Support Staff, Specialists and AI in a Legal Team

Summary

Every firm has work that demands precision and consumes hours. Compiling the court book. Numbering annexures. Building an index. Reading two thousand documents to find the four that matter. Turning a rough note into a letter that can go out.

It has to be done, it has to be done properly, and it takes an enormous amount of professional and support time to do well.

This guide is about how that work gets routed: to specialist teams, to outsourced help, and increasingly to AI. But it starts with an unfashionable claim: none of those options are available to a firm that hasn't documented its work. You cannot delegate, outsource or automate a process that exists only as a habit in somebody's head. The firms getting real value from AI aren't the ones who bought the best tool. They're the ones whose work was already captured.

We should be direct about our own position here: Hivelight has no built-in AI. What it has is structure and context about the work, which turns out to be the missing link for both delegating well and getting anything useful out of AI. Neither an experienced paralegal nor the best AI on the market can do good work without knowing what a matter actually requires and what it's building toward.

Contents

• Precision work at volume
• The ladder shifts up, and nobody's rung disappears
• What you actually gain is fatigue, not headcount
• Why your firm's AI probably isn't helping yet
• The progression: document, classify, then choose
• What's worth a first pass today
• Confidentiality isn't an afterthought
• Letting agents do the work in your systems
• Using AI to build the system itself
• Where to start
• Key takeaways
• Frequently asked questions

Precision work at volume

Start with an honest observation about the work itself, because the usual framing gets it wrong.

A meaningful share of legal work is exacting rather than discretionary. Bundling. Indexing. First-pass reading. Boilerplate drafting. These tasks don't call for legal judgement so much as they call for accuracy, consistency and completeness, sustained over a great many pages.

And they are genuinely skilled. Someone good at assembling a court book knows what a properly built one looks like, spots the annexure that's out of sequence, and knows from experience which small errors cause trouble down the track. That expertise is real and mostly invisible: a well-assembled bundle draws no attention at all, which is precisely the mark of it being done well.

The difficulty isn't the work. It's the ratio. These tasks consume hours out of all proportion to the judgement they require, and precision at volume is exactly where human attention is most expensive to sustain. Not through carelessness, but because the four hundredth page can never have quite the freshness of the fourth. The expertise sits in knowing what right looks like. The hours go into producing it.

Which is a familiar problem in an unfamiliar place. It's the delegation argument one rung further out: the scarce resource is judgement, and spending it on execution is what caps what a firm can produce.

Historically this work went to support staff, and that was the right answer. It often still is. But the range of options has widened, and firms that can take advantage of that are competing against firms that can't.

The ladder shifts up, and nobody's rung disappears

Before going further, it's worth heading off the anxious version of this conversation.

The framing that gets pushed, that AI replaces paralegals, is both wrong and unhelpful. What's actually happening is more interesting: these are tasks support staff used to first-pass, and now support staff and lawyers alike can have AI first-pass for them.

The whole ladder shifts up. The paralegal who used to spend a day building an index now reviews a draft index and spends the rest of the day on work that was previously sitting with a junior lawyer. The junior lawyer moves up in turn. Nobody's rung disappears; everyone's rung moves.

And it's worth being clear about what that review actually requires, because it's easy to assume checking is the lesser job. It isn't. Reviewing a first pass demands more expertise than producing one, not less. You have to already know what right looks like in order to see what's wrong, and machine output fails in ways that are quieter and more plausible-looking than human error. The person who has assembled two hundred court books is exactly the person you want reading the draft, because they are the only one who will notice the thing that is subtly out of place.

So the expertise doesn't become redundant. It becomes the part of the job that matters most, applied to more matters than one person could previously have touched.

What you actually gain is fatigue, not headcount

Which raises the obvious question. If a person still has to review everything, what has the firm really gained?

The best answer is an analogy from outside law. Think about supervised self-driving in a car. The driver hasn't stopped driving. They're still responsible, still watching the road, still the one who has to react. What's changed is that they're no longer spending attention on the unremarkable parts: holding a steady speed, staying in the lane, mile after mile after mile. And because those miles no longer draw down their concentration, they arrive with something left in reserve for the moments that genuinely need them.

