Anthropic suspended access to its advanced Claude models in January 2025 for users in countries subject to US export controls. No warning. No migration window. Firms relying on those models for contract review, discovery triage, or intake workflows woke up to dead endpoints and half-finished work.

The suspension lasted days, not hours. For a law firm running client intake through a third-party AI platform built on Claude, that meant every after-hours call went to voicemail. Every form submission sat in a queue. Every conflict check stalled. The firms that lost the most revenue weren’t the ones using AI poorly, they were the ones using it well and trusting a vendor whose infrastructure sat outside their control.

This isn’t a story about Anthropic. It’s a story about what happens when you move sensitive, time-critical work to a platform you don’t control and the vendor’s compliance obligations override your operational needs. For law firms, the lesson is blunt: if you’re automating intake, document review, or matter triage with AI, you need to know where the model runs, who controls access, and what happens when geopolitics or export rules change overnight.

The Real Cost of a Service Interruption

Most firms think about AI vendor risk in terms of data security or accuracy. Those matter, but the sharper edge is availability. When your intake agent goes offline at 6 PM on a Friday and doesn’t come back until Monday, you’re not just losing leads. You’re losing the high-intent calls that were ready to retain you tonight.

We see this pattern across firms doing $2M to $15M annually. After-hours intake represents 30 to 40 percent of total inbound volume. Conversion rates for those calls are lower than business-hours leads, but the ones that do convert tend to be higher-value matters. Someone calling a family law firm at 9 PM about custody isn’t comparison shopping. They want help now.

When that call hits voicemail because your AI vendor’s model access got suspended, the caller moves to the next firm on the search results page. You don’t get a second chance. The cost isn’t just the lost retainer, it’s the cumulative leakage across dozens of calls over a weekend. For a mid-sized firm, that’s $12K to $30K in forgone revenue every time your intake system goes dark for 48 hours.

The Anthropic suspension was short. The next one might not be. Export controls are tightening, not loosening. The US government added new restrictions on AI model exports in late 2024. Other jurisdictions are following. If your vendor relies on a model developed in a country subject to those controls, or serves clients in restricted regions, you’re one regulatory update away from another outage.

Where the Model Runs Matters More Than What the Model Does

Most firms evaluate AI vendors by asking what the model can do. Can it summarise a deposition transcript? Can it flag non-standard clauses in a commercial lease? Can it route a new matter to the right associate? Those are the wrong first questions.

The right first question is: where does the model run, and who controls access to it?

There are three deployment models in enterprise AI. Cloud-hosted, where the vendor runs the model on their infrastructure and you send data over the wire. Hybrid, where some processing happens locally but the model itself still lives in the vendor’s environment. And on-premise, where the model runs inside your boundary, on hardware you control or a private cloud instance that doesn’t share infrastructure with other tenants.

For law firms handling privileged communications, trade secrets, or personal injury files with medical records, on-premise deployment isn’t a nice-to-have. It’s the only architecture that keeps you in control when a vendor’s access to a third-party model gets cut off.

VIDIZMO’s response to the Anthropic suspension illustrates the difference. Because their enterprise video AI platform supports on-premise deployment, customers running the software inside their own data centres weren’t affected. The model was already local. No external API call. No dependency on a vendor’s ability to access a foreign-origin model. When Anthropic’s access got suspended, VIDIZMO’s on-premise customers didn’t notice.

Firms using cloud-hosted platforms built on Claude did notice. Their workflows stopped. Some vendors scrambled to swap in a different model, but that introduced new accuracy problems and required re-training agents that had been tuned for Claude’s behaviour. The firms that came out clean were the ones that had either deployed on-premise or had contract terms guaranteeing failover to a secondary model with no service interruption.

What an On-Premise Intake Agent Looks Like

Let’s make this concrete. A five-attorney family law firm in the Midwest handles 60 to 80 inbound calls a week. Thirty of those come in after 5 PM or on weekends. Before automation, the firm used an answering service that took a message and emailed it to the intake coordinator. The coordinator followed up the next business day. Conversion rate on those after-hours leads was under 15 percent.

