Virtual assistant vs AI agent: what to delegate to whom

Virtual assistant, AI agent, RPA and workflow automation: what each one is

A virtual assistant is a remote person who handles your administrative work: inbox, calendar, documents, data entry, follow-up and basic bookkeeping. An AI agent is software that uses a language model to decide its own next steps and complete a task on your behalf. RPA and workflow automation are rule-following software that repeats a fixed sequence you define. The virtual assistant vs AI agent question is really about which tasks need judgment, which need rules and which need an accountable person.

• Virtual assistant (VA). A remote human doing the work an in-house administrative assistant would do. The Bureau of Labor Statistics lists that role's duties as answering phones, scheduling, handling mail, preparing memos, invoices and reports, editing documents, maintaining databases and performing basic bookkeeping.¹ A VA does the same list remotely, often part time.
• AI agent. OpenAI's guide for builders defines agents as systems that independently accomplish tasks on your behalf, and says plain chatbots and single-turn tools that do not control a workflow are not agents.⁷ Anthropic draws the line the same way: in an agent, the model directs its own process and tool use.⁶
• Workflow automation. Anthropic's term for model and tool steps that run along predefined code paths.⁶ In a tool such as Zapier, a trigger (a new form entry) fires set actions (create a CRM record).
• RPA. Robotic process automation, which IBM describes as software robots that perform repetitive office tasks such as extracting data, filling in forms and moving files, and which can follow only the processes a user has defined.⁸ Our explainer on how RPA works and where it stops goes further.

The useful distinction is who decides the next step. RPA and workflows decide nothing; you decided when you wrote the rule. An agent decides within limits you set. A VA decides with context and can be held to account.

Task delegation matrix: what to give a VA, an AI agent or a workflow

Give fixed, repeatable steps on structured data to workflow automation or RPA. Give reading, drafting and sorting of messy inputs (emails, PDFs, research) to an AI agent, with a person checking the output. Keep anything that moves money, commits you to a customer or depends on a relationship with a person, and make the VA the owner of the exceptions.

The rule follows the builders' own guidance. OpenAI says agents earn their place on work with nuanced judgment, rules too tangled to maintain, or heavy reliance on unstructured data, and that otherwise "a deterministic solution may suffice." The same guide says sensitive, irreversible or high-stakes actions, such as authorizing large refunds or making payments, should trigger human oversight.⁷ Workers draw a similar line. In a Stanford survey of 1,500 workers across more than 844 tasks, people welcomed agent automation for 46.1% of tasks, and the most common reason, given in 69.4% of pro-automation answers, was freeing time for higher-value work.⁹

Almost no row has the agent working alone: it drafts or extracts and a person decides. That is the practical answer to what to delegate to AI. The VA inherits the exceptions, conflicts and anything involving money, the part that does not shrink when you add software.

Virtual assistant vs AI agent cost comparison

An AI tool stack for one person costs tens of dollars a month: about $74 for a Copilot seat, a Zapier plan and a Zapier agent add-on at annual prices, before the Microsoft 365 license Copilot needs. Outsourced back office support, the closest published proxy for a remote VA, averages under $25 an hour on Clutch, so at that average, 40 hours a month comes to under $1,000. A US in-house administrative assistant at the BLS median costs about $70,400 a year with benefits, about $5,860 a month, by our rough upper estimate.

BLS puts the median secretary or administrative assistant at $48,310 a year, or $23.23 an hour, in May 2025; executive assistants sit at $76,590.¹ For office and administrative support in private industry, the June 2026 employer cost survey shows $11.51 an hour in benefits on top of $25.22 in wages, a ratio of about 45.6% (our calculation).² Applied to the median, that is about $70,400 a year, or $33.80 an hour. Treat it as an upper estimate, because $2.77 of the benefit figure is paid leave, which an annual wage already covers.

Virtual assistant cost is lower still. Clutch's back office outsourcing guide, updated September 2026, puts the average cost of back office outsourcing services at less than $25 an hour, and places the United States, India and the Philippines all in that band.³

Software looks close to free. Microsoft 365 Copilot Business lists at $21 a user a month on an annual plan ($18 on a promotion running through December 2026, $25.20 month to month), and it needs a qualifying Microsoft 365 license underneath.⁴ Zapier's Professional plan starts at $19.99 a month billed annually for 750 tasks, and its agents product starts at about $33.33 a month on an annual Pro plan for 1,500 activities.⁵

The chart flatters the software: a subscription is not a finished job. Someone designs each workflow, checks output and fixes what breaks; price those hours. Tools win on high-volume, rule-bound work; a VA wins on varied work where a wrong answer costs more than the task.

The hybrid model: a virtual assistant who runs the AI

For most small teams the answer is both: a VA who owns the admin function and uses AI agents and workflows to do more of it. The software drafts, extracts and routes; the VA sets the rules, reviews the output, handles exceptions and owns the result.

In the Stanford study, equal partnership between worker and agent was the dominant preferred level of involvement in 45.2% of occupations.⁹ It also fits what current agents can do. On TheAgentCompany, a benchmark set in a simulated small software company, the best agent tested finished 30% of tasks on its own in its September 2025 results.¹⁰

• Morning inbox. An agent labels and drafts replies overnight. The VA sends the routine ones and flags the few that need you.
• Data flow. Workflows move form entries into the CRM. The VA works the exceptions: duplicates, missing fields, anything the rule could not place.
• Books. Bookkeeping software applies categorization rules; the VA or a bookkeeper checks the outliers before the monthly close. Heavier finance work suits a dedicated back office specialist.
• Customer messages. An agent answers common questions from your playbook. Refunds and cancellations go to the VA, in line with OpenAI's advice on human oversight for high-risk actions,⁷ and so do complaints.