That is much closer to what AI does with exacting legal work than "it does the task for you." The person remains responsible and remains in the loop. What lifts is the fatigue of sustaining precision across four hundred pages, which means the same person covers more ground at the same standard, and is still sharp for the judgement calls waiting at the end of it.

This is why the benefit is so often mis-stated. Firms look for the hours saved and miss the more valuable effect: the same people, working at the same standard, further into the day. Fatigue is the real constraint on precision work, and it's the constraint that lifts first.

Key point: The gain isn't that a human stops checking. It's that they arrive at the checking with their attention intact.

"This is the next iteration of effective collaboration: professional staff, support staff, and AI supporting them both."

— Ashley Kelso, Hivelight

That's the useful way to think about it, and it's also the accurate one. A firm that treats AI as a headcount question will get a one-off cost saving. A firm that treats it as a capacity question will get a higher ceiling, which is worth considerably more.

Key point: AI doesn't remove the bottom rung of the ladder. It moves everyone up one, provided you actually redeploy the time you free up.

Why your firm's AI probably isn't helping yet

Most firms reading this already have AI. It came with the practice management system, or with the office suite, or someone signed up to something. And for a lot of firms, the honest assessment is that it hasn't changed much.

That's rarely the tool's fault, and the reason is worth understanding because it determines whether the next tool will work either.

Every AI product that promises to learn from your firm's data carries an unstated precondition: the data has to exist. Not documents; those you have. What's usually missing is the structured record of the work itself: what happens on this matter type, in what order, who does each part, what "done" looks like, what stage this particular file is at. If that was never captured, there's nothing for the AI to learn from and nothing for it to plug into. It can summarise a document you point it at. It can't tell you what should happen next, because nobody ever wrote down what happens next.

"Documenting the work as you go gives your support staff, and any AI you use, the context to do the job properly. Both need the same thing: to understand what they're doing and what it feeds into."

— Ashley Kelso, Hivelight

This is the same argument as the rest of this hub, one rung further out. A new paralegal and an AI agent need remarkably similar things: a clear task, an instruction, and an understanding of what the work feeds into. A firm that captures its work progressively is building that context for both at once, and a firm that doesn't will struggle with both for the same reason.

Why adding AI can make a firm busier

There's a second-order effect worth understanding, because it explains a common and demoralising experience.

Every participant in a piece of work carries a coordination cost. Someone has to brief them, check what came back, and route it onward. In a firm that coordinates manually (by email, by conversation, by remembering) that cost is paid in a person's time, and it grows with each new participant.

An AI agent is a new participant. So dropping one into a firm that coordinates manually doesn't reduce the coordination load; it adds to it. Somebody now assembles the context by hand, pastes it in, reviews the output, and carries the result back to wherever the work actually lives. The mechanical task got faster and the surrounding admin got heavier, which is why so many firms report that AI feels like a lateral move.

When the work is structured, that overhead disappears, because the agent picks up from the same place everyone else does, and puts its output back there too. No assembling context by hand, because the context is already sitting on the task.

Key point: AI isn't underperforming in your firm because it's not clever enough. It's underperforming because it has no context, and someone is paying, in their own time, to supply by hand what the system should have provided.

The progression: document, classify, then choose

Here's the sequence, and the order isn't optional.

1. Document the work. Capture what actually happens on a matter, as it happens.
2. Classify it. Not just individual tasks, but types of task. This is the step firms skip, and it's the one that unlocks everything below. Once you can see that forty matters each contain the same category of work, you're looking at a workload rather than forty separate errands.
3. Specialise. Stand up internal teams that work across matters, handling a class of task in bulk. Doing thirty of the same thing well is a different job from doing thirty different things adequately, and people get markedly better at it.
4. Outsource. Certain task classes can go to external or offshore teams, which is only safe once the process is defined well enough to hand over and check.
5. Automate. Some classes can go to AI or to automated workflows.

Steps three, four and five are simply unavailable to a firm that hasn't done one and two. You cannot bulk-allocate, outsource or automate work you can't describe. This is why "we should use more AI" so often stalls: it's a step-five ambition in a firm that hasn't done step two.

Figure 1 — Every routing option depends on the two steps most firms skip.

It's also the strongest argument for treating documentation as an investment rather than an overhead. It isn't admin. It's the thing that makes every subsequent option possible.