The firm deployed an Intake Voice Agent through the AI audit for law firms we ran in late 2024. The agent answers every call, asks the caller about their matter, runs a conflict check against the firm’s case management system, and books a consultation directly into the partner’s calendar. The entire interaction takes three to four minutes. The agent sends a confirmation text with the appointment details and a link to upload documents before the consultation.

That agent runs on-premise. The firm’s IT provider set up a small server in their office. The voice model, the conflict-check logic, and the calendar integration all run locally. When the caller’s audio hits the system, it never leaves the building. No API call to a third-party vendor. No dependency on a model hosted in another country.

The firm’s conversion rate on after-hours intake jumped to 48 percent. The partner books four to six consultations a week that wouldn’t have happened under the old answering-service model. At an average retainer of $4,500 for a contested custody matter, that’s an extra $90K to $135K in annual revenue from leads that used to go to voicemail.

When Anthropic suspended Claude access in January, this firm didn’t feel it. Their agent kept running. The model was local. The only thing that could take it offline was a power outage or a hardware failure, and both of those are risks the firm already managed for their case management system and email server.

Compare that to a firm using a cloud-hosted intake platform. When the vendor’s access to Claude got cut off, the platform switched to a fallback model that wasn’t trained on legal intake workflows. The agent started asking irrelevant questions, missed conflict flags, and double-booked consultations because the calendar integration broke during the model swap. The firm turned off the agent and went back to the answering service for a week while the vendor sorted it out. They lost 11 inbound calls that weekend. Three of those calls were high-value estate planning matters. Total forgone revenue: roughly $38K.

Document Review and Discovery Don’t Tolerate Downtime

Intake is the most visible place where an AI outage costs you money, but it’s not the only one. Document review and discovery workflows are just as time-sensitive, and the cost of a suspension there shows up differently.

A litigation boutique with eight attorneys handles commercial disputes and employment cases. They run 40 to 60 active matters at any time. Discovery in a typical case involves 2,000 to 8,000 documents. The firm used to assign first-pass review to junior associates, who billed the work at $250 an hour. A first pass on 5,000 documents took 18 to 24 hours of associate time. That’s $4,500 to $6,000 in billed hours, but clients pushed back on the cost and the firm often wrote down 20 to 30 percent of the review time to keep the relationship intact.

The firm deployed a Document Review Agent that performs first-pass review, flags responsive documents, summarises key clauses, and produces a memo ranking documents by relevance. The agent runs on a private cloud instance inside the firm’s AWS environment. The model is hosted in the firm’s own virtual private cloud. No shared infrastructure. No dependency on a vendor’s ability to access a third-party API.

The agent cut first-pass review time from 20 hours to four hours of associate time. The associate reviews the agent’s output, spot-checks flagged documents, and finalises the memo. The firm bills the work at the same rate, but the client sees a faster turnaround and the associate spends the saved 16 hours on higher-value work like drafting motions or prepping witnesses.

When the Anthropic suspension hit, this firm kept running. Their agent didn’t call an external API. The model was already inside their boundary. A competitor firm using a cloud-hosted document review platform lost access for three days. They had a summary judgment motion due that Friday. The associate had to do the entire first pass manually. The firm missed the filing deadline, requested an extension, and billed the client for 28 hours of emergency review work. The client refused to pay the full amount and the firm wrote off $4,200.

The difference wasn’t the quality of the AI. Both firms were using capable models. The difference was where the model ran and who controlled access to it.

What to Ask Your Vendor Before You Sign

If you’re evaluating an AI platform for intake, document review, or matter triage, these are the questions that matter:

Where does the model run? Cloud-hosted, hybrid, or on-premise? If it’s cloud-hosted, does the vendor offer an on-premise option for an incremental cost?

What model does the platform use, and where was it developed? If it’s a foreign-origin model, what happens if export controls restrict the vendor’s access?

Does the contract include a service-level agreement with penalties for downtime? What’s the guaranteed uptime percentage, and what counts as an outage?

If the primary model becomes unavailable, does the platform fail over to a secondary model automatically? How long does failover take, and will the secondary model produce comparable results?

Can you run the platform in your own cloud environment or on your own hardware? If yes, what’s the implementation timeline and what’s the incremental cost?

