CyberSecureRIA

AI for RIAs, session 1

Setting up Claude for an RIA

Our recommended settings for a Claude Team account, admin side first, then each person's own settings. The short version comes first. The detail, with links to Anthropic's documentation, follows.

These are starting points for a typical small-to-mid RIA. Your compliance program and your clients decide the final answer.

Current as of November 19, 2026. Anthropic changes these screens often.

The short version

On turn it on / leave onOff turn it off / leave offYour call depends on the firmSkip not needed for most RIAs

Firm: organization and access

SettingRec.Why, in one line
Team name and logoOnStaff can see at a glance they're in the firm account, not a personal one.
Verify your domainOnUnlocks SSO and stops new personal accounts on your email domain.
SSO and "Require SSO"OnClaude inherits your Microsoft or Google MFA and sign-in rules.
User provisioningInvite onlyEvery seat is a deliberate, documented decision.
Organization instructionsOnBrief Claude on who you are and your rules. Useful context, not a control.

Firm: usage and billing

Extra usage creditsYour callPrepaid credits for people past their limit. Upgrade the seat if someone hits limits often.
Spend limitsSet lowOn Team the firm-wide cap may be your only brake on a runaway task.

Firm: data and privacy

Share chatsOnSharing stays inside the firm on Team. Need-to-know still applies.
Share chats that use connectorsOnLean on, but a shared chat can show a coworker data they can't open themselves.
Share Claude Code sessionsOffOff until someone at the firm actually uses Claude Code.
Share projectsOnPeople will want shared projects. Chats inside them stay private.
Data exportYour callOwner export for records; member self-export off unless needed. Export on a schedule.
Rate chats (thumbs up/down)OffFeedback is the exception to "not used for training." Rated chats are kept up to 5 years.
Location metadataYour callCity-level from your IP, not GPS. Handy for local questions.

Firm: capabilities

Web searchOnNeeded in practice. Results are still outside content.
Visual and interactive contentOnMaps and image results. Google and Bing are involved.
Built-in browser (Cowork)OnUseful for Cowork tasks.
Import cookies from other browsersOffThe most important switch on the page. It hands Claude your logged-in sessions.
Site permissionsYour callAllow all is defensible only while Claude in Chrome is off. One list covers both.
Artifact connectors, inline visualizationsOnEach viewer uses their own connections.
Allow "Automatically approve" modeOnAvailable for people who choose it, once connector tools are dialed in.
Start new tasks in Automatically approveOffNobody should land in auto-approve by accident.
Allow "Skip all approvals"OffIn Skip, nothing checks Claude's actions at all.
Allow "Always allow" for connector toolsOnPrevents alert fatigue. Lock down write tools first.
Code execution and file creationOnNeeded for documents, spreadsheets and artifacts.
Network egressLimitedPackage managers plus a short list of trusted sites (your website, sec.gov).
AI-powered artifactsOnRuns on the viewer's plan and connectors, not the author's.
Computer accessOffOff most of the time. No sandbox between Claude and your apps.
Scheduled tasksOnValuable, but they run unattended. Watch modes and spend.
MemoryOnMakes Claude far more useful. Staff in older orgs must turn it on themselves.

Firm: other admin pages

Claude in ChromeOffThe single most dangerous setting. Acts as the signed-in user in their own browser.
Claude TagSkipSlack only for now.
CoworkOnIt merges with Chat on December 1 anyway. Set the guardrails first.
DispatchOffRemote control of a workstation from a phone, often with nobody at the desk.
Artifacts: external sharingOffAnything shared with a client is a record you need to keep. Team has no audit log.
Artifacts: other settingsOnInternal sharing and creation are fine.
Claude DesignYour callThe standalone app closes December 14. Move anything you built.
PluginsDefaultsLeave the defaults. Treat third-party plugins like any software install.
ConnectorsYour callMost useful and most risky. Trusted vendors only, per-tool permissions, in layers.
Desktop extensionsOffRun locally with the user's own access, so the other layers don't apply.
Members, Claude Code, Git providers, Cloud environmentsSkipOnly for firms doing custom development.