What's worth a first pass today

Practical territory, kept at the level of task classes rather than tools, because tools change every few months, and the categories don't.

Bundling and preparation. Compiling court books, numbering annexures, building indexes. High volume, exacting, with a well-defined right answer, which is exactly what makes a first pass checkable.

Reading at scale. Discovery review, summarising long documents, searching a large set for the material that's actually relevant. The classic needle-in-haystack work that consumes days.

Drafting the repetitive parts. The boilerplate and structural elements of pleadings and standard correspondence: the parts that are the same every time, leaving the judgement-heavy sections to a lawyer. Also turning rough file notes into letters that can actually go out.

Research and process. First-pass legal research and, often overlooked, working out the procedural requirements: what a particular court or government body actually requires, from practice notes and published guidance.

Capture. Dictation with automatic formatting, which quietly removes the need to learn dictation as a distinct skill. That's a bigger deal than it sounds for anyone who never got fluent at it.

In every case the pattern is the same: AI does the first pass, a human does the judgement pass. The output is a draft that someone checks, not a finished product. Firms that get this wrong in either direction, treating output as final or ignoring it entirely, get worse results than firms that treat it as an eager junior whose work always gets reviewed.

Confidentiality isn't an afterthought

No firm can sensibly adopt any of this without dealing with client confidentiality and privilege first. It belongs near the front of the decision, not in a policy document written afterwards.

The core question is straightforward: what happens to your client's information after the AI has processed it? Some services retain inputs, and may use them to improve their models. For legal work, that is usually the wrong answer.

Two approaches address it. Run models locally, so nothing leaves your environment. Or use hosted services that contractually offer zero data retention: inputs processed and discarded, not stored, not trained on. Whichever route, the requirement is the thing to hold firm on, and it's worth confirming in writing rather than inferring from a marketing page. Providers' terms change.

Your professional obligations are a separate matter from the technical ones, and they're not ours to interpret for you. The Federal Court has issued a practice note on the use of generative AI, several state Supreme Courts have their own, and the law societies have published guidance for practitioners. Those are the authoritative sources, so go to them directly, and make sure whoever leads this in your firm has read them properly rather than relying on a summary.

Key point: Decide your data-handling position before you choose a tool, not after. It rules out options, and it's much cheaper to rule them out early.

Letting agents do the work in your systems

Here's where things are heading, and it's worth being precise rather than breathless.

The interesting shift isn't AI that answers questions. It's AI that picks up assigned work, does it, and reports back, participating in the same flow of work as everyone else, rather than sitting in a separate window waiting to be asked something.

That requires the work to exist somewhere an agent can reach. To be clear about Hivelight's position: it has no built-in AI features. What it has is a documented public API, available on every plan, and a published Zapier app, and that turns out to be enough.

The Hivelight Zapier integration is the clearest illustration, because you can go and look at it rather than taking our word for it. It exposes triggers for Matter Created, Task Created, Task Status Updated, Milestone Created and Milestone Status Updated, and around twenty actions including applying a workflow to a matter, creating matters, updating task and milestone status, and adding notes to tasks.

Read that list as a loop rather than a feature list. An agent can be woken by Task Created, do the work, and report back via Update Task Status and Create a Task Note, visible to human colleagues in exactly the same place they'd see a paralegal's progress. The same surface connects to Clio, Actionstep, Smokeball, Outlook and the rest of a firm's stack.

Two honest caveats. This is configuration work. Every firm has different processes and a different mix of apps, so there's no download-and-go version. And it only becomes worth doing once your workflows are documented well enough that there's something for an agent to pick up. Which brings us back to step one.

Using AI to build the system itself

One more application, and strategically it might be the most useful of all.

The biggest obstacle to systematising a firm has always been the sheer work of writing the workflows down. Everyone agrees it should be done. It requires someone to sit with a matter type and articulate every step, and that person is invariably the busiest person available. So it doesn't happen, or it happens for two matter types and stalls.

That obstacle is now much smaller. Point an AI at a handful of real, completed matters and ask it to describe the process actually followed: the stages, the tasks, the order, the typical timings. What comes back isn't a finished workflow. It's a first draft of one, which is a dramatically easier thing to edit than a blank page is to fill.

