How a Paralegal VA Handles Legal Research

It is 8:40 on a Tuesday night and a commercial attorney in Sandton is on her eleventh browser tab.

The client wants a view on whether a restraint of trade clause in an employment agreement will hold — the employee has moved to a competitor across the road, the clause runs for eighteen months, and the client is asking whether it’s worth going to court. It is not a hard question. She knows roughly where the law sits. She knows the reasonableness test, she knows Basson, she knows the enquiry turns on protectable interest and the scope of the restraint in time and area.

What she does not have is four uninterrupted hours to find the six most recent judgments that apply that test to a set of facts resembling her client’s, read them properly, check that none has been overturned on appeal, and reduce the whole thing to a two-page note her client can actually use.

So she does what most practitioners do. She spends ninety minutes half-reading, produces a memo that is competent rather than good, writes off two of the four hours because she cannot in good conscience bill a client for time spent orienting herself, and goes to bed at eleven feeling like she has not moved a single matter forward.

Multiply that Tuesday by roughly two hundred working days and you have the actual economics of a small legal practice. Not the economics in the pitch deck. The real ones.

Here is what is strange about it. There is an entire category of work inside that four hours that does not require an admitted attorney to do it. It requires someone trained, careful, supervised, and available — and almost none of it requires the person holding the practising certificate. The profession has known this for decades. It’s the whole reason paralegals exist. What has changed is that the person doing it no longer has to sit down the corridor.

The Hours Nobody Bills: What Legal Research Actually Costs

Start with the raw numbers, because they are worse than most practitioners realise until they see them written down.

The American Bar Association’s profile work has consistently put legal research at around 17% of the typical lawyer’s working time, with solo practitioners running higher — the ABA’s 2024 technology survey put solos at 23% of working hours against 19% at larger firms. Bloomberg Law’s attorney workload surveys have put associates at roughly fifteen hours a week on research alone. That is not a rounding error. That is nearly two full working days out of every five, spent on a task where most of the elapsed time is not analysis at all. It is searching, opening, skimming, discarding, re-searching, and formatting.

Now put that next to what firms actually capture. Clio’s 2025 Legal Trends Report found that for every eight hours worked, only about 2.9 are billed — a utilisation figure of roughly 38%, meaning something like five hours of every lawyer’s day disappears before it reaches an invoice. Once collection losses are accounted for, the effective capture drops closer to 2.4 hours per lawyer per day. Firm benchmarking data tells the same story from another angle: a solo running at 26% utilisation captures about 2.1 billable hours daily, while an attorney in a twenty-plus person firm captures 3.6 — roughly 71% more billable time out of the identical eight-hour day.

That gap is not explained by solos being lazier. It is explained by structure. The larger firm has someone else doing the first three hours of the four-hour research task.

An attorney at a twenty-person firm captures 71% more billable time from the same working day than a solo does. The difference isn’t talent or effort. It’s that somebody else did the searching.

Research is also the single most written-down category of work in most practices. Clio’s own analysis of write-downs identifies legal research specifically as difficult to estimate in advance and awkward to defend on an invoice. An hour comes off here because the client queried it. Thirty minutes comes off there because the partner felt it looked excessive. Nobody logs these as losses — they are absorbed as the cost of doing business. Across a year, in a practice billing R4 million, a five-percentage-point improvement in realisation is R200,000 that was always yours and that you simply gave away.

And then there’s the part that never enters the accounting at all: the mental cost of context-switching between research and everything else. A practitioner who breaks off a research task to take a client call does not resume where she stopped. She resumes twenty minutes earlier, re-reading the judgment she had already read, because the thread is gone.

What “Legal Research” Actually Contains

The reason so many practitioners resist delegating research is that they picture the wrong thing. They picture the moment of judgement — reading a paragraph in a judgment and recognising that it either does or does not help the client. That moment is genuinely non-delegable. It is what the practising certificate is for.

