Claude Enterprise just shipped spending controls, and the timing isn’t a coincidence. Law firms that deployed AI agents for document review and legal research are watching their monthly bills climb faster than expected. When you’re running flat-fee immigration matters or fixed-price M&A diligence, an agent that burns through $800 in API calls on a $5,000 engagement kills your margin before you invoice.

The new controls let you set hard spending thresholds per project, restrict which models an agent can call, and receive alerts when usage crosses a percentage of budget. For law firms, this matters most in three places: document review workflows that process discovery batches, intake agents that handle after-hours calls, and research agents that pull case law and statutes. Each of those can spin up costs quickly if you don’t fence them in.

Why Law Firms Hit Budget Surprises First

Legal work has a strange cost structure. A personal-injury firm might spend 40 hours on a case that settles for a contingency fee. A corporate practice might quote a flat $15,000 for a standard contract review. In both scenarios, the firm eats any cost overrun. When an AI agent is doing the work, that overrun shows up as a line item on your Anthropic invoice.

Most firms we work with started with a single use case, usually document review or intake triage. They’d spin up an agent, test it on a handful of matters, see good results, and then scale it across the practice. Three months later, the bill is $4,000 a month and climbing. The agent works, but no one set a ceiling or tracked cost per matter.

The problem compounds when you’re using Claude’s larger models. Opus and Sonnet 3.5 are powerful, but they’re also expensive per token. If your document review agent is summarising 200-page discovery responses and you didn’t specify a cheaper model for the first pass, you’re paying top-tier rates for work that a smaller model could handle. We’ve seen firms burn $1,200 in a weekend because an agent was set to “always use the best available model” and someone uploaded a 50-file batch on Friday afternoon.

Anthropic’s new controls let you cap spending at the project level, so you can allocate $500 to a specific matter and the agent stops when it hits that limit. You can also restrict which models are available, forcing the agent to use Haiku for summaries and only escalate to Sonnet when it detects complex legal reasoning. That two-tier approach typically cuts costs by 60% without a meaningful drop in output quality.

Where Spend Controls Matter Most in Legal Practice

Three workflows generate the bulk of AI spend in law firms, and each benefits from a different control strategy.

Document review and discovery. This is where most firms start. A junior associate used to spend 12 hours doing a first-pass review of a discovery batch, flagging responsive documents and summarising key exhibits. Now a Document Review Agent does that work in 90 minutes. The agent reads every page, applies your review protocol, and produces a memo with citations.

The cost risk here is volume. Discovery batches can be 500 pages or 5,000 pages, and the firm often doesn’t know until the client uploads the files. If you’re quoting a flat fee for discovery support, you need a spending cap on the agent so it doesn’t consume your entire margin on an unexpectedly large batch. Set a threshold at 70% of your budgeted cost, and configure the agent to flag when it’s approaching that limit. You’ll have time to either request a scope change from the client or manually finish the review.

Intake and lead qualification. An Intake Voice Agent answers every call, runs a conflict check, captures the caller’s matter, and books a consultation. It’s handling after-hours, lunch breaks, and weekends when no one else is available. The average personal-injury or family-law firm loses 30 to 40% of after-hours leads because they don’t respond fast enough. An intake agent fixes that.

Spend control here is less about capping total cost and more about preventing runaway usage. If the agent is misconfigured and starts making outbound calls or sending follow-up texts without a limit, your bill can double in a week. Set a per-interaction budget and a monthly ceiling. Most firms we work with allocate $800 to $1,200 a month for intake voice, which covers 200 to 300 inbound calls depending on average handle time.

Legal research and memo generation. A research agent pulls case law, statutes, and secondary sources, then drafts a memo with citations. It’s faster than a junior associate and doesn’t miss obvious precedent. The risk is that the agent will over-research. If you ask it to “find all relevant case law on fiduciary duty in Delaware,” it might pull 40 cases when you only needed the top five. Each case costs tokens to retrieve and summarise.

