Legal AI Agent vs Paralegal vs Outsourcing: Small Firm Math
The question behind “should we be using AI?”
When a partner at a six-lawyer firm asks about law firm AI automation, they’re almost never asking a technology question. They’re asking a capacity question: we are behind, the work is piling up, what is the cheapest reliable way to get more done without wrecking quality?
There are exactly three answers to that question. Hire a person. Rent a person (outsourcing, contract paralegals, virtual assistants, records-retrieval vendors). Or build leverage with software — which, as of 2026, increasingly means AI agents that carry out multi-step tasks rather than chatbots that answer questions.
They’re not interchangeable, and the honest comparison is more useful than another list of ten AI tools.
What each option is actually good at
An AI agent is software that can be given a goal, pull the relevant documents or records, take a sequence of actions across your systems, and produce a defined output — a first-draft chronology, a coded review set, a drafted response, a matter opened in your practice management system. The mature way to run one is to connect the assistant to the systems you already have through governed integrations (MCP, the open Model Context Protocol, is the emerging standard for this), and to package the firm’s instructions as reusable skills so it does the job the same way every time.
A paralegal is judgment plus accountability plus hands. They can call the clerk, notice that the client sounded off on the phone, catch that the caption is wrong because they filed in that court last month, and take responsibility for a task end-to-end.
Outsourcing deserves its own line item, not a footnote
Outsourcing buys elastic throughput without a payroll commitment: overflow document review, records retrieval, after-hours intake answering, overnight transcription, e-filing support.
Cost structure is the part firms misjudge. Vendors price three ways — per hour (staffing-style, cheapest headline, hardest to forecast), per unit (per page, per record request, per call answered, which makes budgeting easy and rewards you for sending clean inputs), or per seat with a monthly minimum (you pay for the floor whether or not you use it). Compare like with like by converting everything to cost-per-completed-unit, and add your own management time, because someone at your firm will be answering the vendor’s questions.
Quality varies by person, not by contract. The same vendor can produce excellent work from one reviewer and unusable work from the next, so ask who specifically is assigned, whether the team is stable, and what the sample-and-check process looks like. Turnaround is a genuine advantage — time-zone coverage means work moving while you sleep — but confirm the cutoff times, not the marketing promise.
Diligence you owe. ABA Formal Opinion 08-451 addresses a lawyer’s obligations when outsourcing legal and nonlegal support services; read it before you sign. Practically: run a conflicts check on the vendor’s other clients where relevant, get written confidentiality terms, ask whether work is subcontracted further and where data physically sits, and confirm what the client needs to be told.
And note what is not a point of difference: ABA Model Rule 5.3 already makes lawyers responsible for the conduct of nonlawyer assistants — contract paralegals and outsourced vendors included — and Formal Opinion 512 applies the same supervisory logic to generative AI. You are accountable in all three scenarios. The options differ only in how you discharge that duty.
How AI is actually being used in law firms right now
Setting aside the marketing, the deployments that hold up in small and mid-size firms tend to cluster in a few places: intake triage and lead follow-up; conflict-check preparation; document review against a fixed rubric; deposition and medical-record summarization into chronologies; discovery triage and first-pass responses; billing-narrative cleanup; and drafting from the firm’s own precedent library. The pattern is the same in each: a bounded first pass on a large pile, reviewed by a human who is faster at checking than at creating.
What firms run to do it splits four ways: general-purpose assistants (Claude, ChatGPT, Copilot); AI features built into the practice management platform you already pay for (Clio Duo, Smokeball Archie); legal-specific products (CoCounsel, Harvey, Spellbook, and the litigation platforms); and custom agents wired into the firm’s own systems. We break the options and their real price ranges down in what AI law firms use and what it costs. For most small firms the honest sequence is: use what’s already in your stack, buy a point product where one clearly wins, and build custom only where your process is genuinely your own and the volume justifies it.
