A partner's words on the call. Projects, skills, bounded context, tool approval, caching, a usage view, training: what each saves and how to set it.
[ KEY TAKEAWAYS ]
A partner on a call last week: "we're at a pretty high burn rate. If we could put in the right parameters and skills that we don't know how to build, someone with experience with law firms might give us some guidance." His four-attorney firm is on Claude Enterprise with Microsoft 365 connected and four pilot users. The bill he feels is not a token invoice, because Enterprise is sold per seat with allowances; it is users hitting limits, chats that crawl, and the API spend behind anything that runs without a person. Nine guardrails bring all three down, and seven of them are settings and habits, not code.
guardrails. Seven are settings and habits. Two need the connector the firm already planned.
habits behind most waste: one chat for everything, whole folders attached, templates re-pasted.
days of per-user usage in the Enterprise CSV export, refreshed daily.
of the standard input price for cache reads on the API, per Anthropic's documentation.
"We need everybody on our team to understand how to use it successfully, so they don't junk it up or sit there and spin and cost us tokens." The partner, on why training is a line item.
Context, read again on every turn. A lawyer keeps one chat open from Monday, pastes the engagement-letter template into it on Tuesday, attaches a 180-page record on Wednesday to ask one question, and by Thursday every short question is carrying all of that. Multiply by four pilot users and you get the partner's sentence: "they sit there and spin and cost us tokens." Nothing about the model changed. The chat got heavy.
What the bill actually is Enterprise is sold per seat with usage allowances; there is no token invoice for chat. What a firm feels is users hitting limits, slow answers on bloated chats, and the API bill for anything that runs unattended. The guardrails target those three.
1
One Project per matter, one for firm standards
Every chat opened inside a Project sees the same files and instructions without re-uploading. The firm's templates and rules live in one shared Project with no client facts; each matter gets its own.
What it saves. The same 40-page template stops being pasted into every chat. Context per turn drops to what the matter needs.
How to set it: Any member creates Projects; the owner sets sharing. One week to set up for a four-attorney firm, including the templates.
2
Skills for the five documents you produce most
A skill is a reusable instruction set: letter, memo, brief, motion, intake form. The user names the document type; the skill carries the style, structure and checks.
What it saves. Shorter prompts, fewer retries, the same output from a paralegal and an associate.
How to set it: Written from three to five real examples of each document in the firm's preferred form.
3
Attach the document, not the drive
Connectors can return a matter's whole document list into a chat. Train users to ask for the one document, and keep SharePoint retrieval scoped to a folder.
What it saves. The single biggest context saver we see. A chat that holds 200 pages it never uses costs on every turn.
How to set it: A one-line rule in the skills library plus the approval step below.
4
Connector write tools on needs approval
Org-wide, each tool a connector exposes is set to always allow, needs approval, or blocked. Writes and bulk reads go to needs approval.
What it saves. Stops runaway tool calls and the context they drag in. Also the matter-safety control, so it pays twice.
How to set it: Owner setting, five minutes, no code.
5
New chat per task
A chat that has lived all week carries all week. Start a new one inside the matter's Project for each task; the Project keeps the knowledge, the chat stays light.
What it saves. Faster answers, fewer limit hits, and no facts from Tuesday's matter in Thursday's letter.
How to set it: Training and a reminder in the skill text. The hardest habit to change, and the one worth the most.
6
Unattended jobs on the API, with caching
The morning summary and the Monday report run from the firm's own cloud through the API. The instruction prefix and the templates are cached, so repeated runs read them at the cached rate.
What it saves. Anthropic documents cache reads at roughly a tenth of the standard input price. The jobs also run under zero data retention, which chat does not offer.
How to set it: Built once in the connector service the firm already owns.
7
Right model for the job
Enterprise lets the owner decide which models members can use. Drafting a cover letter and reasoning through a 300-page record are not the same job.
What it saves. Fewer wasted turns on the heavy model for light tasks, and vice versa.
How to set it: A line in the skills library telling users which model each skill expects.
8
A monthly usage view for the administrator
Enterprise usage analytics: active members, chats per day, projects created, per-user CSV for 90 days, an Analytics API if the administrator wants a dashboard.
What it saves. You find the two people generating most of the usage and fix a habit instead of buying seats.