Personal settings (each person)

Personal instructionsOnYour role, working style and preferred format.
Memory: search chats, generate memory, sensitive topicsOnAll three. Do client work in a project per client or household.
Capabilities: artifacts, AI-powered artifacts, chartsOnYour toggles can only narrow what the firm allows.
Customize: connectorsYour callConnect only what you need. Tighten tool settings as you like.
VoiceYour callTry the voices and pick one.
Meetings (beta)WaitNeeds Google Calendar today; most RIAs run on Microsoft.
NotificationsOnHow you find out a task is waiting for approval.
Design systems, skills, Claude CodeLaterFuture sessions.
Chat vs Cowork, models, effort, per-chat connectorsKnow themStart on the default model; turn off connectors a task doesn't need.

Every way data can leave Claude

If Claude reads something hostile (a web page, an email, a review), instructions hidden in it can try to steer Claude. These are the doors it could use to send data out. Close or narrow each door rather than trusting the model to refuse. Orange cards can act or send on their own; blue cards leak what Claude reads or shares.

Code network egressCode Claude runs can call out to the internet.Control: package managers plus a short domain list.
Cowork built-in browserBrowses and fills in sites.Control: site permissions; cookie import off.
Claude in ChromeActs as the signed-in user.Control: off until staff are trained.
Connector write toolsSends, posts, edits, deletes.Control: Blocked or approval; remove scopes in Entra.
The user's computerDesktop extensions, computer access, Dispatch.Control: off; these skip the vendor and Entra layers.
Connector read toolsSearches go to the vendor.Control: only vendors you would contract with.
Web search and fetchA URL can carry data.Control: web search toggle; per-chat switches.
External sharingAnyone with the link.Control: artifact external sharing off.
Claude CodeTerminal and web fetch.Control: leave unused unless someone develops.

Plugins and scheduled tasks are not new doors; they reuse these.

Firm settings in detail

For the Owner or Admin of the firm's Claude Team account, in the order they appear in the admin console. Menu paths are from Anthropic's help center and may differ slightly on your screen.

Organization and access

Team name and logo On

Set the firm's name and logo. It takes fifteen seconds and tells staff they are working in the firm's account, where the firm's controls apply, rather than a personal one.

Docs: Account management FAQs

Domain verification On

Organization and access > Domains > Add or edit domains > Verify

Verifying your email domain does two things. It lets you turn on SSO, and on Team it lets you block anyone from creating a new personal Claude account with a firm email address. We have seen staff do that by accident.

  • Blocking only stops new accounts. Anyone who already signed up with a work email keeps that personal account; claiming those accounts is Enterprise-only. Ask your team before you flip it on.
  • It does nothing about staff using a personal Gmail address. That is a policy item.

Docs: Claim and migrate accounts on your domain

Single sign-on and "Require SSO for Claude" On

Organization and access > Authentication > Setup SSO

With SSO, Claude sits behind your Microsoft or Google sign-in, so it inherits your MFA and conditional access rules. Also turn on Require SSO for Claude: without it, people can still choose "Continue with email" and skip your MFA. Anyone not assigned to the Claude app in your identity provider will be locked out, so check assignments first.

Docs: Set up single sign-on (SSO)

User provisioning: Invite only vs JIT Invite only

Invite only (the default) means you send each person an invitation. Just-in-time (JIT) gives an account to anyone assigned to the Claude app in your identity provider the first time they sign in. For an RIA, invite only keeps every seat a deliberate, documented decision.

Docs: Set up single sign-on (SSO)

Organization instructions On

Instructions Claude follows in every conversation for everyone at the firm. Owners only; up to 3,000 characters; changes can take up to an hour to apply.