The fix is to set a token budget per research task and configure the agent to prioritise the most relevant results. You can also restrict it to a smaller model for the initial search and only use a larger model for the final memo. We typically see research costs drop by half when firms move from an open-ended “research everything” prompt to a scoped task with a 10,000-token ceiling.

How Spending Thresholds Protect Flat-Fee Margins

Flat-fee and contingency work is where budget overruns hurt most. If you’re billing by the hour, an extra $300 in AI costs might just mean you bill an extra hour. If you’re working on a $7,500 flat-fee retainer, that $300 comes out of your profit.

Most small and mid-sized firms operate on a 35 to 45% profit margin after associate salaries and overhead. A $7,500 engagement might have $3,000 in margin. If your Document Review Agent burns $800 because the discovery batch was larger than expected, you just lost a quarter of your profit. Do that twice in a month and you’re underwater on the engagement.

Spending thresholds let you allocate a fixed AI budget to each matter and enforce it at the agent level. When you open a new flat-fee case, you set a $400 spending cap for all AI work on that matter. The agent tracks usage in real time and stops when it hits the limit. You get an alert at 75% and 90%, so you have time to decide whether to increase the budget or finish the work manually.

This approach works best when you tie the spending cap to your engagement letter. If you’re quoting a client $10,000 for contract review and diligence, you know your internal cost budget is around $4,000. Allocate $600 of that to AI and $3,400 to associate time. If the AI work comes in under budget, your margin improves. If it hits the cap, you’re still within your cost estimate and you haven’t blown the engagement economics.

For firms that want a practical starting point, we put together a worksheet that walks through intake workflow design and cost allocation. You can grab the AI Client Intake Checklist and use it to map your current intake process against the cost structure of an agent-based system.

What Model Restrictions Actually Do

Claude Enterprise now lets you restrict which models an agent can call. This is useful when you have a multi-step workflow and not every step needs the most expensive model.

Take document review. The first pass is usually a skim for relevance and privilege. That doesn’t require deep reasoning, it’s pattern matching and keyword spotting. You can run that step on Haiku, which costs about one-tenth as much as Opus. The second pass is where the agent reads flagged documents in detail, extracts positions, and writes a summary. That step benefits from Sonnet or Opus.

If you don’t restrict models, the agent defaults to the best available option for every task. That means you’re paying Opus rates for work that Haiku can handle. Over a month, the cost difference is $1,200 versus $4,000 for the same volume of document review.

Model restrictions also reduce the risk of accidental overspend. If an associate accidentally uploads a 300-page filing to the wrong workflow, and that workflow is set to use Opus, you’ll burn $600 before anyone notices. If the workflow is restricted to Haiku with manual approval required to escalate, the cost caps at $60 and you get a notification.

Most firms we work with use a three-tier model strategy. Haiku for intake, triage, and first-pass review. Sonnet for research, memo generation, and detailed document analysis. Opus only for complex reasoning tasks like contract negotiation strategy or multi-jurisdictional compliance questions. That structure keeps 70% of tasks on the cheapest model and reserves the expensive ones for work that justifies the cost.

The Omni Approach to Spend Governance

We build agents with spending controls baked in from the start. Every agent we deploy for a law firm has a monthly budget, a per-task cost ceiling, and a two-tier model strategy. The firm gets a dashboard that shows cost per matter, cost per agent, and a rolling 30-day burn rate.

When we run the AI audit for law firms, we map your current workflows and estimate the cost of moving each one to an agent. That estimate includes model selection, expected token usage, and a monthly spending range. You’ll know before you deploy whether an agent will cost $400 a month or $2,000 a month, and you’ll have levers to adjust that cost if it’s too high.

The audit also identifies where you’re losing money to manual work that an agent could handle for less. Most firms we work with are losing $80,000 to $250,000 a year to unbilled time, intake delays, and slow document review. An agent that costs $1,500 a month is a bargain if it’s recovering $15,000 a month in billable hours or converting 30% more intake leads.

