Anything with a definite trigger and a definite output can be automated outright; anything that produces a representation about a person can only be assisted; and anything that is legally somebody’s obligation stays with a named human. Sort your process into those three buckets and the build list writes itself.
Key takeaways
Most automation projects at agencies fail for the same reason: someone automated a step that was actually a judgement, discovered it produced nonsense in the awkward cases, and switched the whole programme off. The fix is a sorting rule applied before anything is built.
Automate. The step has an unambiguous trigger, an unambiguous output, and no opinion in the middle. Nobody is represented, assessed or rejected.
Assist. The step produces writing that goes to a human being — a summary, a note, an email, an ad. A model drafts; a person approves and sends.
Never. The step decides something about a person, makes a representation about them, or discharges an obligation the law places on a named party.
These are the ones that pay back quickly and carry almost no risk, because the output is a record or a reminder rather than an opinion.
Record creation from an application. Parse the CV into structured fields, create or match the record, attach the document, log the source. Parsing is not screening — that distinction matters, and it is the subject of its own article.
Deduplication and merge suggestions. The single highest-value background job at most agencies, and the one that makes every later tool work better.
Availability and confirmation loops. Interview confirmations, reschedules, reminders the day before, address and parking details, and the follow-up when nobody replies.
Timesheet and assignment administration on a contract desk. Submission reminders, approval chasing, missing-hours flags, and end-of-assignment prompts for the redeployment conversation.
Compliance document expiry. Certifications, tickets, work authorisations and insurance documents with an expiry date sitting in a field. A calendar-driven watch is trivial to build and prevents an ugly Monday.
Statutory clocks. Ontario now requires an employer to tell an interviewed applicant whether a hiring decision has been made within 45 days of the interview, and that applies equally where a recruiting firm interviews on the employer’s behalf. It is a date arithmetic problem with a written output — exactly what automation is for. The same section requires postings and associated application forms to be retained for three years after public access is removed, which is an archiving job, not a filing habit.
Posting distribution and version archiving. Push once, archive every version automatically, retain the evidence.
Everything here is drafted by software and sent by a person. The rule is not ceremonial: the moment a message goes out unread, the agency has made a statement it has not checked.
Job ads to a compliant template. Candidate summaries built from facts you supply. Interview confirmations with role-specific detail. Rejection notes, which is the category most improved by automation because it is the category most often skipped. Client status updates. Debrief notes from a recorded call. Weekly desk summaries.
For anything sent in volume, the sending system — not the drafting tool — has to handle consent. Canada’s anti-spam legislation requires consent, identification information and an unsubscribe mechanism for a commercial electronic message, with the onus on the sender to prove consent. A cadence that fires four times regardless of reply is where an agency turns a productivity tool into a complaint.
Three categories, and each has a reason rather than a superstition.
The decision to reject. A filter can order a list. A person decides who is out. The reasons are set out in our article on AI resume screening and the bias risk, and they are legal reasons in Canada, not merely ethical ones.
Any representation about a candidate. Under the ACSESS Code of Ethics, member agencies take all reasonable steps to give clients accurate information about a candidate’s qualifications and experience, and present only candidates who authorised the agency to represent them. A generated profile sent without review breaches both.
Accepting an order. The same code commits members to refuse a client order that is discriminatory in any way. That is a human reading a requirement and saying no.
More generally, the ethical AI principles ACSESS endorses put it as “human in command”: systems designed to augment human capability, with clear processes keeping them under human direction and control at all times, and accountability remaining with whoever deploys them.
Every automation needs five things written down before it is built, and the fifth is the one that gets skipped.
Trigger — the exact event, in the exact system. Action — what happens, including what it must not do. Exception path — what happens when the record is incomplete, the email bounces, or the date is missing. Owner — a named person, not a team. Log — a record that it ran, with enough detail to reconstruct what was sent to whom.
Without the exception path you get the characteristic failure: an automation that quietly does nothing for three weeks because a field changed name, and nobody notices because success and silence look identical. Alert on non-execution, not just on errors.
Duplicate contact, when two automations own the same nudge and a candidate gets both. Stale data, when a sequence addresses someone by a title they left in 2023. Cadence creep, when four separate sequences each fire twice and the aggregate reads as spam. And orphaned automations, built by a consultant who has since left, still emailing from a mailbox nobody reads. Put a review date on every automation at the moment you build it.
A light-industrial temp desk in the Greater Toronto Area places 90 to 130 workers a week. Four automations, built in sequence over a quarter.
First, timesheet chasing: a reminder at 4pm Friday to any assignment worker with no submitted hours, a second at 9am Monday, and an exception list to the coordinator rather than a third message. Second, compliance expiry: a 30-day and 7-day warning on any certification field, with the assignment flagged if the date passes. Third, end-of-assignment: a redeployment prompt to the consultant seven days out, with a drafted message to the worker awaiting approval. Fourth, the interview outcome clock for the permanent desk, with a drafted notice at day 30 and an escalation at day 40.
None of the four makes a decision. Two of them produce drafts a person sends. The coordinator’s Friday afternoon went from chasing 40 timesheets to reviewing an exception list of six, and the permanent desk stopped leaving interviewed candidates without an answer.
Write down every recurring task on one desk for a week, then sort each into automate, assist or never. Build the two highest-frequency items in the automate column first, because frequency beats complexity for payback. Do not start with screening; it is the most interesting item on the list and the worst first project. A sign-off workflow is what keeps the assist column honest as it grows.
Is this AI or just workflow automation?
Most of the automate list is ordinary workflow automation and works better for it — deterministic, testable, auditable. AI belongs in the assist column, where the output is language. Confusing the two is why projects overrun: people buy a model to do a job a rule would have done more reliably.
Can we automate the rejection email?
Draft it automatically, send it deliberately. In Ontario the outcome notice after an interview is a statutory obligation with a 45-day deadline and a three-year retention requirement, so the record of what was sent matters as much as the sending.
How do we stop candidates receiving four messages in a day?
Give one system ownership of outbound contact and let every automation request a send rather than send directly. Rate-limit per contact. This is also what makes consent and unsubscribe handling under the anti-spam rules possible to evidence.
A 30-minute call is enough to tell you whether AI pays for itself here.