But that moment occupies perhaps ten minutes of a four-hour task. Break the task down honestly and it looks more like this:

Layer one — scoping and framing. Turning “can we enforce this restraint” into a set of searchable legal questions: what is the applicable test, which court’s decisions bind us, what timeframe matters, what factual features are legally significant. This is attorney work. It stays with the attorney.

Layer two — retrieval. Running searches across SAFLII, Juta, LexisNexis, government gazettes, regulator publications and, where relevant, foreign persuasive authority. Building a candidate list. Pulling full texts. Downloading and naming files consistently. This is not attorney work. It is careful, methodical, trainable work.

Layer three — triage. Reading enough of each candidate authority to determine whether it is on point, tangential, or irrelevant, and sorting accordingly with a one-line note explaining the call. This is where a well-trained paralegal earns their keep. It requires legal literacy, not legal authority.

Layer four — verification. Checking whether each authority is still good law: subsequent history, appellate treatment, whether the statute has been amended, whether a subsequent judgment has distinguished or overruled it. Utterly critical. Utterly mechanical. Completely delegable to someone trained to do it properly.

Layer five — summarisation. Producing a structured research note: the question, the authorities in order of weight, the holding of each in two or three sentences, the pin cites, the factual distinctions that matter, and the open questions. Delegable.

Layer six — application. Reading the note and deciding what it means for this client, on these facts, given this client’s appetite for risk and cost. Attorney work. Non-delegable. And now it takes forty minutes instead of four hours, because the attorney is starting from an organised, verified, cited summary rather than a search box.

A paralegal virtual assistant handles layers two through five. That is the entire proposition. It is not exotic and it is not new — it is precisely the division of labour that every large firm has run internally for a century. What has changed is that a five-person practice in Durban can now access it without hiring, housing, and carrying a full-time employee.

How a Paralegal VA Actually Runs a Research Task

The theory is easy. What separates a paralegal VA who saves you time from one who creates more work is the workflow. Here is what a properly run research instruction looks like in practice.

The brief in, not the question in. The attorney does not send “look into restraints of trade.” She sends a structured instruction: the legal question, the client’s key facts, the jurisdiction and court level that matters, the date range that matters, the deadline, and the output format. At VAConnect this is a standing template, because the quality of a research note is capped by the quality of the instruction. Two minutes spent framing saves an hour of the wrong search.

Acknowledgement and scoping check within the hour. A good paralegal VA reads the instruction and comes back with either “understood, delivering by 14:00 Thursday” or one clarifying question. Not five. One. If a task generates five clarifying questions, the brief was underspecified and the VA should say so directly rather than guessing.

The search log. This is the part that separates professional research support from someone Googling on your behalf. Every search run is logged: the database, the exact query string, the date, and the number of results reviewed. When the attorney later asks “did we check whether anything post-2023 changed the position,” the answer is documented rather than remembered. It also means the research can be picked up by someone else if the VA is on leave — which, in a managed model, is a scenario that has been planned for rather than discovered.

Triage with reasons. Each authority pulled gets a one-line disposition: on point, analogous, distinguishable on facts, superseded, not relevant — retained for completeness. The attorney can scan a triage table in ninety seconds and immediately spot if the VA has miscalibrated. That’s the supervision loop working as designed.

Verification, always, without being asked. Every authority in the final note has been checked for subsequent treatment. Every statutory provision has been checked against the current consolidated text, not against whatever version appeared in a 2019 blog post. This is non-negotiable, and it is the single most important habit to train for.

The research note. Structured, consistent, and short. Question. Short answer. Authorities in descending order of weight with holding, pin cite and relevance. Contrary authority — flagged deliberately, because a research note that only contains helpful cases is worse than useless. Gaps and open questions. Sources consulted.

The handback. The attorney reads, applies judgement, and — critically — gives the VA feedback on the triage calls. Over eight or ten matters, the VA’s calibration to that specific attorney’s practice area and preferences tightens dramatically. This is the compounding return that a one-off freelancer engagement structurally cannot deliver.

