Amazon’s director of artificial general intelligence research said something last week that every law firm partner needs to hear. The problem with enterprise AI isn’t capability anymore. It’s reliability.

Most systems can handle 95% of tasks correctly. That sounds good until you realise the other 5% includes telling a prospective client you don’t handle their practice area when you do, or missing a conflict check that torpedoes a six-figure engagement three months in.

For law firms deploying AI in client intake or document review, the math is brutal. You need 99%+ accuracy before you can remove human oversight. We’re not there yet, and most firms won’t be for another 12 to 18 months.

That doesn’t mean you wait. It means you deploy AI as an assistant, not a replacement. The firms that figure this out now will capture the 30 to 40% of after-hours intake that currently walks to competitors. The firms that wait will keep bleeding $80K to $250K annually in missed opportunities and unbilled hours.

The reliability gap in legal AI

Here’s what reliability means in practice. A voice AI that answers your intake line can transcribe perfectly, route calls to the right person, and even book consultations. But if it misclassifies one in twenty callers, you’ve just told a personal injury lead you only do estate planning, or worse, you’ve onboarded a client with a conflict you didn’t catch.

The cost of that error isn’t just the lost engagement. It’s the malpractice exposure, the ethics complaint, the time your partners spend cleaning up the mess. One mistake can cost more than a year of saved intake hours.

This is why most legal AI deployments right now should keep a human in the loop. The AI does the heavy lifting, but a paralegal or intake coordinator reviews the output before it goes live. You get 80% of the efficiency gain with 5% of the risk.

Document review has the same problem. An AI can flag clauses, summarise positions, and produce an associate-grade memo in minutes. But if it misses a liability cap or misreads an indemnity provision, you’ve just advised a client based on incomplete analysis. The partner who signs off on that work is the one who owns the error.

The firms we work with through the AI audit for law firms typically start with two questions. First, where are we losing the most money to manual work right now? Second, where can we deploy AI without creating unacceptable risk?

The answer to the first question is almost always client intake and document review. The answer to the second is almost always “with human oversight for at least the next year.”

Where law firms leak money every week

Let’s walk through the actual dollar impact. A three-partner firm doing $2.5M annually typically loses 4 to 6 unbilled hours per attorney per week. That’s time spent on intake calls, document review, and matter admin that never makes it onto a billable invoice.

At $400 per hour, that’s $1,600 to $2,400 per attorney per week. Across three partners, you’re looking at $250K to $375K annually in time that could be billed but isn’t.

Most of that time falls into three buckets. Intake delays are the biggest. A high-intent call comes in at 6pm or Saturday morning. No one answers. The caller tries two more firms and books with whoever picks up first. You never even knew they called.

The second bucket is document review. A junior associate spends three days on first-pass review of a discovery batch. The work is necessary, but it’s slow, expensive, and hard to scale. At $200 to $400 per hour of associate time, a single matter can burn $10K to $15K before the partner even sees the file.

The third bucket is matter triage. Intake forms and emails sit in a shared inbox. Someone has to read each one, figure out the practice area, check for conflicts, and route it to the right partner. That’s 20 to 30 minutes per submission, and most firms get 15 to 25 submissions per week.

Add it up and you’re looking at $80K to $250K in leakage for a typical firm in this revenue band. The firms at the higher end of that range are usually the ones with the most after-hours intake volume and the least structured triage process.

What assistant-mode AI looks like in practice

Here’s how we deploy AI for intake right now. The Intake Voice Agent answers every call. After-hours, lunch, weekends, doesn’t matter. It captures the caller’s name, matter type, and availability. It runs a basic conflict check against your client list. It books a consultation directly into the partner’s calendar.

Then it sends a summary to your intake coordinator. Name, matter, conflict status, appointment time. The coordinator reviews it in two minutes, confirms the conflict check, and either approves the booking or flags it for follow-up.

The AI did 90% of the work. The human did the 10% that carries the risk. You’ve just captured an after-hours lead that would have gone to a competitor, and you’ve done it without exposing the firm to a missed conflict or a misclassified matter.

The Matter Triage Agent works the same way. It reviews incoming form submissions and emails, classifies the practice area, scores fit based on your intake criteria, and routes it to the right partner with a one-paragraph brief attached. The partner sees it in their inbox with enough context to decide in 30 seconds whether to take the call or pass.

