Sixty applications, and they all say the same thing. Proficient in Google Workspace, Asana, Canva, Slack, HubSpot. Excellent written and verbal English. 65 words per minute. Detail-oriented, proactive, fast learner. Three have a certificate from an online VA academy you have never heard of. You asked for skills, you got a list of skills, and you are no closer to knowing which of these people will still be useful in a year.

The list measures the wrong thing. Almost every item on a standard VA job post is either learnable in a week or unverifiable from a document.

The skills worth screening for when hiring a virtual assistant are not software proficiencies. They are three behaviours: reading a brief accurately when the brief is incomplete, telling you when something looks wrong instead of quietly working around it, and finishing the boring last step of a task without being chased. Those three decide whether delegation gives you time back or just moves the work sideways, and all three can be tested in about four hours of paid work. That testing is the only part of this process that reliably tells you anything.

Tool proficiency is the weakest screen you have

Think about what you are buying when you hire for “proficient in Asana.” A competent adult learns enough Asana in two afternoons to run your board. Same for Trello, ClickUp, Notion, Calendly and whatever replaces them by 2028. Tool fluency depreciates fast and transfers easily, which is why it is cheap and why everyone claims it.

What does not transfer is judgment about incomplete instructions. Your briefs will always be incomplete. You will write “send this to the client” without saying which of two contacts, because in your own head the answer is obvious. How the assistant handles that gap, ask, guess, or stall, decides more about your supervision load than anything else. So tools drop to “comfortable learning new tools, we train you on ours in week one,” and the top of the list becomes evidence of behaviour under ambiguity. Two ways to get that: watch them do real work, or ask questions that resist rehearsed answers.

The research here gets quoted loosely, so handle it carefully. Schmidt and Hunter’s much-cited 1998 paper in Psychological Bulletin put work samples and structured interviews near the top for predictive validity. Sackett, Zhang, Berry and Lievens revised those estimates downward in 2022 in the Journal of Applied Psychology, showing that the standard corrections for range restriction caused substantial overcorrection and inflated published validities by roughly .10 to .20. The ranking largely survived, with structured interviews highest. Two caveats: none of that work studied a solo owner hiring one assistant, and even the best methods have modest effect sizes. What transfers is method rather than magnitude. A planned test of real work plus a fixed question set beats a friendly chat and a CV.

Three qualities that predict a good virtual assistant

Run these as paid work with two or three finalists. Each is real work you need done, so nothing is wasted even if you hire nobody.

Test one: comprehension under ambiguity

Write a brief that is 95% clear and deliberately unclear in one place, where a careful reader notices and a careless one does not. Send five real enquiry emails with a reply template and one instruction: “Reply to each using the template. Anyone asking about price gets the rate card attached.” Make one of the five an existing client on an old, lower rate who is asking about price. The rule says attach the rate card. Doing that shows a current client your prices went up, in an email from a stranger, with no context.

You are not grading the replies. You are grading the first twenty minutes. What you want is one message: “Email three is from someone already in your system on what looks like a different rate. Standard rate card, or hold that one for you?” Then they do the other four and wait.

Attaching the rate card to all five and saying nothing is the commonest outcome and it is a decline, because that is what will happen every time your instructions are imperfect, which is always. Eleven questions before starting is a different failure, often fixable, but it says this person needs a fully specified brief. Halting all five instead of finishing the four unambiguous ones is a mild negative: every blocker becomes your blocker.

Test two: whether they tell you something is wrong

Give a task with a small planted problem in the source material. Not a trick, just the kind of defect real work is full of. Take a reconciliation or data-entry job of about 40 rows and seed three faults: an invoice dated before the job it bills for, a row where line items do not add up to the stated total (off by a plausible amount, not an obvious one), and a reference link in your instructions that 404s. Ask for the output with no mention of accuracy and no “check for errors.” The instruction is just: produce the summary.

A strong candidate delivers the work plus a short note. Three things looked off, here they are, here is what I assumed, tell me if I got any wrong. A weak one delivers a clean file that has silently averaged over the broken figure. A third flags everything and finishes nothing, which is workable but expensive.

This is the quality most worth a premium and the one nobody screens for. An assistant who surfaces problems early turns your biggest delegation risk, work going wrong invisibly for six weeks, into small annoying messages. Those messages are the product.

Test three: whether they finish

Design a task with four or five steps where the last one is dull and invisible at a glance. Research 25 contacts into a sheet, that is the interesting part. Then: rename the source files to a stated convention, log each entry in your tracker with today’s date, move the original emails to an archive folder.