  • Good content: firm name, SEC or state registration, client base and niche, services, house style for client communications, and the regulatory frame ("we are a fiduciary; avoid guarantees and performance promises; flag anything that might be marketing under the SEC Marketing Rule").
  • Results are mixed. These are background context sitting alongside each person's own instructions, not hard rules. Never rely on them as a compliance control; the tool permissions below are the control.

Think of it as the briefing you'd give a smart new hire on day one.

Docs: Set organization instructions

Usage and billing

Extra usage credits and seat types Your call

This is where you load prepaid credits for people who go past their seat's limit. Extra usage is billed at standard API rates, and limits are per person; they can't be pooled.

If someone hits their limit regularly, upgrade their seat from Standard to Premium rather than buying more credits. The time they lose waiting costs more than the seat difference, and the most capable models use limits quickly.

Docs: Extra usage for Team and seat-based Enterprise · What is the Team plan? · Pricing

Spend limits Set low

Organization settings > Usage > Usage and spend limits

You can always cap total extra spend for the firm. Per-person caps exist "where available," and some Team organizations (ours included) don't have them. Without per-person caps, one runaway scheduled task can burn tens or hundreds of dollars quickly. Set the firm-wide cap low and raise it on purpose.

Docs: Extra usage for Team and seat-based Enterprise

Data and privacy

Background: Anthropic does not train on Team chats by default. Team chats are kept until the person deletes them. Custom retention, audit logs, the Compliance API and SCIM are Enterprise-only, so on Team your data export is the main record you can hand an examiner.

Share chats On

On Team, a shared chat is visible only to people in your organization; public chat links exist only on personal plans. A shared chat can carry client information, so the same need-to-know rules apply as for any file.

Docs: Share and unshare chats · Public links for shared chats

Share chats that use connectors On

This does not stop anyone using connectors; it controls whether chats that used them can be shared. Viewers see the conversation and Claude's answers, not the raw tool data, but an answer can still quote or summarize a mailbox or folder the viewer couldn't open themselves. Lean on, and make it a policy point.

Docs: Share and unshare chats

Share Claude Code sessions Off

Claude Code comes with every Team seat, so anyone could start using it. Leave sharing off until someone at the firm has a real need; "off by default, on when needed" is easy to defend.

Share projects On

People will want shared projects. Sharing a project shares its instructions and knowledge files, not the chats inside it: those stay private unless each one is shared.

Docs: Manage project visibility and sharing · Control project sharing

Data export Your call

  • Owner export: every member's chats, including incognito chats, plus memory. Original uploaded files are not included.
  • Anything deleted, by a member or by retention, is gone from later exports, and Team has no retention lock. If books-and-records matter to you, export on a schedule.
  • Incognito chats are kept 30 days and appear in owner exports within that window. Incognito hides a chat from your history; it does not hide it from the firm.
  • Member self-export: leave off unless someone needs it for a period.

Docs: Export your organization's data · Export your Claude data · Use incognito chats

Rate chats Off

Feedback is an explicit exception to "not trained on by default." When someone clicks thumbs up or down, the whole related conversation can be stored for up to five years (de-linked from user IDs) and used to improve models.

Docs: Manage user feedback settings · Is my data used for model training?

Location metadata Your call

Coarse city or region from your IP address, used for things like local search. It is not GPS. Handy when traveling. Members can also turn off their own.

Docs: Does Claude use my team members' location?

Capabilities

Web search On

You need it in practice. Remember that search results are outside content Claude reads, like any web page.

Visual and interactive content On

Maps (including weather and Google Maps) and image results (Bing). Your prompts are not sent; queries become generic search requests. Google receives IP addresses. Note the third parties in your vendor records.

Docs: Use visual and interactive content

Built-in browser On and Import cookies Off

The built-in browser opens in the Claude desktop app during a Cowork task; Claude opens sites, reads pages, clicks, types and fills forms while you watch. Keep it on.