If you want the automation layer built properly, that is a job for AI workflow automation, with the VA as its daily operator.

Risks of delegating to a VA or an AI agent

An AI agent's main risks are errors that compound across steps, too much access, and following instructions hidden in the emails or files it reads. A VA's main risks are access and continuity: shared passwords, payment authority and a single person holding the knowledge. The controls overlap and are cheap: named accounts with least access, human approval for anything that moves money, and a written process that survives turnover.

• Compounding errors. Anthropic warns that agent autonomy brings higher costs and the potential for compounding errors, and recommends testing in sandboxed environments with guardrails.⁶
• Excessive agency. OWASP's 2025 list for language model applications names excessive functionality, excessive permissions and excessive autonomy as the root causes of agents doing damage, and recommends human approval for high-impact actions and permissions limited to the minimum necessary.¹²
• Hidden instructions. An agent that reads your inbox reads whatever strangers send you. OWASP classes this as indirect prompt injection, where a model accepts input from external sources such as websites or files. Keep such agents away from payment and send rights.
• Trust and accuracy. Among Stanford respondents with concerns, 45.0% named a lack of trust in AI accuracy or reliability, the top concern.⁹ That argues for review, not avoidance.
• VA access. Give a VA their own login with a limited role, never your password, and keep payment release with you.
• Continuity. One person holding every process is a single point of failure. Write the checklist down and keep every account and automation in your name.

How to decide: five questions for each task

Run each recurring task through five questions: are the steps fixed, is the input structured, what does a mistake cost, does it touch money or a relationship, and how often does it happen. Fixed, structured, frequent, low-stakes work goes to a workflow. Messy but low-stakes work goes to an agent with review. High-stakes, rare or relational work goes to a person.

1. Are the steps the same every time? Yes points to a workflow or RPA. Anthropic's advice is to find the simplest solution possible and add complexity only when it is needed.⁶
2. Is the input structured? Form fields and spreadsheets suit rules. Emails, PDFs and phone notes suit an agent, since handling unstructured data is where AI goes beyond RPA.⁸
3. What does a mistake cost? Cheap, easy-to-spot errors: automate and sample. Expensive or public errors: a person reviews every output.
4. Does it move money or commit you to someone? Payments, refunds and cancellations keep a human approval step,⁷ and so should contract terms.
5. How often does it happen? Twice a year is not worth automating; fifty times a week usually is.

AI automation for small business is still early. The Census Bureau's survey found AI use at 17% to 20% of US firms between December 2025 and May 2026, and under 20% among firms with four or fewer employees, against 37% at firms with 250 or more.¹¹ Demand for the role persists: BLS projects about 314,400 openings a year through 2035, all to replace workers who leave, even as employment edges down 2%.¹

If the matrix leaves a pile of varied, judgment-heavy admin work, hire a dedicated virtual assistant and give them the tools. For agents beyond off-the-shelf tools, our guide to how AI agents work in production covers custom builds, and our view of AI in operations management shows how the pieces fit at larger scale.

Gunjan Paul

Head of Human Resources, Resourcifi

Gunjan Paul is Head of Human Resources at Resourcifi. She runs hiring and vetting for the engineering teams, sets the standard each role is held to, and reviews the roles and engagement details we publish.

Resourcifi on LinkedIn →

Kanika Mathur

Reviewer. Head of Service Delivery, Resourcifi

Kanika Mathur is Head of Service Delivery at Resourcifi. She leads the engineering pods that scope, build and run client software, from mobile apps to AI systems, and she reviews the process and figures in our engineering guides for accuracy.

Resourcifi on LinkedIn →

Keep reading

Related guides worth your time

← →

Models & sourcing In-house vs outsourced development team: cost, control and when to use each An outsourced development team vs in-house hires, compared on cost, speed, control, IP and risk, with a sourced cost tabl... Read guide → Models & sourcing Outsourced bookkeeping cost: what's included and how it works Outsourced bookkeeping cost explained: what the service includes, what it costs per month by business size, how the hando... Read guide → Models & sourcing Outsourcing to India Guide Learn how to outsource software development to India: engagement models, rate ranges by region, risk controls, and partne... Read guide → Models & sourcing Staff Augmentation Guide What is staff augmentation and when should you use it? This guide covers IT staff augmentation models, rates, and how to... Read guide → Models & sourcing Staff augmentation vs outsourcing Staff augmentation vs outsourcing: control, cost, IP, and risk compared. Find out which model fits your team, plus when a... Read guide → Models & sourcing US-based, offshore, or hybrid: which software team model fits your project Compare US-based, offshore and hybrid software teams on rates, overlap hours, quality control, IP and compliance, with a... Read guide → Agency & white-label Agency AI Pricing What AI development cost depends on: ranges by phase (POC $75k to $150k), the six pricing models, scoping with a paid POC... Read guide → Hiring AI Engineering Team Structure AI engineering team structure explained: core AI team roles, AI engineer vs ML engineer, the minimum viable team, and how... Read guide → Hiring AI staff augmentation: cost, models and when it beats hiring AI staff augmentation explained: roles you can add, in-house vs augmented cost per seat, time to productivity, engagement... Read guide →