Short answer: AI agent vs virtual assistant comes down to the kind of work. An AI agent is software that drafts, files and prepares repeatable admin inside your tools. A virtual assistant is a person who handles tasks that need conversation or judgement. You can also use both: the agent prepares, a person reviews and decides.
This guide is for owners of small accounting, bookkeeping, law, consulting, clinic and agency teams in Toronto/GTA and Vancouver. Related reads: shadowing admin before you automate, admin tasks an AI agent can take on and scheduling with an AI agent.
Key takeaways
- Score each task on four questions, then mark it automate, assist or keep human.
- Decide who reviews what, and when work escalates, before either one starts.
- An agent's cost is mostly setup and review time; an assistant's is ongoing hours, onboarding and cover.
- A named person always owns advice, money approvals and commitments.
What is an AI agent, and how is it different from a virtual assistant?
An AI agent is software that runs the admin steps you map; a virtual assistant is a person who handles work that needs conversation or judgement.
The agent drafts a document, files it or proposes a time, then hands results to a person to check or spot-check. The difference that matters: an agent only knows the steps you mapped. A virtual assistant can improvise, read a client's mood and pick up the phone.
Which admin work fits an AI agent vs a virtual assistant?
| Admin area | AI agent | Virtual assistant | Person still owns |
|---|---|---|---|
| Client intake paperwork | Drafts forms, checks missing fields, files packs | Chases clients for missing items | Accepting the file |
| Document prep and naming | Fills templates, renames and files | Assembles unusual packs by hand | Advice content and signatures |
| Inbox triage | Labels, drafts follow-ups, flags urgency | Handles sensitive or awkward threads | What actually gets sent |
| Scheduling and reminders | Proposes slots, sends reminders, updates holds | Resolves conflicts and exceptions | Final calendar calls |
How do you score your admin tasks?
List every admin task from one or two busy weeks and score each from 1 to 3 on four questions. On every question, 3 means safe to automate and 1 means a person must do it, so higher totals suit an AI agent.
| Question | 1 (needs a person) | 2 | 3 (safe to automate) |
|---|---|---|---|
| Repeatability: does it run the same way each time? | Rarely | Mostly, with variations | Same steps every time |
| Judgement: does someone weigh options? | Every time | Sometimes | Almost never |
| Client contact: does the output reach a client? | Directly, as a commitment | After a person reads it | Never, internal only |
| Sensitivity: what does it touch? | Advice, payments or money approvals, health or privileged files | Client personal details | Routine internal records |
Copy these columns into a spreadsheet: Task | Repeatability | Judgement | Client contact | Sensitivity | Total | Result | Owner
Add the scores, then check the override and the client-contact rule:
- 11 to 12: automate. The agent does it; a person spot-checks.
- 7 to 10: assist. The agent prepares; a person finishes.
- 4 to 6: keep human. A virtual assistant or staff member owns it.
- Override: any 1 on sensitivity (advice, payments or money approvals, health or privileged files) keeps a task human.
- Client-contact rule: if it scores 1 on client contact, the agent may draft but a person sends.
What does this look like in a small firm?
Fictional example, not an 852 Tangram client: a five-person bookkeeping practice with one partner, three bookkeepers and an office coordinator.
| Task | Scores (R, J, C, S) | Result | Who does what |
|---|---|---|---|
| Renaming and filing client statements | 3, 3, 3, 2 = 11 | Automate | Agent files; coordinator spot-checks |
| Chasing missing month-end documents | 3, 2, 2, 2 = 9 | Assist | Agent drafts; coordinator sends and handles replies |
| Booking year-end review meetings | 3, 2, 2, 3 = 10 | Assist | Agent proposes times; coordinator confirms |
| Setting up a new client's files | 2, 1, 2, 2 = 7 | Assist | Agent prepares a checklist; bookkeeper completes it |
| Reassuring a client about a tax notice | 1, 1, 1, 1 = 4 | Keep human | Partner |
Client statements score 2: personal details, no payment approvals. On this sheet the practice would not need a new assistant. The agent takes filing and first drafts, and the coordinator moves to the exception queue. A practice whose clients mostly phone could land the other way.
Who reviews what when an AI agent and an assistant work together?
Agree these in writing first:
- Nothing reaches a client unreviewed at first. A named person approves every outbound draft. Loosen that one task at a time, in writing.
- One review queue, one owner. Agent output lands in a single queue the assistant or coordinator clears.
- The agent hands to a person when information is missing after two reminders, a client sounds upset, or the case falls outside the mapped steps.
- The assistant escalates to the owner anything about money, deadlines, advice or a new commitment.