Turn "Import cookies from other browsers" off. It is on by default on Team. It lets the built-in browser reuse your logged-in sessions from your normal browser, which for an advisor can mean the custodian portal, the CRM and email. On first use every site starts selected.

Docs: Set up browser use in Cowork · Use the built-in browser

Manage site permissions Your call

Choose a default for all sites (allow all or block all), then add exceptions to the allow or block list. One list governs both the Cowork built-in browser and Claude in Chrome, and you can't separate them today.

  • Built-in browser only, cookie import off: allow all is defensible. Blocklist custodian, banking and money-movement sites.
  • Claude in Chrome on: allow all is much more dangerous, because Chrome has everything the user is logged into. Anthropic recommends a separate browser profile without sensitive accounts.

Docs: Claude in Chrome admin controls · Use Claude in Chrome safely

Artifact connectors and inline visualizations On

Both help a typical RIA get value from Claude. Artifact connectors are all-or-nothing for the org, but each person connects their own apps, even inside a shared artifact.

Docs: Artifacts admin guide

Approval modes: how they interact with connector permissions

Cowork has three modes, and they change what the admin's per-tool settings mean:

  • Manual: Always-allow tools run; Needs-approval tools prompt; Blocked is denied.
  • Automatically approve: Needs-approval tools run unless an automated check flags them. Anthropic says the check "can miss risky actions." Blocked is denied.
  • Skip all approvals: Needs-approval tools run with no prompt and "nothing checks its actions automatically." Blocked is still denied.

Docs: Get started with Cowork · Cowork on Team and Enterprise

Allow "Automatically approve" On · Start new tasks in it Off

On Team both are on by default. Leave auto-approve available for someone who chooses it per task, but turn off the default so nobody lands in it by accident. Revisit once your connector tools are set to Needs approval or Blocked.

Allow "Skip all approvals" Off

Owners can turn Skip off entirely. When off, people can't choose it for scheduled tasks or the built-in browser.

Docs: Set up the unified Claude experience for Team

Allow "Always allow" for connector tools On

Off by default. Without it, people click Allow on every prompt, get alert fatigue, and approvals stop meaning anything. An approval prompt you see fifty times a day is not a control.

Order matters: set write tools (send, delete, forward) to Blocked or Needs approval at the firm level first, then turn this on, so people are only prompted for the actions that deserve attention.

Code execution and file creation On · Network egress Limited

Keep code execution on so Claude can create documents, spreadsheets and artifacts. For network egress there are four levels: off, package managers only, package managers plus specific domains, and all domains (which Anthropic calls "the riskiest option").

  • Choose package managers plus a short list: your own website, sec.gov, and other sites outsiders can't post to or receive data at. Every allowed domain is a possible exit.
  • Anthropic's own words: hidden instructions in a file or site could make Claude send "information from its context" to a third party; "disabling network access eliminates this risk entirely."
  • Egress does not cover connectors. Those are controlled separately.

Docs: Create and edit files with Claude

AI-powered artifacts On

Mini apps that call Claude. Each viewer's use counts toward their own plan and uses their own connectors, so a shared app does not expose the author's mailbox to colleagues.

Docs: Prototype AI-powered apps with artifacts

Computer access Off

Off most of the time. Computer use "has no sandbox between Claude and your applications," Claude can see anything on screen, and Anthropic says not to give it access to sensitive apps such as banking. For an advisor, that means the custodian portal and CRM. (Anthropic's docs disagree on whether Team has full computer use; read the label on your own screen.)

Docs: Let Claude use your computer in Cowork

Scheduled tasks On

Scheduled tasks run in the cloud on their cadence, even when the laptop is asleep. That is where an auto-approve mode can do damage with nobody watching, and where runaway spend comes from. As far as we can tell you cannot limit a task to specific connectors; it runs with whatever the person has connected. Use the task's approval mode and firm-level Blocked tools.