"Once a firm documents its work centrally, the opportunities to use AI well start to identify themselves: more billable output, and fewer hours spent on work that doesn't need a person doing every step of it."

— Ashley Kelso, Hivelight

There's a pleasing circularity to it. The thing that makes AI useful to a firm is documented work; and AI is now one of the better ways to get the work documented in the first place.

Where to start

Not with a strategy, and definitely not with a committee.

Pick one class of task your team does forty times a month. Something repetitive, well-defined, and consuming more hours than the judgement in it warrants. Write down how it's actually done, not how it should be done, how it is done. Ask the people who do it; they'll know things about it you don't. Then decide whether it's best handled by a specialist internal group, an outsourced team, or automation.

You'll learn more from routing one task class properly than from six months of evaluating platforms. And whatever you learn, the documentation you produced along the way is valuable on its own, which makes it the rare experiment that pays off even when the answer is "not yet."

Key takeaways

• Exacting, high-volume work is a routing question, and firms have more options for it than they used to. The problem was never the work. It's the hours it consumes relative to the judgement it needs.
• The ladder shifts up; nobody's rung disappears. Support staff and lawyers both get a first pass done for them.
• Reviewing a first pass takes more expertise than producing one. You have to know what right looks like to see what's wrong, so the people who do this work well become more central, not less.
• The real gain is fatigue, not headcount. Like supervised self-driving, the human stays responsible and stays watching. They just aren't spending their concentration on holding the lane. Same standard, more ground covered, attention intact for the judgement calls.
• AI underperforms in most firms because it has no context, not because it isn't capable. Documented work is the context.
• The progression is strictly sequential: document → classify → specialise → outsource → automate. Steps three to five are locked until one and two are done.
• Think in task classes, not tools. Tools change constantly; the categories of work don't.
• AI does the first pass; a human does the judgement pass. Always.
• Settle your data-handling position first — local models or contractual zero data retention, then go to the courts and law societies directly on professional obligations.
• Hivelight has no built-in AI. It has a documented API and a published Zapier app, which is what lets agents participate in the same flow of work as everyone else.

Want to see what structured work looks like in practice? Take a look at Hivelight →

Frequently asked questions

Will AI replace paralegals?

No, and the framing misses what's actually happening. These are tasks support staff used to first-pass, and now support staff and lawyers alike can have AI first-pass for them. The whole ladder shifts up rather than losing its bottom rung. Reviewing a first pass also demands more expertise than producing one, because you have to already know what right looks like to spot what's wrong, and machine output fails in quieter and more plausible-looking ways than human error does. The people who do this work well become more central, not less.

Which legal tasks can AI genuinely help with today?

Think in task classes rather than tools, because tools change every few months and the categories don't. Bundling and preparation such as court books, annexures and indexes. Reading at scale, including discovery review and document summarising. The repetitive parts of drafting, like boilerplate and structural elements, and turning rough file notes into letters. First-pass research, including procedural requirements from practice notes. And capture, such as dictation with automatic formatting. In every case AI does the first pass and a human does the judgement pass.

Is it safe to use AI on client matters?

Only once you've settled what happens to the information after it's processed. Some services retain inputs and may use them to improve their models, which is usually the wrong answer for legal work. The two approaches that address it are running models locally, so nothing leaves your environment, or using hosted services that contractually offer zero data retention. Confirm that in writing rather than inferring it from a marketing page, because terms change. Your professional obligations are a separate question, and the authoritative sources are the courts and the law societies rather than any vendor, including us.

Does Hivelight have AI built in?

No. Hivelight has no built-in AI features, and we'd rather say so plainly than imply otherwise. What it has is a documented public API available on every plan, plus a published Zapier app whose triggers and actions let an AI agent pick up assigned work, do it, and report progress back where human colleagues can see it. Configuration is bespoke per firm, because every firm has a different mix of processes and apps.

Why hasn't the AI we already have made much difference?

Almost certainly because it has no context. Every product that promises to learn from your firm's data carries an unstated precondition: the data has to exist. Documents you have. What's usually missing is the structured record of the work itself, which is what tells an AI what a task is, what standard it's held to, and where it sits in the matter. There's also a second-order effect worth knowing: an AI agent is another participant, and in a firm that coordinates manually, adding one adds coordination work rather than removing it.