Your Tools Already Tell You What Happened. This Closes the Loop.
Anthropic's enterprise-search plugin finds things across all your connected tools and hands you a tidy, sourced answer — then stops. You still triage the queue, write the reply, file the ticket, fix the board, send the rollup and update the CRM by hand. Enterprise Actions runs that loop for you across Slack, Gmail, Linear, Jira, Asana, Notion, Confluence, Intercom, Salesforce and Fireflies — and never takes a risky action without your say-so.
Built directly on Anthropic's enterprise-search plugin. Its three retrieval skills are kept byte-identical and called as the context engine; /search and /digest are retained verbatim. Enterprise Actions adds the six acting agents and a deterministic safety gate on top. They retrieve; we close the loop — safely. This is a working acting plugin you can run today; autonomous writes to your systems activate when you connect your own MCP sources.
Why 'AI for Work' Mostly Stops at the Answer
The numbers are consistent across sources. Asana finds 58–60% of the day goes to 'work about work' — communicating, chasing status, switching tools — versus about a third on skilled work. McKinsey's split is 28% email plus 14% internal communication against only 19% searching. Microsoft puts communication at roughly 60% of the day. Search tools, including Anthropic's own enterprise-search, attack the 19% and stop at the answer. The 42%+ — triaging the queue, drafting the reply, filing the ticket, fixing the board, sending the rollup, updating the CRM — is still done by a human, one tab at a time. The failure of most 'AI assistants' is not that they retrieve badly; it is that they hand the work back.
What most 'AI for work' tools leave on your plate
- ✗Search returns the matching messages, and you still read, sort, label and route 200 unreads by hand.
- ✗A digest summarizes the thread, and you still write the reply, check the ship date, and chase the right source.
- ✗The meeting ends with decisions and owners in the transcript — and 54% of the room still leaves unsure who owns what.
- ✗The board says 'in progress' while chat says 'shipped', and nobody reconciles the drift until the sprint review.
What Enterprise Actions does instead
- ✓It triages and routes the queue, then acts — labels, priority, assignment, snooze — and hands back a sorted, owned queue, not a list.
- ✓It drafts the reply in your voice with every claim cited, stages it internally, and holds any external send for your confirm.
- ✓It turns the meeting into filed tickets with owners and due dates and a posted recap — the action items become tracked work.
- ✓It detects chat-vs-tracker drift continuously and writes the status fix, flags blockers with evidence, and drafts the nudges.
One Inbound, From Signal to Side Effect — No Manual Loop
Here is what actually happens when something lands in your stack. Follow it through all five steps of the acting loop. The gates are not suggestions — they block, hold for confirmation, or escalate.
- 1
Triage — rank and own the queue
The agent pulls the inbound across your connected chat, email, ticketing and tracker into a single queue and classifies each item by urgency, type and owner. The reversible, confident moves — applying a label, setting priority, snoozing an FYI, routing a thread to the right owner — fire and are logged so you can undo them. Bulk or destructive operations are held for confirmation. You get back a sorted, owned queue and a short open-loops list.
- 2
Synthesize — gather the cross-tool context (this step is enterprise-search)
For each item that needs an action, the agent calls Anthropic's retained retrieval skills byte-identical: search-strategy gathers the full thread plus related docs and account history across tools, knowledge-synthesis dedupes and attributes it with confidence scores, and source-management confirms which connectors are live and in-scope. No new retrieval engine — Anthropic's own, used as substrate.
- 3
Draft — produce the artifact, every claim cited
The agent drafts the reply, the recap, the rollup or the CRM note in the right register, and every factual claim must cite a real synthesized source. This is anti-hallucination for actions, not just answers: a draft whose evidence IDs do not resolve to real retrieved records is rejected before it ever reaches the act step.
- 4
Act — fire the side effect, through the safety gate
Every write or external action is classified by reversibility, blast-radius, confidence and whether its claims are cited. Reversible, confident, fully cited actions run and are logged. The five hard-gated classes — anything sent to an outside party, anything that overwrites a trusted record, commitments on someone else's behalf, numbers put in front of leadership, and bulk or destructive operations — are never auto-fired; they wait for your explicit confirmation. Uncited or low-confidence actions are blocked or escalated with the conflict shown.