Docs: Schedule recurring tasks in Cowork

Memory On

The owner toggle only makes memory available; each person controls their own. In orgs created before October 31, 2026, memory starts off for every member until they turn it on, so tell staff to do it.

  • Each project has its own memory, which helps keep one client's work separate.
  • Owners can't view or edit a person's memories. Incognito chats don't create memory. Memory is included in the owner export.

Docs: Chat search and memory

Other admin pages

Claude in Chrome Off

The single most dangerous setting. The extension runs in the person's own Chrome, so it can reach whatever they are logged into, and it keeps working when they switch tabs. Anthropic calls it "enhanced with our safety classifiers but is still risky."

  • We have seen it start acting in a different browser, on a different computer, where the extension was signed in. Sign it out of browsers you aren't using.
  • Turning it on is what makes an "allow all" site list dangerous (see site permissions).
  • Turn it on later, after you've trained staff.

Docs: Getting started with Claude in Chrome · Use Claude in Chrome safely

Claude Tag Skip

Claude in Slack. Most RIAs run on Microsoft Teams, so it doesn't apply yet.

Cowork On

Chat and Cowork become one experience on Team on December 1, 2026. An owner can hold it off, but the separate experience retires February 1, 2027. The real decision is the guardrails above: Skip off, auto-approve not the default, connector tools dialed in. Set those before December 1.

Docs: Set up the unified Claude experience for Team

Dispatch Off

Dispatch lets someone message Claude from a phone and have it work on their desktop, using their connectors and, through computer use, their screen directly. For an RIA that is unattended remote control of a workstation holding client data. Anthropic: "mistakes (or malicious content the model encounters along the way) can have real consequences."

Docs: Assign tasks from anywhere

Artifacts: external sharing Off · other settings On

External sharing lets anyone with the link open an artifact. Anything shared with a client is a client communication, subject to books-and-records retention and possibly the Marketing Rule, and Team has no audit log to show what was shared. Keep it off; leave the internal settings on.

Docs: Artifacts admin guide

Claude Design Your call

The standalone Claude Design app closes December 14, 2026; design moves into regular Claude. Projects can move, chats and comments cannot. If your firm built a brand design system there, move it before the 14th.

Docs: Move from standalone Claude Design

Plugins Defaults

A plugin bundles skills (saved instructions, like a mini SOP), connectors and commands so a team gets the same setup in one install. For each plugin an owner picks hidden, available, installed for everyone, or required. A plugin is third-party code and instructions running with the person's own tools, so treat it like any software install: Anthropic's own first, others only after someone has read them. Leave the defaults for now; skills get their own session.

Docs: Manage plugins · One Plugins page

Connectors Your call

A connector gives Claude access to another tool: email, files, CRM, almost anything. Most use MCP (Model Context Protocol), an open standard Anthropic published in 2024. Connectors are what make Claude so useful, and one of the biggest risk areas.

  • Directory vs custom. Directory connectors are listed by vendors like Microsoft, Google and Box. Custom connectors are any server URL; that is the riskier category.
  • A connector can change at any time. It runs on the vendor's server, so the vendor can add or change tools without anything changing on your side, and today's per-tool settings don't cover tomorrow's new tool. Stick to vendors you trust and review tool lists periodically.
  • Low-risk fun connectors are fine. Allowing something harmless (a hiking or audiobook app) helps adoption and keeps people off personal accounts. The test: no write tools, no client data. Outside content is still a possible source of hidden instructions.
  • Per-tool permissions. Read-only tools on Always allow; write tools on Needs approval; anything that sends outside the firm or deletes on Blocked unless there's a clear need.
Set permissions in layers.
1. Claude's admin settings (Always allow / Needs approval / Blocked). The bluntest layer, and it only holds until the tools change.
2. The user's own sign-in. Most connectors use each person's own credentials, so Claude can't exceed their access. Ask how a connector authenticates: a shared service account gives everyone its access. And fix oversharing at the source, because Claude will find everything a person can open.
3. The platform's own permissions. For Microsoft 365, connect Claude, then remove the permissions you don't want (for example Mail.Send) in Microsoft Entra, and require admin consent so they don't come back. If Anthropic changes a default, your Microsoft settings still hold.

