This is general operational guidance, not legal, clinical or regulatory advice.
Most AI agent projects start with the happy path. A patient completes an assessment, books a consult, gets the right information and moves to the next step. The agent handles it. Everyone claps.
Then it goes live, and the work moves somewhere nobody designed: the pile of things the agent couldn't finish.
A booking is half done. A form is submitted twice. A patient is under the required age. A result is missing. A clinician has to look at something before the workflow can continue. A privacy request arrives through the wrong channel.
The agent does its part and hands those cases to a person. Then they sit. That's the real bottleneck in most AI rollouts I see. The model is usually fine. The approval queue behind it has no owner, no clock and no rules.
I'll use telehealth as the main example because that's where the stakes are highest. The same pattern applies to billing, onboarding, claims, service desks and any service business where a person has to say yes before something happens.
An exception queue is the controlled place where cases go when they leave the routine workflow. Something about the case means a rule, a missing dependency or a risk needs a human.
Every item in that queue should carry five things:
If the queue lives in an inbox, a spreadsheet and three Slack threads, every review starts with a scavenger hunt.
Agents work best when their authority is bounded. Before an agent can safely act, you need to know:
Skip those decisions and an instruction like "handle patient follow-up" is far too broad. A safer version reads like this: find incomplete admin steps, send the approved reminder, route the case to the right queue, and wait for a person on anything outside that list.
Build the queue first and the agent has somewhere safe to put everything it shouldn't touch.
The agent, the reviewer and your reporting should all work from the same customer, patient, invoice or case record. In HubSpot, that means clear identifiers, status, owner and history on the record itself. If the agent reads one system while the reviewer updates another, the queue drifts and every exception turns into reconciliation work. (I wrote about why messy records cause this in Your CRM is a mirror.)
Map the normal sequence first: enquiry, intake, booking, consult, follow-up. You can't spot an exception until the routine path is clear.
Then name exceptions as things you can observe: a missing required field, a duplicate submission, a failed payment, an unconfirmed booking, an identity mismatch, or a requested action outside the agent's authority. "Something looks off" isn't an exception type.
"Open" and "done" aren't enough. Useful states include:
Each state gets one owner, one next action and one clock.
A high-confidence result can still carry a high consequence. A low-confidence one can be harmless. Route on both:
In healthcare, the line is tighter again. An admin agent can organise, extract, route, notify and prepare. It shouldn't diagnose, interpret results, recommend treatment or prescribe. Named clinicians keep clinical judgement.
"The team will review it" isn't ownership. Name a primary reviewer, a backup and a daily review window. Set time-based escalation so ageing items surface before they hold up billing, delivery or patient care.
HubSpot workflows can notify owners, create tasks and escalate overdue items. Ticket pipelines or a custom object can keep the work visible instead of buried in email. I covered a few of these patterns in 5 HubSpot workflows that replace 20 hours a week.
Every approval or rejection should record why, from a short controlled list: missing identifier, duplicate record, policy exception, pricing mismatch, incomplete evidence, agent recommendation wrong.
That tells you far more than a generic accuracy score. It shows whether the problem is the model, the source data, the process or the policy.
Record the input, the knowledge source, the action, the timestamp, the rule applied and who approved it. That's your evidence for quality reviews, privacy enquiries and incidents.
Don't let the agent quietly learn from every human decision. Review the reason codes, sample outcomes, and have a person approve changes to rules, prompts or knowledge. People approve the boundary as well as the exceptions.
If you run a health service, two questions sit underneath all of this.
Is the software doing a clinical job? The TGA's guidance on clinical decision support system software explains that this kind of software can be a regulated medical device. Some of it is exempt if it meets all three exemption criteria, but exempt software is still subject to TGA oversight, including notification and adverse event reporting. Keeping agents on admin tasks is the simplest guardrail.
Is a program using personal information in a significant decision? Changes to the Privacy Act legislated in December 2024 bring in new transparency obligations from 10 December 2026. Under APP 1, your privacy policy will need to explain where a computer program uses personal information to make a decision, or to do something substantially and directly related to making one, that could significantly affect someone's rights or interests. The test reaches past the final decision, so a person clicking "approve" at the end doesn't automatically take a system out of scope. Map where software ranks, blocks, prioritises or recommends.
These numbers connect what the agent does with what the business feels.
The pattern I keep coming back to is simple: one record, self-service where it helps, bounded agents, and people approve. A good agent reduces coordination work and keeps whatever is left in plain sight.
If you've already got an agent handing work to people, look at the queue before you look at the model. That's usually where the time goes. As I said in the honest truth about replacing a hire with AI, the handshake between the agent and the human is where it either works or it doesn't.
Thinking about an agent for one of your workflows, or stuck with one that keeps piling up exceptions? See how we design bounded AI agents and the queues behind them, and let's map the approval path before you build.