- 5
Log — write the audit trail
Every action it takes is appended to a full action ledger: what it did, why, the evidence behind it, whether it was reversible, and the gate's decision. 'Why did the agent do that?' is always answerable, and reversible actions can be undone. Then the loop is ready for the next inbound.
Six Things Enterprise Actions Does That a Search Tool Cannot
Every Enterprise Actions agent runs the same five-step loop — triage, synthesize, draft, act, log. The synthesize step is enterprise-search. The act step is new, and it always produces a concrete side effect in a system of record: a label set, a ticket filed, a reply staged or sent, a status changed, a rollup posted, a CRM record updated. Each agent is rejected from this set if it only retrieves.
/triage — clears the queue, doesn't just read it
Pulls your unread inbox, Slack mentions, assigned tickets and board items into one ranked queue, classifies each by urgency, type and owner, then acts: it sets labels and priority, assigns, snoozes and routes. The done-artifact is a mutated queue plus a short open-loops list — not a prettier summary. Bulk, destructive or auto-archive operations are always held for your confirmation. Example: instead of 40 minutes of inbox archaeology, the three things that actually need you are at the top, the FYIs are filed, and two threads are routed to the right owner.
/draft-reply — the cited reply, written and ready to send
For any message, email or ticket it synthesizes the full thread plus related docs and account history, then drafts a reply in your voice with every factual claim cited to a real retrieved source. Internal drafts are staged automatically; anything sent to an outside party waits for your explicit confirmation — always, no exception. Example: a customer asks where their integration stands, and the reply quotes the actual ship date from the Linear ticket, not a guess. You read it, tweak one line, hit send.
/meeting-to-work — turns the meeting into tracked work
Takes a meeting transcript or notes, extracts the decisions and action items, and acts: it files real tickets with owners and due dates, drafts the follow-up emails, and posts a clean recap to the channel. Committing other people to work — assigning tasks or due dates — is held for confirmation. Example: the planning call ends and, by the time you are back at your desk, four decisions are a posted recap and six action items are tickets with owners, each waiting for your confirm before it pings anyone.
/board-sync — keeps the board honest
Compares what your chat says is happening against what the tracker shows, finds the drift — 'this shipped Tuesday but the ticket is still In Progress' — and acts: it updates statuses, flags real blockers with the evidence attached, and drafts nudges to the owners. Moving or closing a ticket, or marking work done, always asks first. The status write is the headline, not a footnote to a drift report — a run that only produces a report is an incomplete run. Example: five tickets whose Slack threads say 'done' while the board says 'in progress' get corrected, one genuinely blocked item is flagged with a link to the blocking thread.
/rollup — the stakeholder update, drafted and delivered
This is not /digest. /digest summarizes activity for you and stops. /rollup synthesizes the week across tools and metrics, drafts it in your stakeholders' register, and delivers it — it posts to the channel or sends the email. Because it can put numbers in front of leadership, every metric and every wide send is held for your confirmation. A drafted-but-undelivered rollup is an incomplete run, by design. Example: /rollup --weekly assembles the exec update from Linear, Notion and your analytics, drafts it the way your VP likes to read it, and shows it to you with the two metrics highlighted before it posts.
/crm-log — the CRM hygiene nobody wants to do
Logs your calls, emails and meetings against the right account and updates stage, next-step and contacts — the data-entry chore most sellers call their number-one time sink and almost none of them trust. Overwriting a stage or forecast always asks first, and it only ever touches records you already have permission to edit. Example: after a customer call it reads the Fireflies transcript, logs the notes to the opportunity, sets the next step and date, and proposes moving the stage — then waits for your nod before changing any forecast number anyone relies on.
Built on Anthropic's enterprise-search — and Honest About It
Enterprise Actions does not reimplement search and does not replace enterprise-search. It keeps Anthropic's three retrieval skills byte-identical and calls them as its synthesis substrate, and it retains the /search and /digest commands verbatim — your out-of-the-box cross-tool search and daily or weekly digest still work exactly as before. The acting layer is a consumer of retrieval, never a substitute for it. Every command, skill and example in enterprise-search is retrieval and synthesis; it ships no write or send capability — verified against the plugin source. That is precisely the line Enterprise Actions crosses: they retrieve; we close the loop. Here is what enterprise-search does in each area, and what the acting layer turns it into.