Most vendors will tell you their platform is secure and reliable. Push for specifics. Ask to see the SLA. Ask what happened to their customers during the January Anthropic suspension. Ask whether they’ve stress-tested failover and whether they can demonstrate it working in real time.

The vendors who can’t answer those questions cleanly are the ones who’ll leave you in the dark the next time a model provider pulls access. The ones who can answer them are the ones who’ve thought through the operational risk and built redundancy into their architecture.

The Omni Approach: Build It Inside Your Boundary

We build agents that run where you need them to run. For most law firms, that means on-premise or in a private cloud instance that doesn’t share infrastructure with other customers. The model, the logic, and the data stay inside your boundary.

An Intake Voice Agent handles every inbound call, conflict-checks the caller, captures the matter details, and books the consultation into your calendar. It runs 24/7. It doesn’t take holidays. It doesn’t get sick. And it doesn’t go offline when a third-party vendor loses access to a model.

A Matter Triage Agent reviews form submissions and intake emails, classifies the practice area, scores the lead, and routes it to the right partner with a one-paragraph brief attached. It cuts intake coordinator workload by 60 to 70 percent and gets high-intent leads in front of a partner within 15 minutes instead of the next business day.

A Document Review Agent performs first-pass review on contracts, discovery batches, and matter files. It flags clauses, summarises positions, and produces an associate-grade memo. It doesn’t replace your associates. It gives them a head start so they can focus on the judgment calls that actually require a law degree.

We don’t sell you a subscription to a black-box platform and hope it keeps working. We build the agent, deploy it in your environment, and hand you the keys. You control the infrastructure. You control the data. You control uptime.

No deck. No sales pitch. Three outputs: a workflow map, an agent design spec, and a deployment options memo. Sixty minutes.

A Practical Checklist for Client Intake

If you’re not ready to deploy an agent but want to tighten up your intake process in the meantime, we built a worksheet that walks you through the manual steps most firms miss. The AI Client Intake Checklist for Law Firms covers conflict checking, lead scoring, follow-up timing, and calendar integration. It’s a one-page PDF you can print and hand to your intake coordinator or use as a training guide for new staff. Grab it here: download the checklist.

The Dollar Reality of Leakage

A six-attorney firm doing $3.5M annually loses $80K to $250K a year to intake delays, unbilled admin time, and document review inefficiency. That’s not a guess. It’s the pattern we see when we run the numbers during an audit.

Intake delays cost you 30 to 40 percent of after-hours leads. If you’re getting 30 calls a week after 5 PM and converting 15 percent of them, you’re booking four or five consultations a month. If an agent pushed that conversion rate to 45 percent, you’d book 13 consultations a month. At a $4,000 average retainer, that’s an extra $432K a year.

Billable-hour leakage costs you four to six hours per attorney per week. That’s time spent on intake admin, document review, and matter coordination that never makes it onto an invoice. Across six attorneys billing at an average of $275 an hour, that’s $6,600 to $9,900 a week in unbilled time. Over a year, that’s $343K to $515K.

Document review costs you $200 to $400 an hour of associate time. If your associates spend 15 hours a week on first-pass review across all active matters, that’s $3,000 to $6,000 a week in review costs. An agent cuts that time by 60 to 75 percent. You save $1,800 to $4,500 a week, or $94K to $234K a year.

Add it up and the leakage is real. The firms that close the gap aren’t the ones with the most sophisticated technology. They’re the ones that automate the repetitive work, deploy the automation inside their own boundary, and don’t depend on a vendor’s ability to keep a third-party model online.

What Happens Next

The next export control update is coming. The next model suspension is coming. The next vendor outage is coming. The question isn’t whether your AI platform will go dark. The question is whether you’ll notice when it does.

If your intake agent, document review workflow, or matter triage system depends on a cloud-hosted model you don’t control, you’re one regulatory change away from a weekend of lost revenue and a Monday morning scramble to explain to clients why their work didn’t get done.

If your agent runs on-premise or in a private cloud instance you control, the next suspension won’t touch you. Your intake keeps running. Your document review keeps running. Your clients don’t notice, and your revenue doesn’t take a hit.

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.

Or keep reading about how other firms are using AI to close the leakage gap on our insights page and explore the full Omni platform at omni. The next suspension is coming. Make sure you’re ready.