Most people do the research adequately. The renaming and logging is where you find out what you hired, so open the folder and read the tracker. A busy owner would not, which is exactly the point. Partial completion is not a small flaw either: a process that is 80% done still needs you to hold the other 20% in your head, and holding it in your head was the cost you were trying to remove.

Do not shortcut this with a personality quiz. Barrick and Mount’s 1991 meta-analysis in Personnel Psychology did find conscientiousness to be the one Big Five dimension predicting performance across occupational groups, roughly what “finishes things” points at, but self-report measures are weak on their own and trivially gameable by anyone who wants the job.

Pricing a test task fairly, and what a refusal means

Pay the full rate they quoted, for the time the task genuinely takes, and state the number before they start. Upwork’s published guidance puts typical virtual assistant rates at $10 to $20 an hour with a median of $13, rising to $20 to $35 for marketing, bookkeeping or CRM-capable work, though Upwork notes these are historical contract ranges worldwide rather than a wage survey. So a two-hour test costs $20 to $40 a candidate, and three finalists across three tasks is a few hundred dollars against a bad hire that costs you a month.

Cap each task at 90 minutes to two hours and pay the fixed fee regardless of how long they take, so nobody is punished for being careful. Give a two or three day deadline rather than “today,” because urgency is a separate variable. Never run a test with no intention of hiring, which is theft dressed as diligence and the reason candidates are wary of tests at all.

If someone declines, ask why before writing them off. Two answers are legitimate: fully booked, or burned by unpaid “tests” and wanting scope and fee in writing. Give them both. What should end the conversation is a refusal on principle, where the portfolio is meant to be enough. Almost every VA portfolio is unverifiable by design, because the work happened inside someone else’s accounts under an NDA. Someone who will not spend two paid hours proving they can do the job is telling you how they will react the first time you ask for something inconvenient.

Read their writing, not their certificates

Written communication is the one item on the standard skills list that genuinely predicts performance, because in a remote role it is the entire interface. If their text needs rewriting, you have hired a first draft.

Do not assess it from a claimed IELTS band or a VA-academy certificate, which measure test performance and attendance. Assess it from material they wrote for you, which you already have. Put one instruction in the job post that needs a real answer: “in two or three sentences, tell me what you think the hardest part of this role will be.” Read for whether they answered that question or pasted a paragraph about being a detail-oriented self-starter, whether anything is specific to your post, and whether you understand it on one read. That last point matters far more than flawless grammar. Then one live check: send a 600-word document, ask for five bullets capturing what matters. Summarising well is close to unfakeable, and it is what you will ask for constantly.

The interview questions that differentiate

Ask every candidate the same questions in the same order and write the answers down. That is what “structured” means, and it is the cheapest upgrade available to your process.

• “Tell me about a time you thought a client’s instruction was wrong. What did you do?” You want a story about raising it once, clearly, then either complying or being thanked. Worrying: “I never question the client.” That person executes your mistakes at speed. Equally worrying: they quietly did it their own way and never said so.
• “What would you need from me to do this job well?” Good answers are concrete and slightly demanding: access on day one, someone to ask when you are offline, clarity on which decisions are theirs. “Nothing, just send me the tasks” is not humility. It is inexperience, or a plan to stay invisible.
• “Let me describe the main process. Now explain it back as if I am the one who does not know it.” Listen for whether they restructure it in their own words and ask about an edge case. Reciting your sentences back is the failure mode, and it is the same failure as test one.

Most reference calls are theatre, and reference checks were never a strong predictor anyway. One question redeems them. Ask a previous client: what did you have to keep checking? Everyone says the person was lovely and hardworking. Almost nobody has an answer ready for that one, and what comes out is the residual supervision cost, which is the number you are trying to estimate. “Nothing, after the first month” is the best reference you will ever get. “Just the odd typo in client emails” tells you which work to keep away from clients.

If you are a freelancer hiring your first helper

The economics only work above a utilisation threshold. A VA at $12 an hour for 20 hours a week is about $1,040 a month. If you bill $70 an hour, you have to convert roughly 15 of those 87 freed hours into invoiced work, about 17%, just to break even. That sounds trivial and often is not, because freed hours come back in fifteen-minute fragments that do not reassemble into billable blocks. Capacity is also usually not the constraint. If you are turning work away, a VA pays for itself fast. If your calendar has gaps, freeing admin hours produces more gaps, and the fix is sales, not staff.

“I can do it faster myself” is true and mostly irrelevant. Of course you are faster, you have done it two hundred times. The comparison that matters is your speed against theirs after twenty repetitions, and teaching time is a one-off cost per process, not a recurring tax. The rule that keeps it honest: delegate anything you will do more than ten more times, and leave genuine one-offs alone unless they are long and mechanical. Recording a screen capture for a task you will never repeat is a net loss, and pretending otherwise is how people conclude VAs do not work.