View Anthropic's enterprise-search plugin on GitHub
search-strategy (retained, byte-identical)
enterprise-search (out of the box)
enterprise-search uses search-strategy to plan and run cross-tool retrieval — deciding what to look for and where, across every connected source.
With Enterprise Actions
Enterprise Actions calls the same skill unmodified as the context-gathering step of every acting loop. The moment the acting layer edited it, the 'built on Anthropic's own retrieval' claim would become a reimplementation — so it stays untouched and is simply consumed.
knowledge-synthesis (retained, byte-identical)
enterprise-search (out of the box)
enterprise-search uses knowledge-synthesis to dedupe, attribute and confidence-score the gathered context into a single coherent answer with source attribution.
With Enterprise Actions
The acting layer reuses that same attribution and confidence-scoring to decide whether an action is even allowed to fire. Every factual claim in anything it drafts inherits this skill's source attribution — and an action whose evidence does not resolve to a real retrieved record is rejected.
source-management (retained, byte-identical)
enterprise-search (out of the box)
enterprise-search uses source-management to know which connectors are live and in-scope before retrieving from them.
With Enterprise Actions
Enterprise Actions consults the same skill to check which connectors are authorised and in-scope before acting on them — and it never acts outside your own connector permissions. If you cannot do it by hand, neither can the agent.
/search and /digest (retained, verbatim)
enterprise-search (out of the box)
/search runs a cross-tool search and returns a sourced answer. /digest renders a daily or weekly summary — it even classifies activity into action items, decisions and mentions and draws checkboxes — while writing nothing to any system.
With Enterprise Actions
Both commands are retained verbatim, so nothing you rely on today changes. /digest still summarises for you and stops. /rollup — the acting equivalent — synthesises the week, drafts it in your stakeholders' register, and delivers it (posts or sends, gated). The difference between the two is the terminal write, and it is load-bearing.
Is This the Right Tool?
Best Fit
- Individual knowledge workers — ops, PM, support, CS, sales, engineering managers — drowning in the queue inside a company already on Cowork or Claude Code with MCP connectors authorized
- Teams who already use Anthropic's enterprise-search and want to add the acting layer on top — keeping /search and /digest unchanged — without ripping anything out
- Anyone whose real bottleneck is doing the work after the answer: triage, drafting, meeting follow-through, board hygiene, weekly rollups and CRM upkeep across more than one tool
- Buyers who want cross-vendor acting without a separate index, separate seats or an admin rollout — a plugin that rides the seats and connectors they already pay for
Not the Right Fit
- —Teams with no connected sources — with nothing authorized in chat, email, ticketing or CRM, there is nothing to act on, so the agents mostly propose
- —Anyone who wants full hands-off autonomy with zero review — every external or irreversible action here is gated, evidence-bound and logged on purpose
- —Teams looking for a heavyweight standalone platform with its own index and seats — this runs inside Claude Code or Cowork as a plugin, not as a separate product
From Install to First Closed Loop in Four Steps
- 1Install in Claude Code or Cowork with one command — under two minutes
- 2Connect your stack: Slack, Gmail, Linear or Jira, Notion, Intercom, Salesforce, Fireflies
- 3Run /triage on a real queue — watch it sort, label and route, with the risky external send held at the gate
- 4Turn on the acting agents you want — /draft-reply, /meeting-to-work, /board-sync, /rollup, /crm-log — each gated, evidence-bound and logged
Stop Doing the Work After the Answer.
One command to install. Run /triage on a real queue and watch it sort, own and route — with every risky action held at the gate. Free, open source, and built on Anthropic's enterprise-search plugin.
No spam, ever. Your email is stored securely so we can send you updates about new use cases and workflows.
Step-by-Step: Install the Plugin
Add the plugin
In Claude Code or Cowork, run: claude plugins add knowledge-work-plugins/enterprise-actions — this installs the six acting agents and their commands, the retained Anthropic retrieval skills and /search and /digest verbatim, the deterministic action-safety gate, and the append-only action ledger.