The math nobody shows you
Here’s the model. Every entry below is a blank you fill in from your own measurements — there are no benchmark numbers here, and you should distrust any vendor who hands you one.
| Variable | What it means | Your number |
|---|---|---|
| H | Hours per week the firm currently spends on this task | ___ |
| S | Share of H an agent can plausibly take on a first pass | ___ |
| R | Review and QA time you add back per week | ___ |
| Net | H × S − R = net hours recovered per week | ___ |
Upper bound on annual value = Net × working weeks × your effective hourly rate × realization — and only if those hours actually get refilled with billable or business-development work. Against that, put the total cost of the agent option: build or subscription, plus upkeep, plus the staff hours spent maintaining it. Then compare against the fully loaded cost of a hire, and against outsourcing converted to cost-per-unit with minimums included. Our ROI calculator walkthrough has the fuller version of this arithmetic.
Two things make this model honest. First, R is real and often larger than people expect at the start; quality review of AI output is work. Second, recovered hours are worth nothing until they’re reallocated. If a paralegal saves six hours a week and those six hours become slack, you saved zero.
An hour an agent gives back only becomes money when someone decides in advance what that hour is for.
Why the fee model matters more than the tool
A lawyer’s income is a product of four variables: effective rate × matters (or hours) delivered × realization × collection. You can push rate, push volume, or change the fee model so income isn’t capped by your own clock. Write those four out for your own practice before you write a software check — most firms discover the binding constraint is realization or collection, and no agent fixes those.
The fee model is where automation genuinely changes the business. Flat-fee and contingency work reward you for lowering cost-to-serve; hourly work punishes you for it. A firm with a reliable first-pass agent on a high-volume, fixed-fee matter type keeps the margin. A firm billing hourly for that same task has to reinvest the hours elsewhere or the savings evaporate.
The related “80/20” idea lawyers cite — that a minority of clients, matter types, and tasks drive most of the value — is the right filter for choosing what to automate first. Find the task that eats the most hours across the most matters, not the task that’s most fun to demo.
The disadvantages worth taking seriously
None of these argue against using agents. They argue for scoping them to work where a human reviewer can verify output quickly, and for a defined oversight model — we compare three of them in supervising AI agents. ABA Formal Opinion 512 is explicit that competence, confidentiality, communication, and supervisory duties all still apply when generative AI is in the workflow, and state regulators have issued their own guidance — the Florida Bar’s Ethics Opinion 24-1 and the State Bar of California’s Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law are two examples. Check your own state bar’s technology guidance, and its current version, before you scale anything.
How to decide without a six-month evaluation
-
Measure the queue for two weeks
Log where non-billable and low-value billable hours actually go, by task type. Not estimates — a tally. This is the single step firms skip and the one that determines everything downstream. -
Sort tasks into three buckets
Repeatable with a fixed output (agent candidates). Judgment, relationship, or accountability-heavy (human). High-volume but low-skill and spiky (outsourcing candidates). -
Pilot the top agent candidate with what you already own
Before buying or building, run the task in an assistant your firm already has — but check three things first: whether your inputs are used to train the provider’s models by default (and whether you can turn that off), what the data retention and deletion policy is on your tier, and whether you’re on a business or enterprise agreement with written confidentiality terms rather than a consumer plan. If the matter involves protected health information, confirm the provider will sign a business associate agreement before any PHI goes near it. If the task fails under those constraints, it will fail more expensively in a custom build. -
Write the skill down
If the pilot works, convert your instructions into a documented, versioned skill — the rubric, the output format, the exclusions, the escalation rules. This is the asset, not the tool. -
Connect it to your systems only once the output is trusted
Integrations (via MCP or a vendor API) turn a drafting aid into an agent that reads matter data and writes back. Do that after the quality question is settled, and start with read-only access. -
Re-run the math at 90 days
Compare actual recovered hours and actual review time against the hiring and outsourcing alternatives. Be willing to conclude that this task needed a person.
Where is your firm losing billable hours?
Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.
Get a free automation auditRelated guides
AI for Immigration Forms: Docketwise vs INSZoom vs Custom
A vendor-neutral comparison of immigration case management AI, general legal AI, and a custom MCP-connected agent for filling USCIS forms at a small firm.
Read guide →
AI Dictation for Lawyers: Dragon vs Word vs Custom Agent
Compare Dragon legal dictation, built-in Word/Microsoft 365 dictation, and a custom AI agent that turns dictation into filed drafts, time entries, and tasks.
Read guide →
Medical Records Retrieval: Services vs a Custom AI Agent
Compare outsourced medical records retrieval services, paralegal-run processes, and a custom AI agent — with an honest look at where agentic law firm automation fits.
Read guide →