An inbox contains a routine appointment request, a customer complaint and an email asking to change payment details. They all look like administration, but they require different information, permissions and decisions.
That is why choosing between an AI agent and a virtual assistant works better at task level than job-title level. Some steps need only a fixed rule. Others suit AI assistance, a trained person or a specialist with authority to decide.
Use the following framework to allocate work before buying another tool or expanding a role. It is a practical management proposal, informed by primary sources checked on 23 September 2026.
Define the task and its finish point
A virtual assistant is a person working remotely. Their suitable responsibilities depend on experience, training and authority. A Philippines-based team member may coordinate routine work or bring specialist expertise; location does not determine the level of judgement they can exercise.
An AI agent combines a model with tools and a workflow to pursue an objective and take actions. Its permitted actions matter as much as the quality of its answers. ACSC: agentic AI explained.
Write a task as an observable result: “prepare three appointment options that meet these confirmed constraints”. Then separate preparing options from making a booking. A drafting tool may be sufficient for the first step; access to a live calendar introduces another decision.
Ask four questions before allocating the work
- Ambiguity: Are the inputs complete and the rules stable, or must someone clarify conflicting instructions and understand context?
- Permissions: What must the worker or tool read, change or send? Can the task be completed with narrower access?
- Consequences: If the result is wrong, can it be corrected before anyone relies on it? Consider commitments, money, personal information and technical decisions.
- Exception frequency: How often does the normal procedure fail to cover the case? Are exceptions recognisable, or does recognising them itself require judgement?
These questions are discussion prompts, not a numerical safety score. The National AI Centre similarly describes its screening questionnaire as an early check without scores or automatic outcomes, followed by further assessment where needed. National AI Centre: AI screening questions.
Keep the answers beside the task description. Revisit them when the tool receives more access, the workload changes or new types of request appear.
Choose a starting option for each task
This original matrix shows how the questions can change the allocation. Adapt it to your actual procedures and people.
| Task | Starting option | Reason and boundary |
|---|---|---|
| Remind an owner when a recorded due date arrives | Fixed automation | The condition is explicit; check current status to avoid obsolete reminders |
| Flag missing required fields in a standard export | Fixed validation rules | Known fields need predictable checks; a person resolves missing or conflicting facts |
| Draft a reply from approved facts | AI assistance with VA review | Language support may help; sending and customer commitments remain separately controlled |
| Monitor approved public supplier pages for document changes | A narrowly scoped AI agent trial | Multiple retrieval steps may be useful; provide source links and review proposed changes before using them |
| Reorganise appointments after several customers change availability | Trained VA | Coordination may require clarification and negotiation within booking authority |
| Resolve a complaint involving a refund or disputed promise | Authorised person, with VA coordination | Gather evidence and route the decision to someone with the necessary authority |
An agent is unnecessary when a scheduled rule or existing integration already completes the task reliably. For solar-specific examples of assisted document work, see our guide to AI in solar administration.
Match permissions to the selected task
For the public-document monitoring example, the agent might read approved websites and write a comparison into a review folder. It would not need customer records, payment access or permission to overwrite the company's approved procedures.
ACSC guidance recommends minimum necessary permissions, explicit limits on autonomous actions and human control points. It also identifies malicious instructions in external content as a risk. A web page or retrieved document should not acquire authority to expand the agent's job. ACSC: securing agentic AI services.
Write down the allowed inputs, permitted actions, stopping conditions and person responsible for intervention. Confirm the software can enforce the boundary; a sentence in a prompt is not a substitute for access controls.
For a VA, define access and approval authority just as explicitly. Neither a job title nor an account login tells someone whether they can authorise a refund, change a contract or make a technical declaration.
Let exception patterns shape the handover
Imagine a hypothetical supplier-monitoring task. Most weeks, the output is a list of revised public documents. One week, the agent finds two manuals with conflicting revision dates and cannot establish which applies to the equipment your business supplies.
The useful result is an exception containing both links, the conflicting details and the unresolved question. A coordinator can obtain clarification, while the responsible technical person decides which document applies. The agent should not choose a version merely to finish the report.
Review exceptions by cause. Repeated missing inputs may call for a better intake process. Frequent customer negotiation may justify a person owning that queue. A predictable formatting problem may be solved with a fixed rule.
Do not adopt a universal exception-rate threshold. A rare payment error and a frequent harmless formatting correction have different consequences. Track severity and the work needed to resolve each category alongside frequency.
Compare accepted results across the whole task
Test the proposed allocation on representative work, including awkward cases. Record setup, checking, clarification and recovery time as well as the initial processing time. Compare results against the same completion standard.
A published experiment with consultants using GPT-4 found AI assistance improved some task outcomes while reducing correctness on another task. It studied AI-assisted people, not autonomous agents replacing VAs, so it cannot settle this purchasing decision for your business. Read the published task-level experiment.
Choose the arrangement that produces acceptable work with manageable review and clear ownership. Our guide to combining people and AI explains how those choices fit into a wider team structure.
Talk to Nexus about mapping your administration workload and deciding which tasks need a VA, a specialist or a carefully bounded tool.