Read your own client contracts before you hire anyone. This is the real risk in the freelancer case and it gets almost no discussion. Independent-contractor agreements commonly prohibit subcontracting or assignment without the client’s prior written consent, name you as key person, and permit disclosure of confidential material only to employees or agents with a need to know who are bound by equivalent terms. Hand a client’s files to a VA without checking and you may be in breach of something you signed and forgot. If the data concerns identifiable people and the work falls under the GDPR, the requirement is harder: Article 28(2) states a processor “shall not engage another processor without prior specific or general written authorisation of the controller,” and Article 28(4) requires the same data protection obligations to be imposed on that sub-processor by contract, with you still liable if they fail. Platform rules bite too. Upwork’s user agreement requires each person who uses the services to register for their own account, so putting your assistant inside your account is not a neutral workaround.

This is a five-email job, not an impossibility. Ask affected clients for written consent to use a named subcontractor under equivalent confidentiality terms, put your VA under an NDA mirroring your strictest client clause, and keep client data in systems you control rather than forwarding it. Do it before the first task, not after a client asks who sent that email.

The work goes out under your name, and you have no agency brand to absorb a bad deliverable. For the first ten client-facing outputs, nothing leaves without you reading it, then relax the gate by category rather than all at once. Keep “sends email as me, from my address” off the list longer than feels necessary.

Structuring the first engagement, and when to wait

Work arrives in bursts. Two frantic weeks, then a quiet one, so the instinct is per-task with no minimum, paying only for what you use. Wrong call. What you need is availability in the weeks you are drowning, which are the weeks everyone else is drowning too, and an assistant with no retainer from you will service whoever reserved their time. That is rational of them.

So use a floor of reserved hours, eight to ten a week, paid whether used or not, plus an agreed overflow rate beyond the floor with a notice expectation attached (surge hours need 48 hours’ warning where possible). In slow weeks you will pay for hours you did not use. That waste is the premium for having someone free in week nine, and it is cheaper than discovering in week nine that they are booked. Spend the slack rather than resenting it: in month one, put a third of the reserved hours into having the assistant write the SOPs for what you hand over, in their words, reviewed once by you.

Four situations where the honest answer is not yet, and waiting is a business decision rather than a failure of nerve.

• You cannot name eight to ten hours of recurring weekly work. Not tasks you could imagine delegating, work that happens every week. Below that you spend longer inventing assignments than you save.
• Your pipeline is the constraint, not your calendar. Freeing hours does not create demand.
• Nothing is written down and you have no time to record anything. No three hours this month for screen recordings and questions means you cannot afford the hire at any rate.
• You are at a peak and want relief this week. That pressure gets you an untested candidate, no onboarding, and a verdict a month later that it did not work. Get through the peak, hire in the trough.

If that is more process than you want to run alone, it is roughly what a recruitment partner is for. AB7 Solutions places remote professionals and administrative support through recruitment, RPO and staff augmentation, so structured screening, scored work-sample tasks and reference checks happen before a candidate reaches your shortlist rather than on your weekend. To talk through what a role like yours should be tested for, call +1 321 341 7733, email ab@ab7solutions.com or director@ab7solutions.com, or see the support roles covered at www.ab7solutions.com.

Questions that come up next

How many candidates should reach the paid test stage? Two or three. Testing is the expensive part, so filter cheaply first: the application-writing screen removes most of the field and a 20-minute structured call removes most of the rest. Testing eight people means your earlier filters are not working.

Can one task test all three qualities? Combine escalation and completion, since a multi-step task with a planted fault reads both. Keep comprehension separate: with a deliberate ambiguity and a planted error in the same brief, you cannot tell which signal you are reading when they flag only one.

What if they pass every test and fail in month two? Then it is probably onboarding rather than selection, which is good news, because onboarding is yours to fix. Testing predicts capability and disposition. It does not rescue a role handed over as forty undocumented tasks with no feedback for six weeks.

Sources: Sackett, Zhang, Berry & Lievens, “Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range,” Journal of Applied Psychology 107 (2022), 2040-2068; Schmidt & Hunter, “The validity and utility of selection methods in personnel psychology,” Psychological Bulletin 124(2) (1998), 262-274; Barrick & Mount, “The Big Five personality dimensions and job performance: a meta-analysis,” Personnel Psychology 44(1) (1991), 1-26; Regulation (EU) 2016/679 (GDPR), Article 28; Upwork, published virtual assistant hourly rate guidance; Upwork User Agreement, section 1.7.