, which means being able to explain after the fact what was done, by whom, on what basis and with what result. That is the role of the agent’s action log, which also helps you respond when a client asks for the reasons behind an automated decision, as Quebec’s Law 25 allows, or when a regulator asks how your system reached a conclusion. The design of this log and of the other technical controls around an agent in production will be the subject of an upcoming article.
Key takeaways
Before handing a process over to an AI agent, three guideposts shape the decision. First, choose a use case that combines high volume, written rules, available data and reversible errors, keeping traditional automation for the predictable steps. Next, set each agent’s autonomy according to the risk of its task and your tolerance for that risk, even if it means starting low and moving up as the agent proves itself. Finally, frame that autonomy with rules and technical controls, always keeping a person accountable for the decisions that matter.
In practice, a good first step is to list your organization’s sensitive decisions, then choose a file-preparation process where an agent can prove itself at the first or second level of autonomy, with an action log in place from day one.