Again, the AI does the grunt work. The partner makes the judgment call. You’ve cut 20 minutes of admin down to 30 seconds, and you haven’t introduced any new risk.

Document review is where the reliability gap shows up most clearly. The Document Review Agent can perform first-pass review on contracts, discovery batches, and matter files. It flags clauses, summarises positions, and produces a memo that looks like it came from a second-year associate.

But right now, that memo goes to a senior associate or partner for review before it gets filed or sent to the client. The partner isn’t re-reading every document. They’re spot-checking the AI’s work, looking for the 5% of cases where it missed something or misread a provision.

That’s still a massive time save. A three-day review becomes a half-day review. But you’re not removing the human from the loop yet, because the cost of a mistake is too high.

If you want a structured way to evaluate where AI can help your intake process without introducing unacceptable risk, we’ve built a checklist that walks through the decision points. You can grab the AI Client Intake Checklist for Law Firms and use it as a worksheet for your next partner meeting.

The 12 to 18 month window

Here’s the timeline we’re giving clients right now. For the next 12 to 18 months, deploy AI as an assistant. Use it to handle the repetitive, low-risk parts of intake and document review. Keep a human in the loop for the final check.

During that window, two things will happen. First, the reliability of the underlying models will improve. We’re already seeing accuracy gains every quarter. The gap between 95% and 99% is closing faster than most people expected.

Second, your firm will build a dataset of reviewed AI outputs. Every time your intake coordinator approves a booking or your partner signs off on a document memo, you’re creating a training example. After six months, you’ll have enough data to fine-tune the AI specifically for your firm’s practice areas, conflict rules, and intake criteria.

That’s when you can start removing the human from the loop for low-risk tasks. An after-hours intake call for a straightforward estate planning matter? The AI can probably handle that end-to-end by mid-2027. A discovery review for a case with no unusual liability issues? Same thing.

But high-stakes intake and complex document review will still need human oversight for the foreseeable future. The reliability threshold for those tasks is higher, and the cost of a mistake is too steep to automate fully.

The firms that win in this window are the ones that deploy now in assistant mode and use the next 12 months to build the dataset and refine the process. The firms that wait until the technology is “ready” will be 18 months behind on the learning curve.

What an Omni Audit tells you

When we sit down with a law firm for an Omni Audit, we’re looking at three things. First, where are you losing money to manual work right now? We map the actual hours and dollar impact across intake, document review, and matter admin.

Second, where can you deploy AI without creating unacceptable risk? That’s a function of your practice areas, your client mix, and your risk tolerance. A personal injury firm has different exposure than a transactional practice.

Third, what does the roadmap look like? Which tasks can you automate in assistant mode today, which ones need another six months of model improvement, and which ones will need human oversight indefinitely?

The output is three things. A process map that shows where the leakage is happening. A prioritised list of AI deployments ranked by dollar impact and risk. And a 90-day implementation plan that gets the first agent live and starts building your dataset.

It takes 60 minutes. No deck, no follow-up meeting, no multi-week discovery process. You walk out with a plan you can hand to your practice manager and start executing.

Most firms we audit are losing $80K to $250K annually to intake delays and unbilled document review time. The ones that move fast capture 60 to 70% of that back in the first year, even with human oversight still in place.

Book a 60-min Omni Audit and we’ll map the leakage specific to your firm.

Why reliability matters more than capability

The reason Amazon’s AGI director’s comment matters is that it reframes the deployment question. Most law firm partners are asking “Can AI do this task?” The better question is “Can AI do this task reliably enough that I’m comfortable removing human oversight?”

For client intake, the answer right now is no, not for another 12 to 18 months. But AI can absolutely do 90% of the work reliably enough that you should deploy it today in assistant mode.

The firms that get this distinction right will capture the after-hours intake volume that’s currently walking to competitors. They’ll cut document review time by 60 to 70% without introducing new malpractice risk. And they’ll build the dataset and process refinement they need to move to full automation when the reliability threshold crosses 99%.

The firms that wait for perfect reliability will spend another 18 months losing $80K to $250K annually while their competitors build the lead.

If you want to see what this looks like for your specific practice areas and client mix, the fastest way is to book your Omni Audit. We’ll map the leakage, prioritise the deployments, and give you a 90-day plan.

You can also explore more about how we’re building Omni for law firms and the specific agents we deploy for intake, triage, and document review. The technology is ready. The question is whether you’re going to deploy it in assistant mode now or wait until your competitors have an 18-month head start.