Always-On AI Agents vs. On-Demand AI in Law Firms

By Jude Lee · · Comparison

Lawyers and a paralegal reviewing an automated deadline dashboard on a laptop in a small firm conference room

What changed: agents that wait for events, not prompts

Two forces are driving the 2026 conversation. The first is capital: legal AI platforms have raised very large late-stage rounds at multibillion-dollar valuations. The second is packaging: a growing number of vendors now market “always-on” agent hubs aimed at legal teams that want software monitoring inboxes, matter folders, and contract repositories continuously rather than answering one question at a time.

A note on those funding numbers, because it matters for a piece that asks you to verify things. We are not publishing a specific round size or valuation here, because we cannot verify one to a dated primary report as of this writing. If you want to quote a figure in a partner meeting, pull it from the original reporting — Reuters, the Financial Times, and the vendor’s own announcement — and say the date out loud. Funding figures get garbled fast in secondhand summaries, and a stale number cited confidently is exactly the failure mode firms are trying to avoid with AI in the first place.

Strip the marketing and the underlying capability is simple: an agent is a model plus tools plus a trigger. The trigger can be you typing a prompt, or it can be a schedule (every morning at 7), a webhook (Clio fires an event when a matter status changes), or a watcher (a new PDF lands in a NetDocuments workspace). “Always-on” just means the trigger isn’t a human.

An always-on agent isn’t smarter than the one you prompt manually. It’s the same model with a different alarm clock — which means it inherits every one of your governance problems and runs them unattended.

On-demand versus always-on, honestly compared

On-demand AI (you prompt it)
A lawyer or paralegal is in the loop by construction. Output is reviewed the moment it’s produced. Cost is proportional to use. Failure is visible immediately — you see the bad answer. Best for: research, drafting, analysis, anything judgment-heavy or one-off.
Always-on agent (an event prompts it)
Runs whether or not anyone is watching. Catches things nobody had time to look for (a deadline in a scanned order, an unread client email) — though only the things its instructions describe. Failure is silent: a broken connector can mean nothing happens for a week and no one notices. Best for: monitoring, extraction, queue-building, first drafts staged for review.

There is a third option people skip: deterministic automation. If the task is “when a matter closes, send the closing letter template and archive the folder,” you do not need a language model at all. A rule in your practice management system, or a no-code workflow builder, is cheaper, auditable, and won’t hallucinate. Reach for an agent only when the input is unstructured — free text, PDFs, scans, voicemail transcripts — or when the next step genuinely depends on reading and judging content.

Where a background agent actually earns its keep

In our view the honest shortlist for a 5–40 lawyer firm is short:

What does not belong on an unattended trigger, in our opinion: e-filing, sending anything to a client or opposing counsel, trust account movement, and any substantive legal conclusion. The ABA’s Formal Opinion 512 on generative AI tools (July 2024) discusses lawyers’ duties of competence, confidentiality, communication, and supervision when using these tools, and ABA Model Rule 5.3 addresses responsibility for nonlawyer assistance — worth reading alongside your own state bar’s guidance, which varies and should be checked directly.

How firms are actually using AI right now

Across small and mid-size firms, the deployed uses cluster into four buckets: drafting (letters, discovery responses, routine agreements), summarizing (depositions, records, long email chains), searching (finding the right precedent or the right page in a 900-page production), and routing (intake, email triage, task assignment). Research and citation verification remain the areas where firms report the most caution, for good reason — fabricated citations have produced sanctions in multiple U.S. courts, and any research output needs checking against the primary source.

The uses that stall are usually the ones sold as “end-to-end.” Our earlier piece on why AI automation stalls after the first win covers the pattern: one task works, nobody owns the second, and the pilot quietly dies.

Concentrating effort on the workflows that carry the load

The old Pareto heuristic — a small share of inputs drives most of the output — is a useful lens, not a measurement. Applied to automation: list every recurring task in your firm, then automate only the handful that are high-frequency, low-judgment, and unstructured. Everything else is a distraction dressed as a roadmap.

Here’s the economic model to run yourself. Don’t use anyone’s published benchmark — use your own numbers. Assume 48 working weeks a year (52 minus vacation and holidays; adjust to your firm’s actual figure):

That last line is the one vendors omit. An agent that produces output nobody trusts adds a review step rather than removing work. There’s a second cost the demos skip too: a background agent fires on every trigger whether or not anyone reads the result, so your token and licence spend scales with event volume, not with value delivered. A watcher pointed at a busy inbox can quietly cost more in a month than the workflow it replaced.

And on the question that keeps appearing in searches — whether lawyers make $500,000 a year — the honest answer is that individual attorney income varies enormously by practice area, geography, and firm structure; the U.S. Bureau of Labor Statistics publishes occupational wage data for lawyers if you want a defensible reference point. The relevant operational question isn’t the headline number. It’s leverage: how many billable or matter-advancing hours each timekeeper can produce without adding headcount, and whether AI moves that number for your practice area.

Choosing a tool without pretending there’s one best answer

There is no single best AI program for law firms, and any list that claims otherwise is ranking by marketing spend. Practically, you’re choosing among three layers, and most firms end up with more than one. A useful discriminator beyond friction and flexibility: where client data is processed and how long it’s retained — ask each vendor in writing, and confirm the answer against their published documentation.

  1. AI inside the software you already run — Clio Duo, Smokeball Archie, iManage and NetDocuments features. Lowest friction, narrowest scope. Data usually stays inside a system you’ve already vetted and papered, which is the strongest argument for starting here.
  2. Purpose-built legal AI platforms — Harvey, CoCounsel, and practice-specific tools. Strong on legal-specific tasks, priced accordingly, and you work the way the vendor designed. You are adding a new processor to your client-confidentiality map, so the retention and subprocessor terms matter more than the feature list.
  3. A general assistant connected to your systems — Claude or similar, wired to your practice management and document system via MCP, with reusable skills that standardize how a task is done. Most flexible, most configuration work. You control what the agent can reach and what leaves your systems, which is an advantage only if someone actually owns that scoping.

Building a triggered agent without losing control

  1. Pick one trigger and one output

    Start with something like: “when a document is added to a matter in Clio, produce a one-page summary and attach it as a note.” One input, one output, no actions taken outside the firm.

  2. Give it governed access, not credentials in a script

    Connect through MCP so the agent reads only the matters and fields you scope. Our Claude-to-Clio walkthrough covers the mechanics and the permission boundaries.

  3. Write the skill before you turn on the trigger

    Package the instructions — format, required fields, what to do when data is missing, when to escalate — so every run is identical. Run it manually a few dozen times first.

  4. Add a heartbeat and a failure alarm

    Silent failure is the defining risk of unattended agents. Require a daily “ran successfully, processed N items” message to a human. No message is itself the alert.

  5. Decide who reviews, and log it

    Choose an oversight posture deliberately — we compare three in supervising AI agents. Whatever you pick, keep an auditable record of what the agent did and who approved it.

As of 2026, the capability is genuinely new and genuinely useful in a narrow band. Put agents on watching and extracting. Keep humans on deciding and sending. And be willing to conclude that for a given workflow, a rule and a template beat both.

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 audit

Related guides