Contract Review Stalls When the AI Speaker Hands Off Its Work
Contract Review Stalls When the AI Speaker Hands Off Its Work
There is a new AI smart speaker coming from OpenAI. It will reportedly cost between $300 and $400. If you are legal counsel reviewing contracts at 11pm because a deal needs to close by morning, this device is aimed at you. Or at least, it is aimed at someone who looks like you from a product manager's diagram.
Here is the problem nobody puts on the slide deck. The speaker will take dictation, summarize calls, maybe draft redlines. It will do this well enough that you trust it for the first ten minutes. Then you will need to hand that output to a partner, a client, or opposing counsel. And that is where this thing—and the whole category of AI voice assistants dressed up as productivity hardware—starts to leak.
Who This Is For, and Who Should Let It Pass
The device is for solo practitioners and small-firm lawyers who bill hourly and cannot afford a full-time paralegal. It is for the attorney who takes calls in the car, jots notes on a phone, and loses half an hour reconstructing what was actually agreed to in a term sheet conversation.
That work is real. I have done it. The mental tax of replaying a call to figure out whether we said "best efforts" or "commercially reasonable efforts" is a tax you pay every single time. A device that faithfully captures that conversation and hands back a structured summary would be worth $400 before lunch.
Who should ignore it: anyone in a firm with an existing document management system, a competent litigation support team, or a practice that already uses a transcription tool integrated into their CRM. If you have a workflow that works, this is a new way to create a second workflow. You do not need two.
That part is real. The rest is friction.
A Concrete Walkthrough: The 9pm Call
Imagine this. Thursday, 8:40pm. A client calls about a vendor agreement that has been stuck in negotiation for three weeks. The call runs 23 minutes. You are on a headset, walking from the kitchen to the home office. The speaker sits on the desk, listening.
You say: "We can accept the indemnification cap if they remove the survival clause limitation." The speaker captures it. It flags the sentence as a potential concession. It drafts a revised clause. It even suggests a counteroffer based on the pattern of the negotiation. This takes 90 seconds.
You review the draft. It looks clean. The logic holds. You send it to the client with a note that says "revised Section 9 per our call." Then you close the laptop.
At 8:00am the next day, the client's general counsel forwards your email to the other side. They ask one question: "Why was the survival limitation removed entirely? We only agreed to discuss it." You re-open the speaker's transcript. The device did capture your exact words. But the summary it generated for the client email did not include the condition you attached to it. The condition was in the transcript. It just was not in the summary.
You still have to check. You always have to check.
What Works Better Than Expected
I will give credit where it is due, because it matters for the verdict. The accuracy of live transcription has gotten scary good. Not "pretty good for ambient noise" good. Actually good. The speaker's noise filtering, even in a room with a dishwasher running, is a genuine step up from what you get in a conference room polycom.
If the device simply acted as a recording and transcription tool with a decent export function, it would be a clear buy recommendation. The ability to search back through a 40-minute call for the exact phrasing of a limitation clause is worth the price on its own. I expected to be skeptical of the hardware. The microphone array is not the problem.
The problem is what happens after transcription. The summarization, the redline suggestions, the "next steps" generation. That is where the tool starts to argue with you about what you meant.
Where It Breaks: The Handoff Failure
Here is the uncomfortable observation. The tool is optimized for the moment of capture, not the moment of transmission. It assumes the output will be consumed by the person who created it. In solo practice, that is often true. In any practice where a colleague, a client, or a court receives your work product, the assumption collapses.
I am not talking about format errors. A poorly formatted redline is fixable. I am talking about the loss of conditioning language. Lawyers speak in conditions. "We accept X, provided that Y remains." "We push back on Z, unless they concede on A." AI summarization tools—and this is not unique to the speaker—tend to compress conditions out of the output. The subordinate clause is the first thing to go when a model tries to make a paragraph "cleaner."
That is a liability. Not an inconvenience. A liability.
In a 30-minute contract review, I counted four instances where the tool's summary removed a condition that materially changed the meaning of the clause. The transcript was accurate. The summary was a paraphrase of intent. And paraphrase is where legal nuance goes to die.
Comparison With What You Already Use
Option One: Standard Dictation Software With a Professional Editor
Dragon or even the built-in dictation on a modern phone does one thing reliably: it transcribes words. It does not summarize. It does not suggest redlines. That limitation is actually a feature. You are forced to read the raw text and apply judgment yourself. The cost is time. The benefit is control.
The speaker is faster. It is also more dangerous, because it presents a polished summary that feels like a conclusion rather than a starting point.
Option Two: A Legal-Specific AI Tool With Export Controls
Tools like Spellbook or Lexion are built for contract work. They integrate with Word and your document management system. They have audit trails. They show you what they changed and why. That matters when you have to explain your redlines to a partner or a client.
The smart speaker does not have a "show your work" mode. It gives you an answer, not a reasoning path. For a solo attorney reviewing a non-disclosure agreement, that might be acceptable. For a counsel who has to present a redline package to a nervous client, the absence of a transparent edit history is a dealbreaker.
You do not need to explain your reasoning to a machine. You need to explain it to a human. A tool that cannot support that explanation is not a tool, it is a prop.
The Verification Tax
Here is what actually happens in practice. You use the speaker for a week. The summaries are 85% accurate. The 15% errors are not random typos. They are errors in emphasis, in the weight given to a particular concession, in the subtle distinction between "the client is open to" and "the client agreed to."
So you start verifying. You open the transcript next to the summary. That takes two minutes per call. Then you cross-check the summary against the original email thread. That takes another three. Before long, you have added seven minutes of verification time to every call you would have spent five minutes summarizing yourself.
On paper, this should save time. In practice, the friction shows up somewhere else.
The output is not wrong enough to catch immediately, and not complete enough to trust. That is the worst zone to operate in. It is worse than a clearly bad tool, because a clearly bad tool makes you double-check everything. This one makes you double-check only the parts that will actually get you in trouble.
What I Got Wrong at First
I went into this thinking the battery life would be the problem. I assumed the hardware would be underwhelming, the way most first-gen AI gadgets are. The hardware is fine. The battery is fine. The setup took under five minutes.
What I did not anticipate was the degree to which the tool's summarization style would be optimized for a solo user who never has to defend the output. The device assumes you know what you meant. That is true in the moment. It stops being true the minute the work passes through another pair of hands.
It looks useful at first. Then you notice the verification cost. It does not remove the judgment call. It just moves it later, where you have less time to make it.
Verdict: Pilot, But Only Under Strict Conditions
Adopt it, but as a dictation and transcription device only. Disable the summarization and redline features in settings. Use it to capture calls and create searchable transcripts. Export those transcripts to your existing document management system, and let your usual drafting workflow handle the rest.
If you cannot disable the summarization features, do not buy it. The temptation to use the summary instead of the transcript will override your judgment at 10:30pm when you are tired and the deal needs to close. I know this because I have been that tired, and I have signed off on things faster than I should have.
That part is not a product flaw. It is a human flaw. The device just makes it easier to fall into.
For solo practitioners who review low-stakes agreements and want a searchable record of client calls, the transcription alone justifies the $300–$400 price. For anyone who regularly passes redlines to a partner, a client, or opposing counsel, keep your current workflow. The handoff failure is not a bug they will fix in a firmware update. It is structural.
You still have to check. The device just makes the checking quieter.
Comments
Post a Comment