- Review the task log weekly: what the agent got wrong, what the assistant fixed, which tasks should change column.
What should you set up before either one starts?
- Access. Separate logins for the agent and the assistant, limited to the inboxes and folders in scope. Agree who holds each account.
- Approvals. A one-page list of who approves what.
- A task log. Task, who did it, who checked it, what went wrong.
- Templates and naming rules, so both start from the same examples.
- Privacy. Canada has privacy laws that cover client personal information. Depending on your province and sector, a provincial law may apply instead, such as British Columbia's PIPA, or Ontario's PHIPA for personal health information. Check your obligations and your professional body's rules before client files go to a new tool or person. The Office of the Privacy Commissioner of Canada has guidance for businesses.
- An exit plan. Know how to remove access if you change course.
How does cost behave for each option?
Prices vary, so compare how each cost behaves.
| Cost driver | AI agent | Virtual assistant |
|---|---|---|
| Start-up | Mapping and setup time | Hiring and onboarding |
| Each week | Review time, falling as drafts improve | Paid hours, rising with volume |
| More volume | Little extra, unless exceptions grow | More hours |
| Process change | Retuning the steps | Retraining |
| Absence | Keeps running, still needs a reviewer | Needs holiday and sick cover |
What are the signs you picked wrong?
- The reviewer rewrites most agent drafts: the task needs judgement, so move it down a column.
- The exception queue grows faster than it clears: the steps were mapped badly, or the work needs a person.
- The assistant spends the day copying data between systems: that slice belongs to an agent.
- Clients chase replies or comment on tone: too much went out without a person.
- Nobody opens the task log: no one owns review, so fix that before adding scope.
What should you never hand to an AI agent or a virtual assistant alone?
Client advice, confidential judgement, money approvals, and anything your regulator or professional body says needs a person (check with yours). Also keep a named person on emails that create commitments, contract or clinical note changes, hiring and public statements.
Comparing options for your firm?
Prefer to write? The short form is at the end of this post.
How 852 Tangram helps
Our AI operations service starts on site: "We come to your office, shadow your team and follow the day-to-day." The steps are Shadow, Map, Set up, Hand over and tune. The worksheet above is a do-it-yourself Map step.
Our published case is a dance studio, not a professional services firm. ZeekersDanz is a K-pop dance studio in Markham. From its case page: "We spent time in the studio office, shadowing the team and following their day-to-day work. Then we set up an AI agent for the studio, to automate daily paperwork, class scheduling and admin." A studio office and a firm's back office share routine work: forms, records, calendars and reminders. The method transfers; your task list is your own.
On scope: "You decide which tasks and files are in scope, and we agree that with you in writing, including who holds the accounts, before we set anything up."
Frequently asked questions
What is the difference between an AI agent and a virtual assistant?
An AI agent is software that follows agreed steps to draft, file, remind or propose times, then hands results to a person to check or spot-check. A virtual assistant is a person who completes assigned tasks and can handle conversation and judgement. You can use both. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
Can an AI agent replace a virtual assistant in a small firm?
For repeatable paperwork, filing and scheduling prep, often yes, once the steps are mapped. It does not replace judgement, advice, money approvals or difficult client conversations. If most of your admin is exceptions and relationships, a person is the better fit. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
How do I decide which admin tasks to automate?
List the tasks from one or two busy weeks. Score each from 1 to 3 on repeatability, judgement, client contact and sensitivity, where 3 is the end that is safe to automate. High totals go to an AI agent, middle totals to an agent draft that a person finishes, and low totals stay human. A 1 on sensitivity always stays human. If it scores 1 on client contact, the agent may draft but a person sends. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
How does the cost of an AI agent compare with a virtual assistant?
An AI agent costs mostly setup time up front, then review time each week, plus retuning when your process changes. A virtual assistant costs ongoing hours that rise with volume, plus onboarding and cover for absence. Count your own review time in both. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
Do we need technical staff to run an AI agent?
No. We do the setup and show your team how to use it. You do need one named person to own the review queue and task log, and to approve which tasks and files are in scope. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
Where has 852 Tangram set up an AI agent?
At ZeekersDanz, a K-pop dance studio in Markham. We shadowed the studio office, then set up an AI agent to automate daily paperwork, class scheduling and admin. It is a studio, not a professional services firm, but the shadow-first method carries over. Book a free strategy call at https://calendly.com/852tangram/free-strategy-call.
Talk to 852 Tangram
We can help you map your firm's admin and decide what to automate, what to assist and what to keep human.
Or send us a short note below and we'll reply by email.