Note what has not happened anywhere in that workflow. The paralegal VA has not given legal advice, has not decided what the law means for the client, has not signed anything, and has not communicated a legal position to a third party. They have prepared material for an attorney’s judgement. The line is bright and it is easy to hold, provided the workflow is designed to hold it.

The Human in the Loop: Why Unsupervised AI Research Became a Liability

There is an obvious objection to all of this. Why pay a person to do layers two through five when a language model will produce a research memo in forty seconds?

The empirical answer to that question is now one of the best-documented failures in professional services, and it is worth sitting with.

Start with the tools built specifically for this. In a preregistered study published by Stanford’s RegLab and HAI — the first independent evaluation of commercial legal AI research products — researchers tested LexisNexis’s Lexis+ AI, Thomson Reuters’s Westlaw AI-Assisted Research, and Ask Practical Law AI across more than two hundred legal queries, hand-scored by legal experts. These are retrieval-augmented systems, marketed with language about being grounded in trusted content and avoiding fabrication.

They hallucinated between 17% and 33% of the time. Lexis+ AI, the best performer in the study, answered 65% of queries accurately. Westlaw’s AI-Assisted Research was accurate on 42%. Ask Practical Law returned incomplete answers — refusals or ungrounded responses — on more than 60% of queries. General-purpose GPT-4, tested alongside them, sat around 43% hallucination.

The subtler finding got less attention and matters more: a large share of the errors were not invented cases but misgrounding — real citations attached to propositions the cited authority does not actually support. A fabricated case name is easy to catch, because it does not exist. A real case cited for something it never held will survive any check short of someone opening the judgment and reading it.

A fabricated citation is a problem you can catch with a search. A real citation attached to a holding that isn’t in the judgment is a problem you can only catch by reading. That reading is the job.

Now look at what happened when the profession skipped that reading. The public database maintained by Damien Charlotin, a research fellow at HEC Paris, catalogues court decisions worldwide in which a party relied on AI-fabricated material and a court responded. It held roughly 200 cases in mid-2025. By January 2026 it was at 719. By early April, 1,227. By 9 June 2026, 1,598 — with more than a thousand of those in the United States alone.

The penalties have escalated in step. The original 2023 sanction in Mata v. Avianca, where two New York lawyers filed six cases ChatGPT invented, was $5,000. By 2026, US courts had imposed over $145,000 in AI-filing penalties in the first quarter alone, with individual matters reaching six figures. In February 2026 an Omaha attorney filed a divorce appeal brief in which 57 of 63 citations were defective, including twenty entirely hallucinated cases; the Nebraska Supreme Court suspended him in April. In June 2026, a federal judge in the Northern District of Mississippi found fabricated citations in filings from both sides, cancelled the trial, and suspended the two lead attorneys from the district for two years.

There is a pattern running through the sanctions decisions that every practitioner should internalise: the cover-up draws the harsher penalty. Courts have been comparatively forgiving of the initial error and extremely unforgiving of the lawyer who denied using AI, blamed a junior, or defended the fake cases after being warned. Candour with the court is the one obligation that cannot be delegated to software.

None of this is an argument against AI in legal research. AI is genuinely useful for the parts of the task it is suited to: generating search strategies, producing a first-pass summary of a judgment you have already opened, surfacing terminology you had not thought of, drafting the skeleton of a note. The trained paralegal VAs at VAConnect use these tools daily and are faster for it.

The argument is against AI with nobody attached to it. A research note is not a document — it is a chain of accountability that ends at a person who read the judgment. Remove the person and you have not saved four hours; you have converted four hours of research into a contingent liability with an unknown maturity date. The human in the loop is not a nostalgic preference. It is, at this point, the only structure the courts have accepted.

Supervision, Privilege and Getting the Structure Right

The second objection is the serious one, and it deserves a proper answer rather than reassurance: is it ethical, and is it safe?