We typically deploy three agents in the first 90 days. An Intake Voice Agent that answers every call and books consultations. A Matter Triage Agent that reviews form submissions and routes them to the right partner. A Document Review Agent that does first-pass review on contracts and discovery. Each agent has a spending cap, a model restriction, and a monthly cost target. The firm tracks ROI by comparing the cost of the agent to the revenue it generates or the cost it displaces.

If you want to see what that looks like for your practice, book a 60-min Omni Audit. You’ll walk away with a cost model, a deployment plan, and a list of workflows where an agent will pay for itself in the first month.

Practical Steps to Implement Spending Controls

If you’re already using Claude Enterprise, you can add spending controls to your existing agents today. Here’s the sequence that works for most firms.

Start with a cost audit. Pull your last three months of Anthropic invoices and break down usage by agent or workflow. Identify which agents are driving the most cost and whether that cost is justified by the output. If your document review agent is costing $3,000 a month but it’s replacing 40 hours of associate time, that’s a good trade. If your intake agent is costing $1,800 a month and you’re only getting 60 calls, something is misconfigured.

Set matter-level budgets. For every flat-fee or contingency engagement, allocate a fixed AI budget and configure your agents to respect that ceiling. Most firms start with 10 to 15% of the total engagement fee as the AI cost target. A $10,000 engagement gets a $1,000 to $1,500 AI budget. A $50,000 engagement gets $5,000 to $7,500. Track actual spend against that budget and adjust your pricing model if you’re consistently over or under.

Implement model tiers. Review every agent workflow and assign a default model. Intake and triage use Haiku. Research and memo generation use Sonnet. Complex reasoning and strategy use Opus with manual approval. Test the output quality at each tier and only escalate to a more expensive model if the cheaper one produces unacceptable results.

Monitor and adjust weekly. Spending controls are not set-and-forget. Review your usage dashboard every week for the first month, then every two weeks after that. Look for agents that are hitting their caps early, workflows that are costing more than expected, and tasks that could move to a cheaper model. Most firms find they can cut their AI spend by 40% in the first 60 days just by tuning model selection and tightening task prompts.

If you’re not on Claude Enterprise yet, or you’re using a different platform, the same principles apply. Every agent platform now offers some form of spending control, whether it’s a hard cap, a model restriction, or a usage alert. The key is to configure those controls before you scale the agent, not after you get a surprise bill.

Why This Matters Now

The legal AI market is moving faster than most firms realise. Two years ago, document review agents were experimental. Today, they’re table stakes. Firms that don’t have an intake agent are losing 40% of their after-hours leads to competitors who do. Firms that don’t have a research agent are paying associates $250 an hour to do work that an agent can do for $8.

But the firms that deploy agents without cost controls are finding that the economics don’t work. An agent that costs $4,000 a month needs to generate or save at least $12,000 a month to justify the expense. If you’re not tracking cost per matter, cost per task, and ROI by agent, you won’t know whether you’re getting that return.

Anthropic’s new spending controls make it easier to run agents profitably. You can cap costs at the project level, restrict expensive models to high-value tasks, and get alerts before an agent blows through your budget. For law firms working on flat-fee or contingency matters, that’s the difference between an agent that improves your margin and one that erodes it.

We’ve built Omni to handle this complexity for you. Every agent we deploy has spending controls, model restrictions, and a cost dashboard. You’ll know what each agent costs, what it’s saving or generating, and whether it’s paying for itself. The audit is 60 minutes, and you’ll walk away with a cost model and a deployment plan that ties agent spend to business outcomes.

If you’re ready to see what that looks like for your firm, book my Omni Audit and we’ll map your workflows, estimate your costs, and show you where an agent will pay for itself in the first 30 days. You can also explore more about how we approach AI strategy and implementation or see what other firms are building in our insights library.

The firms that figure out cost governance now will have a two-year head start on the ones that wait. Spending controls aren’t a nice-to-have feature, they’re the foundation of a profitable AI practice. Get them in place before your next flat-fee engagement, and you’ll protect your margin while your competitors are still trying to explain why their AI bill is higher than their cloud hosting.