OpenAI just admitted what every law firm experimenting with ChatGPT Enterprise already knows: without hard spending limits, one associate running a discovery batch through the API can rack up a $12,000 bill before lunch.

The company rolled out new cost controls this month that let enterprise customers set per-user spending caps, usage alerts, and project-level budgets. It’s a direct response to the problem firms face when they hand AI tools to a team of associates who don’t think twice about feeding 80,000 pages of deposition transcripts into a model that charges by the token.

I’ve watched three mid-sized litigation practices hit surprise invoices north of $18,000 in a single billing cycle. In each case, the culprit was the same: a junior associate treating the API like a search bar, running the same query fifteen different ways, and uploading entire matter files without chunking or filtering. The associate wasn’t reckless. They just didn’t know the cost structure, and the firm didn’t have guardrails in place.

OpenAI’s new controls are a step forward, but they don’t solve the bigger problem. Most law firms don’t need better cost monitoring for ad-hoc ChatGPT use. They need purpose-built agents that do specific work at a predictable cost, with no risk of runaway spend.

Why Law Firms Hit Surprise AI Bills

The typical pattern looks like this. A partner approves a ChatGPT Enterprise seat for the litigation team. An associate starts using it to summarise depositions, draft discovery responses, and review contracts. The first month’s bill is $400. The second month it’s $1,800. By month three, it’s $6,200, and the managing partner is asking what happened.

What happened is volume. The associate found a tool that works, so they used it for everything. They uploaded full PDFs instead of excerpts. They re-ran queries with minor tweaks. They didn’t realise that a 200-page contract costs $4 to process, and they were processing forty contracts a day.

The new OpenAI controls let you cap that associate at $500 per month. When they hit the limit, the account pauses until you approve more spend or the cycle resets. That’s useful if you’re running ad-hoc queries, but it doesn’t help you scale the work that actually drives revenue.

The real issue isn’t cost monitoring. It’s that most firms are using general-purpose AI for tasks that need a specialist. You wouldn’t hire a generalist associate to do nothing but intake calls, and you shouldn’t use a general-purpose LLM to do work that needs context, routing logic, and integration with your practice management system.

What a Purpose-Built Agent Looks Like

A law firm doing $8 million in annual revenue typically loses 30 to 40 per cent of after-hours intake. A potential client calls at 6:45 PM, gets voicemail, and calls the next firm on their list. By the time your intake coordinator returns the call the next morning, they’ve already signed a retainer elsewhere.

An Intake Voice Agent answers every call. It conflict-checks the caller’s name and opposing party against your database, captures the matter details, and books a consultation directly into the partner’s calendar. The entire interaction costs $1.20 in API spend. The agent runs 24/7, handles twenty calls a night, and never misses a high-intent lead.

Compare that to an associate using ChatGPT to draft an intake summary after the fact. They’re spending $3 to $5 per summary, and they’re doing it after the client is already gone. The work happens, but it doesn’t capture the revenue.

The same logic applies to document review. A Document Review Agent performs first-pass review on discovery batches, contracts, and matter files. It flags problematic clauses, summarises positions, and produces an associate-grade memo. The cost is fixed per document, typically $2 to $6 depending on length and complexity. An associate doing the same work bills at $200 to $400 per hour, and the client often pushes back on the line item.

The agent doesn’t replace the associate. It handles the first pass, and the associate reviews the output, adds judgment, and signs off. The total time drops from six hours to ninety minutes, and the client sees a cleaner invoice.

The Three Places Firms Lose Money to Manual Work

Most law firms leak revenue in three places: unbilled time, intake delays, and document review overhead. The dollar impact varies by practice area, but the pattern is consistent across firms doing $1 million to $25 million in annual revenue.

Billable-hour leakage is the silent killer. Every attorney in the firm spends four to six hours per week on work that never makes it onto an invoice. Intake calls, matter admin, internal coordination, and first-pass document review. It’s necessary work, but it’s not billable, and it adds up to $80,000 to $250,000 per year in lost capacity for a typical firm.

Intake delays cost you clients you never knew you had. A family law practice in our network tracked after-hours calls for three months. They received 140 inbound calls outside business hours. Sixty-three callers left voicemail. Forty-one never called back. That’s $160,000 in retainer revenue that walked out the door because no one picked up the phone.

Document review and discovery work is expensive and hard to scale. Junior associates spend days on first-pass review, billing $200 to $400 per hour. Clients balk at the cost, and partners write off chunks of the invoice to keep the relationship intact. The work still has to get done, but the firm eats the margin.

If you’re running a practice with ten attorneys and you’re losing five hours per week per person to unbilled admin, that’s 2,600 hours per year. At a blended rate of $300 per hour, that’s $780,000 in capacity you can’t invoice. Even if you recover half of that through better process and delegation, you’re still leaving $390,000 on the table.

How We Build Agents That Don’t Blow the Budget

The difference between a general-purpose LLM and a purpose-built agent is scope. A general-purpose tool can do anything, so it has no guardrails. A purpose-built agent does one thing, and it does it the same way every time.

When we build an agent for a law firm, we start with the Omni Audit. It’s a 60-minute working session where we map the manual work, identify the highest-value automation, and spec the agent logic. You walk away with three outputs: a process map, a cost-benefit model, and a 14-day build plan. No deck, no discovery phase, no six-week scoping exercise.

The agents we build most often for law firms are Intake Voice Agent, Matter Triage Agent, and Document Review Agent. Each one targets a specific workflow, integrates with your existing systems, and runs at a fixed cost per interaction.