Connect your systems
Open the connector config and authenticate your chat (Slack), email (Gmail), project tracker (Linear, Jira, Asana, monday.com, ClickUp), knowledge base (Notion, Confluence), customer support (Intercom), CRM (Salesforce), meeting notes (Fireflies) and analytics (Amplitude). It uses the same tool-agnostic connector model as enterprise-search — bring whatever you already use in each category.
Run /triage first
Start with /triage on a real Monday inbox plus Slack plus tracker queue. Watch it rank and own the queue: it labels, prioritises, assigns and routes the reversible items, and holds bulk or destructive operations for your confirmation. This is the fastest way to see the loop and the gate working together.
Turn on the rest of the loop
Enable the other agents as trust builds — /draft-reply, /meeting-to-work, /board-sync, /rollup and /crm-log. The action-safety gate stays active at every stage: the five hard-gated classes always wait for your confirmation, and every action is written to the ledger so you can audit or undo it.
Enterprise Actions for Claude — FAQ (2026)
enterprise-search is an excellent retrieval layer — it finds things across all your connected tools and hands you a single coherent answer with source attribution. Verified against the plugin source: every command, skill and example is retrieval and synthesis, and it ships no write or send capability. Even /digest, which classifies activity into action items, decisions and mentions and draws checkboxes, mutates nothing. Enterprise Actions keeps all of that — the three retrieval skills are byte-identical and /search and /digest are retained verbatim — and adds the layer enterprise-search deliberately leaves to you: six acting agents that triage, draft, file, update, post and send, each with a deterministic safety gate. The honest one-line version: they retrieve; we close the loop.
Only within the gates, and the gates are deterministic code, not adjectives. Every write or external action is classified by reversibility, blast-radius, confidence and whether its claims are cited. Reversible, confident, fully cited actions — a label, a staged draft, a private comment, a ticket assigned to yourself — run and are logged so you can undo them. Five action classes are always held for your confirmation and never auto-fired: anything sent to an outside party, anything that overwrites a trusted record (CRM stage or forecast, closing or 'done'-marking work), commitments on someone else's behalf, numbers put in front of leadership, and bulk or destructive operations. Uncited claims, low confidence or conflicting sources are blocked or escalated with the conflict shown. Every action is written to an append-only ledger, and it never acts outside your own connector permissions. The research anchor: Gartner expects more than 40% of agentic-AI projects to be cancelled by the end of 2027, often on policy violations and cascading errors — the gate is the countermeasure.
It uses the same tool-agnostic connector model as enterprise-search — it works with whatever you have connected in each category: chat (Slack), email (Gmail), project tracker (Linear, Jira, Asana, monday.com, ClickUp), knowledge base (Notion, Confluence), customer support (Intercom), CRM (Salesforce), meeting notes (Fireflies) and analytics (Amplitude). The more sources you connect, the more of the loop it can close. Requirements: Cowork or Claude Code, and at least one connected MCP source.
No — and we are deliberate about saying so. The propose-approve-execute pattern is table stakes among the serious enterprise players: Glean ships propose-preview-approve plus an alignment check across 85+ cross-vendor actions, and Dust ships explicit read/propose/execute separation. The gate is a shared trust bar, well-built and necessary, but not a moat. What is genuinely differentiated is distribution and form-factor: nobody else ships cross-vendor acting at near-zero adoption friction, riding the Cowork seats and MCP connectors you have already authorized, built on Anthropic's own retrieval spine. Heavyweight platforms (Glean, Moveworks, Copilot Studio) need their own index, seat and admin rollout; single-vendor agents (Notion-in-Notion, Atlassian Rovo, Salesforce Agentforce, Sierra) act inside one product. Enterprise Actions is a plugin you install in minutes that acts across your whole stack.
Automated (reversible and confident, fired and logged): sorting, labelling, prioritising and routing your queue; snoozing FYIs; staging internal reply drafts; creating tickets assigned to yourself; adding private comments; detecting chat-vs-tracker drift; drafting recaps, nudges and rollups. Human confirmation required (the five hard-gated classes): sending anything to an outside party; overwriting a trusted record such as a CRM stage or forecast, or closing or 'done'-marking work; committing someone else to a task or due date; putting numbers in front of leadership in a rollup or status; and any bulk or destructive operation. Anything uncited, low-confidence or built on conflicting sources is blocked or escalated rather than guessed.
