Anthropic just published a case study about a national legal services company, 1,100 employees, that put more than 50 AI agents into production and now has roughly 310 people using Claude every day. The easy reaction is the one you have already had: big company, big budget, in-house dev team, nothing to do with my business.
That is not the story. The story is that the people who run the company's operations built the agents themselves, to streamline their own work.
TL;DR
- The first 50 agents were built by a 15 person group from finance, marketing, and operations, each automating the workflows they run every day. All 15 had working agents within a week.
- The agents do boring work: diagnosing rejected court filings, parsing remittance emails, chasing overdue jobs. Boring is where the return is.
- Every agent starts as a recommender and has to earn autonomy by agreeing with humans. Their review agent hit 98% agreement with the compliance team before it got trust.
- Agent costs follow a J curve: underwater at launch, profitable after tuning. Their total spend fell in July while usage grew.
- The hard part was not the AI. It was process discipline: version control, review gates, audit trails. That is a training problem, and training problems are solvable at any company size.
The people who run the work built the agents
The company's CTO did not hand agent-building to engineering. He picked fifteen people from across the business, finance, marketing, operations, the people closest to the processes that needed streamlining, and gave them Claude Code plus a starter template. They knew exactly which parts of their day were worth automating because they were the ones living them.
Within a week, every one of those fifteen had a working agent. Within a month they had trained their own teams and the fleet passed fifty agents. The bottleneck that routes every piece of automation through a development team, the bottleneck almost every business we talk to has, was gone.
Read that against the deployment gap we covered in April: most companies use AI, almost none have deployed it into actual workflows. This company closed that gap by putting the tools in the hands of the people who run the work.
The agents are boring, and that is the point
Nothing in the fleet is a moonshot. One agent fires when a court rejects a filing, reads the job, checks the court's rules, and posts a diagnosis to Slack within a minute, work that used to eat hours every day. One parses remittance emails and builds accounting payment files for one-click approval. One works their attorney network for hearing coverage, emails attorneys, and reads the replies. One pulls reports and delivers PDFs to customer FTP servers daily, replacing an account manager's weekly grind, and a non-technical employee built it in about an hour.
Every one of those is a task you could describe in one sentence to a new hire. That is the selection filter that matters. The companies winning with agents are not automating judgment. They are automating the repetitive, rule-bound work that judgment was drowning under.
Agents earn trust the way employees do
The discipline underneath the fleet is the part worth stealing. Every agent starts in recommend-only mode: it does the work, a human approves it. Only agents that demonstrate consistent agreement with the humans reviewing them get autonomous authority. Their job-review agent reached 98% agreement with the compliance team. That is how it earned the right to run without a hand on it.
And every change to an agent, its prompt, its rules, its config, arrives as a pull request that a person reviews before it merges. The CTO's line is the best one-sentence summary of safe AI adoption we have seen: if you want an agent involved in a decision, make the decision look like a pull request. Version control, approval, and an audit trail, the same controls you would demand for code, applied to automation that touches your money and your customers.
The costs go up before they go down
The case study is honest about the economics, which is rare. Fleet spend climbed through spring as agents launched, then declined in July even as usage grew. New agents start underwater: they run on the biggest models, with untuned prompts, while the team writes evaluations. Then they get optimized onto smaller, faster models and the cost per task drops. The company now measures an efficiency ratio, value delivered against runtime cost, per agent, and has killed tasks that do not justify agent economics.
That J curve is the piece to internalize before you start. If you judge an agent in week two, you will judge it at the most expensive point it will ever be. If you never tune it, you will pay week-two prices forever.
What this means at your size
You do not have 1,100 employees. You also do not need 50 agents. The pattern scales down cleanly because none of its ingredients are enterprise ingredients: pick the boring, describable, repetitive work first, start every agent as a recommender, keep a human approval on anything that touches money or customers, and expect the cost curve to bend only after tuning. A five person office with two well-chosen agents captures the same kind of return per person as this fleet does.
The real barrier the case study exposes is not technical capacity. It is that somebody has to own the discipline: the templates, the review gates, the audit trail. We build and run managed agents for clients with exactly that structure, because the structure, not the model, is what decides whether an agent becomes an asset or a liability.
If there is a repetitive process in your business you could describe in one sentence, that is probably an agent. Get in touch and we will tell you honestly whether the economics justify building it.