The framework is well established. The American Bar Association addressed it directly in Formal Opinion 08-451: a lawyer may outsource legal and non-legal support services provided she remains ultimately responsible for competent representation under Rule 1.1, exercises supervision under Rules 5.1 and 5.3, preserves confidentiality under Rule 1.6, avoids assisting the unauthorised practice of law under Rule 5.5, and bills reasonably under Rule 1.5. Multiple state bars — Florida, Washington, and others — have issued their own guidance confirming the same position. The North Carolina State Bar’s 2024 formal ethics opinion confirmed that the supervision duty extends to all outsourced third parties and software providers alike.

South African practitioners operate under the Legal Practice Act and the Legal Practice Council’s Code of Conduct rather than the ABA Model Rules, but the architecture is the same in substance: the attorney of record carries the professional responsibility, non-admitted staff work under supervision, the reserved work stays reserved, and client confidentiality travels with the information wherever it goes.

Four things make the difference between an arrangement that satisfies this and one that does not.

Scope discipline. The paralegal VA prepares research; they do not advise. Nothing leaves the practice under their name. Nothing is communicated to a client or opposing party by them without an attorney’s review. Put this in writing at the outset — it protects everybody, including the VA.

Real confidentiality infrastructure. Signed NDAs are the floor, not the ceiling. What actually matters is access control: matter-scoped permissions rather than blanket access to the document management system, encrypted transfer, no client documents on personal devices, and a documented offboarding process. For South African practices, POPIA obligations follow the personal information into the VA’s hands, and the operator agreement needs to say so explicitly. For firms serving UK or EU clients, the same applies under GDPR, and South Africa’s POPIA framework was drafted with close reference to it — which is why compliance conversations with European clients tend to be short.

Conflicts screening. A VA working across multiple firms in the same practice area is a conflicts problem waiting to happen. This is one of the clearest arguments for a managed agency over an independent freelancer: VAConnect screens and controls client assignments as a matter of policy, whereas a freelancer’s other clients are a matter you find out about later, if at all.

Documented supervision. The search log, the triage table, the review trail. If a research call is ever questioned, you want to be able to show the process rather than describe it.

The ethics rules have never prohibited delegating legal research. They prohibit delegating responsibility for it. Those are different things, and the second one has never been available to anybody.

The South African Advantage

If you are a firm in London, Dublin, Amsterdam or Dubai looking for research support, or a South African practice trying to hire it locally, the case for South African paralegal talent is stronger than most people outside the country realise. Three factors compound.

The legal system is genuinely compatible. South Africa is a mixed jurisdiction — Roman-Dutch civilian foundations overlaid with English common law procedure — but the working method of a South African legal researcher is common-law method. English became the language of the courts in 1827. The country follows English law in criminal and civil procedure, company law, constitutional law and the law of evidence. Court procedure is adversarial, judgments are fully reported with dissents, and the system runs on stare decisis. A South African law graduate has spent four years reading judgments, extracting ratio from obiter, tracing lines of authority, and distinguishing cases on their facts. Drop that person into English, Australian, Canadian or American case law and the method transfers immediately, even though the substantive law does not. That is a meaningfully different starting point from a research assistant trained purely in a codified civil-law system.

The talent pool is deep, credentialed, and structurally under-employed. This is the uncomfortable part, and it is the reason the quality available at the price is what it is. South Africa produces well over 5,000 LLB graduates a year and has done for a decade. The profession cannot absorb them. In one representative cycle, only about 2,863 articles of clerkship were registered against that year’s output — barely 55%. Law Society of South Africa figures put roughly 6,294 candidate attorneys against 33,929 practising attorneys, with most firms taking one or two candidates at a time because each requires a qualified principal to supervise. Graduate unemployment in South Africa roughly doubled between 2008 and 2023.

The result is a large population of people with four years of formal legal training, strong research and writing ability, and no route into the profession through the traditional door. For a firm in Manchester or Cape Town that needs someone to run authoritative searches and produce clean research notes, this is a labour market almost nobody is competing in properly.