In Claude Code or Cowork: claude plugins add knowledge-work-plugins/enterprise-actions. It installs the six acting agents and their commands — /triage, /draft-reply, /meeting-to-work, /board-sync, /rollup and /crm-log — plus the retained Anthropic retrieval skills and the /search and /digest commands verbatim, the deterministic action-safety gate (lib/action_safety.py and the action-gate hook), and the append-only action ledger. Connect your tools through their MCP servers — each category has a connector slot. Without connectors, the safety gate and the loop still run so you can see the gating behaviour before you wire in live writes.
Because /digest is /rollup's passive baseline, the line has to be crisp. /digest summarizes the week for you and stops — it renders a summary and mutates nothing. /rollup synthesizes the same week across tools and metrics, drafts it in the audience's register (the way your VP or your customers like to read it), and delivers it — it posts to the channel or sends the email. The terminal post or send is the entire difference, and it is non-optional: a drafted-but-undelivered rollup is treated as an incomplete run, not a feature. And because /rollup can put numbers in front of leadership, every metric and every wide send is held for your confirmation.
Yes. Enterprise Actions is free and open source at github.com/mlobo2012/enterprise-actions. You need an active Claude Code or Cowork subscription and your existing tool connectors at their own pricing — Slack, Gmail, Linear, Notion, Salesforce and so on. There is no charge for the plugin itself. AI Heroes offers paid implementation: connector wiring, gate calibration for your risk tolerance, action-ledger and audit setup, and a crawl-walk-run rollout for teams who want a managed deployment rather than self-serve. Book a call to discuss.
Sources & Research
Every claim, statistic and competitor characterization on this page is grounded in a verifiable source — cited inline throughout and listed in full below. Where a figure appears in the copy, it traces to one of these. Last verified 2026-05-24.
Anthropic — enterprise-search plugin (source)
The official enterprise-search plugin source on GitHub. Verified directly: every command (/search, /digest), every skill (search-strategy, knowledge-synthesis, source-management) and every example is retrieval and synthesis with no write or send path. This is the foundation Enterprise Actions builds on and the basis for the 'they retrieve; we close the loop' boundary.
View source ↗Anthropic — Use enterprise search (support docs)
Anthropic's support article describing enterprise-search's scope: it retrieves across connected tools and ends at a single coherent answer with source attribution. Corroborates that the out-of-the-box product is read-only.
View source ↗Asana — Anatomy of Work (58–60% on 'work about work')
Asana's Anatomy of Work index finds workers spend 58–60% of the day on 'work about work' — communicating about work, chasing status and switching between tools — versus roughly a third on the skilled work they were hired for. This is the core motivation: search tools attack the small slice; the acting work is the large one.
View source ↗McKinsey — The social economy (28% email + 14% comms vs 19% search)
McKinsey's interaction-worker time split: about 28% of the workweek on email and 14% on internal communication, versus only 19% searching and gathering information. Search tools target the 19%; Enterprise Actions targets the 42%+ of triaging, drafting and coordinating.
View source ↗Microsoft — Breaking down the infinite workday (~60% on communication)
Microsoft's 2025 Work Trend Index finds communication eats roughly 60% of the day — about 117 emails and 153 messages daily — with interruptions every two minutes. The volume that makes manual triage and drafting unsustainable.
View source ↗Fellow — State of Meetings 2024 (54% leave unsure of next steps)
Fellow's State of Meetings report finds 54% of meetings leave attendees unsure of the next steps or who owns what. The direct justification for /meeting-to-work turning decisions and action items into filed tickets with owners and a posted recap.
View source ↗Salesforce — Sales research 2023 (68% call CRM entry their #1 time sink; only 2% trust it)
Salesforce's sales research finds 68% of sellers call CRM data entry their number-one time sink, and only 2% fully trust their CRM data. The pain /crm-log is built to remove by logging interactions and updating records — gated on stage and forecast overwrites.