Anthropic offers "Claude for Financial Advisors," a plugin with about 20 partner connectors (CRMs, custodians, portfolio tools). Its page recommends Enterprise for RIAs. There is no official SEC or FINRA connector.

Docs: Use connectors · Microsoft 365 connector · Claude for Financial Advisors · Model Context Protocol

Desktop extensions Off

One-click installs of local MCP servers in the desktop app. They run on the person's computer with that person's file and system access, so the vendor sign-in and Entra layers don't apply. Allow specific ones only for staff who know what they're doing.

Members, Claude Code, Git providers, Cloud environments Skip

Relevant only to firms doing custom development.

Personal settings in detail

Each person's own settings. The personal page looks like a smaller copy of the admin page because a personal toggle can only narrow what the firm allows.

Personal instructions On

Applies only to your own conversations, alongside the firm's instructions. Put your role, how you like to work, and your preferred format here.

Memory On

Turn on all three: search and reference past chats, generate memory from chats, and include sensitive topics. In a relationship business, remembering things about yourself and your clients makes Claude far more useful.

  • Do client work inside a project per client or household, so those memories stay with that client.
  • Review your memory now and then: open a topic, edit or delete it. Reset permanently deletes everything, including project memories.

Docs: Chat search and memory

Capabilities On

Artifacts on; AI-powered artifacts on for most people; charts and diagrams in chat on.

Customize: connectors Your call

This is where you connect the tools the firm has approved, signing in with your own account. You can change each tool between Always allow, Needs approval and Blocked for yourself. The firm's choices set the boundary; check on your own screen which options are open to you.

Voice Your call

Try the voices and pick one you like.

Meetings (beta) Wait

Still in beta, and today it needs Google Calendar to join meetings; most RIAs are on Microsoft 365. Any AI meeting capture also raises recording-consent and books-and-records questions.

Notifications On

Leave task and approval notifications on. An unattended task that needs approval may wait for you or finish without the tool.

Design systems, skills, Claude Code Later

Design systems teach Claude your brand and are very powerful; skills are saved instructions Claude loads when a task calls for them. Both get future sessions. Most people can leave Claude Code alone for now.

Main page: Chat vs Cowork, models, effort, microphone, connectors

  • Chat is a conversation; Cowork is delegating a task. In Cowork Claude works through several steps on its own with files, connectors, the browser and scheduled runs, stopping for approval according to the mode. They merge December 1.
  • Models: start on the default. Move up for hard reasoning or long documents; the most capable model uses limits fastest.
  • Effort: higher means Claude thinks longer: better on hard problems, slower, more usage. Default is fine day to day.
  • Microphone vs voice mode: the microphone is dictation into the box; voice mode is a spoken back-and-forth.
  • Per-chat connector toggles: turn off connectors a task doesn't need. Fewer connectors means less outside content and less chance of reaching into the wrong system. In our testing the choice seemed to carry into the next chat, so check before you assume.

Docs: Get started with Cowork

Before December 1

  1. Turn off "Skip all approvals" and "Start new tasks in Automatically approve."
  2. Set connector write tools to Needs approval or Blocked, then turn on "Always allow" for connector tools.
  3. Turn off "Import cookies from other browsers," Claude in Chrome, Dispatch and artifact external sharing.
  4. Set a low firm-wide spend limit.
  5. For Microsoft 365, remove unwanted permissions in Entra and require admin consent.
  6. Tell staff to turn on memory and do client work in per-client projects.
  7. Schedule regular data exports if you rely on them for records.