The time zone does something no other outsourcing destination does. South Africa sits at GMT+2. That is one or two hours ahead of the UK, the same as or one hour off most of Western Europe, and — this is the part that matters — a full working day that overlaps almost completely with a London or Amsterdam working day. A research instruction sent at 9 a.m. in London reaches a paralegal VA already three hours into their day. Questions get answered in real time, not on a twelve-hour delay. Compare that with the standard offshore model, where the practical rhythm is one round-trip per day and every misunderstanding costs twenty-four hours. For US East Coast firms the overlap is narrower but still workable: a South African VA finishing at 5 p.m. local has covered the American morning, which means research briefed the previous afternoon is on the desk before the attorney’s first coffee.

The cost picture then follows from all of this rather than leading it. US Bureau of Labor Statistics data puts the median paralegal wage around $61,010, before benefits and overhead — figures that reach $75,000–$90,000 in major metros and approach six figures fully loaded. UK paralegal salaries average around £24,000 before employer costs. Managed offshore paralegal support typically runs a fraction of that, with the added structural benefit that it is a variable cost that scales with caseload rather than a fixed one that idles when matters are quiet.

But the cost saving is the least interesting part of the argument, and any provider leading with it is telling you what they compete on. The interesting part is that a firm can get common-law-trained research support, in its own working hours, in native-standard English, at a price that makes it viable for a five-person practice rather than only for a fifty-person one.

Managed, Not Matched

Here is where most firms get burned, and it has nothing to do with geography.

The dominant model for hiring remote support is matching. A platform shows you profiles, you interview three, you pick one, and from that moment you are the employer, the trainer, the quality controller, the HR department and the continuity plan. The platform’s job ended at introduction. When the person you hired disappears for a week, gets a better offer, or turns out to have been running four other clients at the same time, that is your problem to solve — usually mid-matter.

For legal research specifically, that model fails in three predictable ways.

Training cost falls entirely on you. Getting a research assistant genuinely useful in your practice area takes six to ten matters of feedback. If they leave at matter eleven, you pay that cost again from zero.

Quality is unverified until it fails. A freelancer’s research note looks correct. You discover it was not on the day opposing counsel points out that the authority you relied on was overturned in 2022.

There is no second person. Deadlines do not pause for illness. A freelancer with flu is a matter that slips.

VAConnect’s model is managed rather than matched. The VA is recruited, vetted, trained and employed by VAConnect, with an account manager who owns the relationship and quality outcomes, structured onboarding into your systems and preferences, documented processes so the work is transferable, and continuity cover when your VA is unavailable. Training investment sits with the agency, which means it compounds rather than resets. VAVarsity, VAConnect’s internal upskilling platform, exists precisely so that capability is built systematically rather than picked up ad hoc.

The founder story behind this is not incidental. Karen van Zyl built VAConnect from 2008 on the premise that South African work ethic and skills were an underused global resource, and that the failure mode of remote work was never talent — it was management. Africa’s largest managed VA agency was built around solving the management problem, not the sourcing one.

For legal work, that distinction is not a preference. It is the difference between an arrangement your professional indemnity insurer would recognise as reasonable supervision and one they would not.

You cannot supervise what you cannot see, and you cannot delegate to someone whose training you did not verify. The hiring model isn’t an administrative detail in legal work. It’s the control.

What Changes in the First Ninety Days

Firms that make this work follow roughly the same sequence.

Weeks one to two: pick the boring work first. Do not start with a complex advisory research question on a live high-value matter. Start with the tasks you already resent: pulling and organising authorities for a matter you have already scoped, verifying the current status of a list of cases you rely on repeatedly, building a precedent bank for a practice area you handle often. The point is calibration under low stakes.

Weeks three to six: build the templates. Standardise the research instruction format and the research note format. Both should be one page. Consistency here is what turns a capable assistant into a fast one, because neither party is re-deciding the format each time.

Weeks six to twelve: extend the scope and start measuring. Move into live research on active matters with proper review. And begin tracking, because otherwise you will have a strong feeling and no evidence. Three numbers are enough:

Hours per research task, before and after.
Research hours billed as a proportion of research hours worked — your research realisation rate.
Turnaround: instruction sent to note delivered.