View source ↗arXiv:2504.11443 — Where AI assistance pays off (email, drafting, support)
Research on where AI assistance measurably pays off: substantially less time on email, faster arrival at a good first draft, and more support tickets resolved per hour. The payoff sits precisely in the acting slice Enterprise Actions automates, not in search.
View source ↗Microsoft — Copilot agent builder (own seat / add-on pricing)
Microsoft's documentation for the Copilot agent builder. Used to characterize the friction of the heavyweight-platform alternative — separate seats and an action add-on — against which Enterprise Actions' ride-your-existing-seats model is the contrast.
View source ↗SalesforceBen — Why Agentforce adoption is slower than expected
Analysis of why Salesforce Agentforce adoption has lagged, including opaque consumption pricing and rollout friction. Supports the point that single-vendor acting carries real adoption cost relative to a plugin on connectors you already have.
View source ↗Glean — How agents work + 85 new actions (propose-approve-execute, cross-vendor)
Glean's agent documentation and its announcement of 85+ cross-vendor actions. Verified that Glean ships the propose-preview-approve pattern plus an alignment check before write actions. This is why the page states plainly that the safety gate is a shared trust bar, not a unique moat.
View source ↗Dust — Product + $40M funding (read/propose/execute separation)
Dust's product and funding coverage. Dust ships an explicit read/propose/execute separation across Drive, Slack, Notion, Gmail, Zendesk and HubSpot — further evidence that gated cross-vendor acting already exists, so the differentiation is distribution and form-factor.
View source ↗Sierra — Product (single-vendor CX agent)
Sierra's product page. Cited as an example of a single-vendor acting agent that operates inside one customer-experience surface, in contrast to cross-stack acting.
View source ↗ServiceNow — Moveworks acquisition (heavyweight platform)
ServiceNow's completion of the Moveworks acquisition. Used to characterize the heavyweight-platform category — own index, own seats, admin program — that Enterprise Actions is deliberately not.
View source ↗Google — Workspace Studio (single-vendor acting)
Coverage of Google Workspace Studio. Another single-vendor acting surface, cited in the competitive landscape that the cross-vendor wedge contrasts against.
View source ↗Notion — 2025 release (Notion-in-Notion agents)
Notion's 2025 release notes for in-product agents. Example of single-vendor acting: powerful inside Notion, but it does not turn a Slack thread into a Linear ticket into a customer reply into a CRM update.
View source ↗Atlassian — Rovo features (single-vendor acting)
Atlassian Rovo's feature set. Acts across the Atlassian estate, cited as single-vendor in the differentiation table against cross-stack acting.
View source ↗ServiceNow — Newsroom (platform acting)
ServiceNow newsroom releases on its agentic platform direction. Part of the heavyweight-platform evidence base for the friction contrast.
View source ↗Salesforce — Agentforce 3 announcement (single-vendor acting)
Salesforce's Agentforce 3 announcement. Cited as the single-vendor CRM acting incumbent that Enterprise Actions is cross-vendor against.
View source ↗TechCrunch — Anthropic brings agentic plugins to Cowork
TechCrunch's coverage of Anthropic bringing agentic plugins to Cowork. The distribution surface the wedge rides: Cowork seats and MCP connectors the buyer already has.
View source ↗Gartner — >40% of agentic-AI projects cancelled by 2027
Gartner's projection that more than 40% of agentic-AI projects will be cancelled by the end of 2027, frequently on policy violations and cascading errors where one bad assumption fans out into many automated downstream actions. The verified headline figure that motivates the deterministic action-safety gate.
View source ↗Salesforce DevOps — Agent observability GA (audit / ledger pattern)
Coverage of Salesforce making agent observability generally available, extending the agentic SDLC. The observability and audit pattern serious vendors ship because agents fail in subtle ways — mirrored by Enterprise Actions' append-only action ledger.
View source ↗Stop Doing the Work After the Answer.
One command to install. Run /triage on a real queue and watch it sort, own and route — with every risky action held at the gate. Free, open source, and built on Anthropic's enterprise-search plugin.
Want This Running for Your Team — Without the Setup Overhead?
AI Heroes configures the full stack: connector wiring across your tools, action-safety gate calibration to your risk tolerance, action-ledger and audit setup, and a crawl-walk-run rollout. Your team closes the loop from week one instead of spending it on integration.