The Intake Voice Agent handles inbound calls. It answers in under two rings, conflict-checks the caller, captures matter details, and books a consultation. The entire interaction costs $1.20. If you’re taking 200 intake calls per month, your total cost is $240. Compare that to a full-time intake coordinator at $4,500 per month, and the math is straightforward.

The Matter Triage Agent reviews incoming form submissions and emails. It classifies the practice area, scores the fit, and routes the inquiry to the right partner with a one-paragraph brief attached. The cost is $0.80 per submission. A firm handling 300 inquiries per month spends $240 on triage, and the partner sees every lead with context already attached.

The Document Review Agent performs first-pass review on contracts, discovery batches, and matter files. It flags clauses, summarises positions, and produces a memo. The cost is $2 to $6 per document, depending on length. An associate doing the same work bills $200 to $400 per hour, and the client sees a line item that’s hard to justify. The agent does the grunt work, the associate adds judgment, and the invoice is cleaner.

We don’t build agents that do everything. We build agents that do one thing well, at a cost you can predict, with no risk of runaway spend. The new OpenAI controls are useful for firms experimenting with ad-hoc use, but they don’t replace the need for purpose-built tools that integrate with your practice management system and run the same way every time.

What Predictable AI Spend Actually Looks Like

The firms that get the most value out of AI are the ones that treat it like any other operational cost. They don’t experiment with general-purpose tools and hope the bill stays reasonable. They deploy agents with fixed costs, clear scope, and integration into the systems they already use.

A litigation practice with twelve attorneys deployed an Intake Voice Agent and a Matter Triage Agent in the same month. The intake agent handled 180 calls in the first 30 days. The triage agent processed 240 form submissions. The total AI cost for the month was $408. The firm captured eleven new retainers that would have gone to voicemail, worth $87,000 in total contract value.

The managing partner didn’t get a surprise invoice. They got a line item that said $408, and they knew exactly what work it covered. That’s what predictable spend looks like.

The same principle applies to document review. A family law practice used a Document Review Agent to perform first-pass review on 60 custody and financial disclosure filings over a two-week period. The total cost was $340. An associate doing the same work would have billed 18 hours at $250 per hour, or $4,500. The associate still reviewed the output and signed off, but the total time dropped to four hours, and the client saw a $1,000 line item instead of a $4,500 one.

The agent didn’t replace the associate. It did the work the associate shouldn’t be doing in the first place, and it did it at a cost the client could understand.

If you’re still using ChatGPT Enterprise for ad-hoc queries, the new cost controls will help you avoid surprise bills. But if you want to scale the work that actually drives revenue, you need agents that do specific tasks at a fixed cost, with no risk of runaway spend. You can explore more about how we build these systems at Omni or dive into the voice and ops capabilities at Omni Voice and Omni Ops.

The Intake Problem Every Firm Has

Most law firms think they have an intake process. What they actually have is a phone number, a voicemail box, and a coordinator who checks messages twice a day. That’s not a process. That’s a gap where revenue disappears.

The typical pattern: a potential client calls at 7 PM. They get voicemail. They leave a message. Your intake coordinator calls back at 9 AM the next morning. The client doesn’t answer. They’ve already called three other firms, and one of them picked up.

You never see that client in your CRM. You never count them as a lost lead. They just vanish, and you assume the marketing wasn’t working.

An Intake Voice Agent solves this by answering every call. It doesn’t matter if it’s 7 PM on a Friday or 10 AM on a Tuesday. The agent picks up, conflict-checks the caller, captures the matter, and books a consultation. The entire interaction takes three minutes, costs $1.20, and the client hangs up with a calendar invite in their inbox.

One family law practice in our network ran the numbers. They were losing 40 per cent of after-hours intake. Over a quarter, that was 52 potential clients who called, got voicemail, and never came back. At an average retainer of $4,500, that’s $234,000 in revenue that walked out the door because no one picked up the phone.

They deployed an Intake Voice Agent in January. By March, they’d captured 31 after-hours leads that would have gone to voicemail. Eighteen of those turned into signed retainers, worth $81,000 in total contract value. The AI cost for the quarter was $720.

If you’re serious about fixing intake, we built a worksheet that walks through the process. You can grab the AI Client Intake Checklist for Law Firms and use it to map your current gaps. It’s a practical tool, not a pitch deck.

Why Cost Controls Don’t Fix the Real Problem

OpenAI’s new spending limits are useful if you’re running ad-hoc queries and you want to avoid surprise bills. But they don’t help you scale the work that actually drives revenue.

The problem isn’t that your associates are spending too much on ChatGPT. The problem is that they’re using a general-purpose tool for work that needs a specialist. You wouldn’t hire a generalist associate to do nothing but answer intake calls, and you shouldn’t use a general-purpose LLM to do work that needs context, routing logic, and integration with your practice management system.

The firms that get value out of AI are the ones that deploy purpose-built agents with fixed costs, clear scope, and predictable outcomes. They don’t experiment with tools and hope the bill stays reasonable. They build systems that do specific work at a cost they can forecast, and they measure the ROI in dollars, not tokens.

If this is the kind of problem agents can help with, the free Working With Claude field guide is the practical next step. Thirty-two pages, no fluff. Get the free guide.

The new OpenAI controls are a step forward, but they’re not a strategy. If you’re serious about scaling your practice without scaling your headcount, you need agents that do real work at a predictable cost. That’s what we build, and that’s what the firms winning with AI have in common.

You can explore more case studies and practical guides in our insights library or learn more about the broader AI implementation process in our guides section.