Most firms find the third number is the one that changes their practice most. When a research note reliably comes back in twenty-four hours rather than “when I get to it,” attorneys start asking questions they previously did not bother asking. The advice gets better because the cost of checking has fallen.

The Gap Is Wider Than It Should Be

What is striking about all of this is how unevenly the profession has adopted something so unremarkable.

Large firms have run leveraged research models for a century. The associate does layers two through five; the partner does layers one and six. Nobody considers this innovative. It is simply how the work is organised, and it is a substantial part of why a twenty-lawyer firm captures 71% more billable time per lawyer than a solo does.

The small and mid-sized end of the profession has largely not had access to that structure — not because the work was different, but because the fixed cost of a full-time employee was too high for a caseload that fluctuates. So thousands of capable practitioners have spent a decade doing three hours of retrieval and verification for every hour of judgement, writing off the difference, and treating this as an immutable feature of practising alone.

It stopped being immutable somewhere around the point that South African legal graduates started working remotely for firms four thousand kilometres away, in the same working hours, at a cost a three-person practice can absorb. The firms that noticed early are not working fewer hours. They are billing more of the hours they already work, answering research questions they used to skip, and turning around client advice in two days instead of two weeks.

The ones that have not noticed are still on tab eleven at 8:40 on a Tuesday night.

DIY vs Freelance vs VAConnect

	DIY Coordination (attorney does everything)	Generic Freelancer (matched, not managed)	VAConnect Paralegal VA (managed)
Time to first useful output	Immediate, but at full attorney cost	2–6 weeks of trial and error	5–10 working days, structured onboarding
Attorney hours per research task	3.5–4.5 hrs	2–3 hrs (review + rework)	0.5–1 hr (scope + apply judgement)
Research realisation	Low — heavily written down	Variable, often unbilled review time	High — delegated hours cost less than they recover
Verification of authority status	Done if time allows	Depends entirely on the individual	Standard step in every task, logged
Search audit trail	Rarely documented	Rarely documented	Documented as standard
Legal-method training	N/A	Unverified	LLB / legal-studies background, vetted
AI use	Ad hoc, unsupervised risk	Unknown and undisclosed	Tool-assisted with mandatory human verification
Confidentiality & conflicts control	Fully internal	NDA only; other clients unknown	NDA, POPIA/GDPR-aligned handling, conflicts screening, access control
Cover when unavailable	None — work stops	None — deadline slips	Managed continuity cover
Who owns quality	You	You	VAConnect account manager
Cost structure	Highest — opportunity cost of attorney time	Low rate, high hidden cost	Predictable monthly; scales with caseload
Improvement over time	Flat	Resets when freelancer leaves	Compounds — training investment retained

Get the Three Hours Back

Legal research is not one task. It is six, and only two of them require you.

If you are still doing all six, you are not being thorough — you are being under-resourced, and it is costing you both the hours you write off and the questions you never get around to asking.

Explore VAConnect’s Paralegal VA service to see how a trained, supervised paralegal virtual assistant fits into your practice, or book a discovery call and we will map which layers of your research workload are delegable this month.

Sources

Stanford RegLab & HAI — Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Magesh et al., published 2025, 22 J. Empirical Legal Stud. 216)
Damien Charlotin, AI Hallucination Cases Database, HEC Paris Smart Law Hub (case counts as at June 2026)
American Bar Association — Formal Opinion 08-451; ABA Profile of the Legal Profession; ABA Legal Technology Survey 2024
Clio — Legal Trends Report 2025; analysis of law firm write-downs
Bloomberg Law — Attorney Workload and Hours Survey
Law Society of South Africa — Statistics for the Attorneys’ Profession (2025); LSSA LEAD Statistics 2017/18
US Bureau of Labor Statistics — Occupational Employment Statistics, paralegals and legal assistants
Norton Rose Fulbright — AI in litigation: Update on Gen AI sanctions in 2026
Georgetown Law Library and NYU GlobaLex — guides to South African legal research and the mixed legal system