When healthcare teams hear that AI can help with special authorization requests, the first question is usually the same one: who is making the decision?

It is a fair question. These requests sit close to patient care. A missing lab result, an incomplete form or an outdated assessment can delay a treatment, service or device a patient needs.

The short answer is simple. A well-designed AI agent does not make clinical decisions. It takes on the paperwork around those decisions. That distinction is the whole design.

How it works in Canada

Canada does not have one authorization process. Requirements vary by province, public program, private insurer and type of care.

A request may involve coverage for a drug, medical equipment, an extended course of therapy, an out-of-province service or another treatment that requires supporting evidence before it is funded. Ontario’s Exceptional Access Program (EAP) and British Columbia’s Special Authority process are examples involving prescription drugs. Private insurers may have separate authorization processes for specialty medications, equipment and certain services. Federal programs such as the Non-Insured Health Benefits program and Veterans Affairs Canada also have their own prior approval or special authorization requirements.

This is different from Health Canada’s Special Access Programs for drugs and medical devices, which allow practitioners to request products not authorized for sale in Canada for patients with serious or life-threatening conditions when conventional options have failed, are unsuitable or are unavailable. This article focuses on treatments, services or products that may be available but require approval before coverage is provided.

Separate the work from the decision

Most authorization workflows contain a large amount of administrative work.

Someone must identify whether authorization is required, find the current criteria and collect the supporting information. Depending on the request, that could include consult notes, lab results, treatment history, imaging reports, therapy assessments, equipment quotes or specialist recommendations. The team must then complete the appropriate form, check for missing information, submit the request and follow up. Renewals may require the process to be repeated.

Then there is the part that needs judgment. Is the requested treatment, service or equipment appropriate for this patient? Does the supporting information accurately represent their condition? Does an unusual case need further explanation? What should the patient be told if coverage is refused?

Those decisions belong to people. The agent prepares the package. The clinician or authorized reviewer makes the decision.

The agent does not recommend treatment or make an independent clinical eligibility determination. It organizes and checks the evidence for a qualified person to review.

What the agent handles

The agent handles four things:

• Finding the route. Does the request involve a drug benefit, medical equipment, therapy, an out-of-province service or another approval pathway? What criteria apply, and which form or portal should be used?
• Gathering the evidence. Pulling relevant notes, labs, treatment history, assessments, imaging reports, specialist letters and other supporting records into one package.
• Checking for gaps. Comparing the package with the applicable criteria and flagging anything missing or out of date before the request goes out.
• Following up. Tracking request status through the channels the organization is authorized to access, flagging anything that has stalled and warning when a renewal is due.

Each step produces something a person can check quickly. That makes these tasks suitable for careful automation. They are also where much of the delay builds up.

What stays with people

Clinical judgment. Whether a treatment is appropriate for a patient is a clinician’s call.

Final review. Nothing goes to a ministry, public program or insurer under a clinician’s or authorized reviewer’s name without a person looking at it first. This is the most important boundary.

Unusual cases. Complex histories, conflicting records or requests that fall outside standard criteria need careful human review. The agent’s confidence is lowest here, and professional context matters most.

Talking to  patients. A patient asking why their treatment or coverage is delayed needs a real conversation, not a status update.

The grey area

The line is not always clean.

Criteria often ask for information that lives in free text rather than structured fields. A drug program may want proof that a patient tried a first-line therapy, for how long and what happened. An equipment request may depend on a functional limitation described in an assessment. A request for additional therapy may rely on a progress note that explains why more sessions are clinically appropriate.

When the agent finds that information and turns it into a summary, it is interpreting, not just retrieving.

That work is still worth automating because it takes substantial staff time. It needs two safeguards. The agent must show its source so the reviewer can open the original record. Interpretation should also remain under closer review than straightforward retrieval.

Retrieval can be spot-checked. Interpretation should be read.

What this could look like in practice

The workflow will vary depending on what is being requested.

Request	How an AI agent could help
Specialty drug	Identify the applicable provincial or private-insurer criteria, gather treatment history and recent lab results and flag missing evidence before the prescriber reviews the submission.
Medical equipment	Collect the prescription, functional assessment, supporting clinical notes and supplier quote, then check whether the required documents are complete.
Extended therapy	Gather progress notes, previous treatment details and the clinician’s rationale for additional sessions.
Out-of-province service	Collect the referral, relevant clinical history and supporting specialist documentation, then prepare the package for review and submission through the applicable approval process.

In each case, the agent organizes the information, checks the package and tracks its status. A qualified person confirms that the request is appropriate, reviews the supporting evidence and approves the final submission.

What oversight looks like in practice

Human in the loop is easy to say. In practice, it means making a few decisions before you build.

• Set the boundaries. Write down which steps the agent can complete, which steps require review and who approves the final request.
• Give escalations an owner. Conflicting records, missing information or a case that does not fit the criteria should go to a named person. An escalation that lands in an unwatched shared inbox creates another delay.
• Keep an audit trail. You should be able to see what the agent pulled, what it summarized, what it flagged and who reviewed it.
• Respect privacy law. The agent should operate under the authority of the organization accountable for the health information, whether that organization is described as a custodian, trustee or another defined entity under provincial law. It should access only the information required for the request and log every access. Applicable provincial privacy requirements, including rules governing service providers and cross-border transfers, must guide its deployment.

Review its work regularly. Criteria and forms change. Sample the output monthly. A rise in requests returned for more information can be an early sign that something has shifted.

Two ways teams get this wrong

Too much trust too early

The agent performs well in a pilot, so review quietly stops. The risk is that automated errors repeat. It is not one mistake. It is the same mistake forty times before anyone notices.

Too little automation

Every step goes to a person, even copying a value from one system to another. Staff click approve on work they cannot realistically check. Review helps only when the reviewer can verify something meaningful.

Start carefully, then adjust

Most organizations should begin with a person checking everything. After a few months, the team will know which steps the agent handles reliably and which ones still need corrections. Reliable steps can move to spot checks. Interpretation remains under closer review. Clinical judgment stays with people.

Human review does not cancel out the time saved. Checking a prepared package takes minutes. Building it from scratch across clinical records, insurer or government requirements and supporting documents takes much longer. The work of nurses, clinic pharmacists and office staff shifts from collecting information to reviewing it.

The takeaway

AI can reduce the administrative work involved in authorization requests without taking over clinical decisions.

It can identify the applicable process, gather supporting records, check for missing information and track the request. The clinician or authorized reviewer remains responsible for confirming that the requested treatment, service or equipment is appropriate and approving the final submission.

Clear responsibilities make the workflow safer. The agent prepares and checks the information. People interpret the patient’s situation, handle exceptions and make the final decision.

When that boundary is clearly defined, teams can submit more complete requests, identify gaps earlier and spend less time chasing documents.

Where those boundaries sit will depend on the organization’s workflow, data access and oversight model. For a practical next step, read our guide on what healthcare organizations need in place before deploying an authorization agent, including a readiness checklist for workflow design, data access, oversight and performance monitoring.