How to set it: Owner access, no setup. Read it the first Monday of the month.
9
Training with acceptance criteria
Two sessions for pilot users, one for the administrator, recorded, with a written guide. Acceptance: each user opens the right Project, uses a skill, attaches one document, and starts a new chat for a new task, unprompted.
What it saves. Everything above depends on this one. It is the deliverable the partner was asking for when he said the team should not "junk it up".
How to set it: In the scope as a deliverable, not a hope.
Want the nine set up for your firm, with the training that makes them stick?
Thirty minutes with a co-founder: your plan, your connectors, your usage dashboard, and which three of these would change your next month.
Users keep hitting limits by mid-afternoon
New chat per task (5), attach the document not the drive (3), Projects (1). Check the usage view (8) for who.
The same letter comes out five different ways
Skills (2) and the firm-standards Project (1). Training (9) with the acceptance test.
The API bill for the morning summary is growing
Caching on the job (6), scope the sources it reads, and pick the model for the job (7).
A connector dumps whole matters into chats
Needs approval on bulk reads and all writes (4). It is also the matter-safety control.
Nobody knows where the usage goes
The Enterprise analytics dashboard and the 90-day CSV (8), read monthly, before any new seats.
Clio Vincent, Thomson Reuters CoCounsel and Harvey all meter their own AI in their own ways, and firms on Filevine already know what "three chats per user per month, then access pauses" feels like. The difference with Claude Enterprise is that the owner holds the controls. Most firms never open them. The partner we spoke to wanted someone to open them with him, which is the right instinct, and it is what the training deliverable in a well-written scope is for.
In every firm we have looked at, three habits explain most of it: one long chat reused for everything, whole folders attached when one document was needed, and the same firm templates pasted into every conversation instead of living in a Project. Each one multiplies the context Claude has to read on every turn. The Enterprise usage analytics show who and which product; the habits show up in the chats themselves.
Enterprise is sold per seat with usage allowances; you do not get a token invoice for chat. What you feel is users hitting limits, slow answers on bloated chats, and the API bill for anything that runs without a person. The guardrails here cut all three. Check your own agreement for the allowance terms, which Anthropic sets per contract.
On the API, Claude can keep a prompt prefix (your instructions, a template, a long document) cached so repeated calls read it at a fraction of the input price, which Anthropic documents as roughly a tenth of the standard rate for cache reads. It applies to the jobs you run on the API from your own cloud (a morning summary, a Monday report), not to a lawyer typing in chat. For chat, Projects do the equivalent job by holding the context once.
Enterprise owners get a usage analytics dashboard: weekly active members, adoption, chats per day, projects created, Code and Cowork usage, and spend; a per-user CSV covering up to 90 days; and an Analytics API for your own dashboards. Look at it monthly and talk to the two or three people at the top of the list before changing any setting.
Set them to needs approval rather than blocked. A connector call that returns a whole matter's documents into a chat is where context balloons; an approval step makes the user ask for the one document they need. Block only tools nobody has a workflow for.
NEWSLETTER
[ WRITTEN BY ]
Petar Jovanović
Co-Founder & Technical Lead
Co-Founder and Technical Lead at Oktopeak. Builds regulated software for legal and healthcare teams, and leads the rescues of codebases other vendors left half-finished.
[ LEGAL TECH ]
LEGAL TECH
Oct 7, 2026 · 8 min read
Projects, per-user connectors, tool approval, a connector in your own cloud. What each isolates for a four-attorney firm on Enterprise, and what does not.
Read ArticleLEGAL TECH
Jun 29, 2026
Lawyers test Claude on Pro, run out after a couple of documents, and jump to Max without knowing what they actually need. Here is how to choose between Pro, Max, the API, and Enterprise for a law firm, by usage limits, cost, and what each one does with your client data.
Read ArticleLEGAL TECH
Jun 14, 2026 · 9 min read
Claude for Legal launched on 12 May 2026 with 12 practice-area plugins and more than 20 connectors, free to install for any paid Claude customer. A 5 to 50 attorney firm doesn't need a Big Law contract to use it. What it is, which plan fits, what the data terms really say, and the setup checklist we run with firms.
Read Article[ GET STARTED ]
30-minute call. No pitch deck. Just an honest conversation about your project.
