The Second Brain Agent. Every cite checkable.

Generic LLMs can't do this.

What it actually does.

Real work the agent does end-to-end. Every step is auditable, every consent gate logged, every escalation routed to a named human.

1. 01

   New-hire onboarding desk

   Day-one expense, payroll, and policy Q&A without pinging a colleague.

   • Hire signs in with SSO on day one.
   • Asks: "How do I file expenses for a client dinner in Mumbai?"
   • Agent pulls the India expense policy, the Concur SOP, and the per-diem table.
   • Answer cites the exact paragraph in each doc with the section header.
   • If the hire's role has no Concur access, the agent flags that and routes to IT.
   Permission boundaryHRIS role and country attributes filter India vs US policy versions.

   Human in the loopCitation requirement: every claim links to a paragraph anchor. No anchor, no answer.
2. 02

   Contract Q&A for legal and procurement

   Find renewal, indemnity, and auto-renew terms across thousands of MSAs in seconds.

   • Legal counsel asks: "What is our auto-renewal stance with Vendor X?"
   • Agent retrieves the active MSA, the last amendment, and the email confirmations of waiver.
   • Returns the clause text, effective date, and the contradicting amendment.
   • Flags conflict between two sources and shows both with dates.
   • Offers to draft a renewal-decline note from the template library.
   Permission boundaryLegal-only DealRoom folders stay invisible to non-counsel users.

   Human in the loopCounsel decides the answer. The agent surfaces the evidence.
3. 03

   Code search across the monorepo

   Show me where the auth middleware lives and what it depends on.

   • Engineer asks: "Where does the JWT refresh path break for SSO users?"
   • Agent pulls files, recent PRs, the on-call runbook, and the last incident postmortem.
   • Returns the function with line range, the upstream config, and the related Jira.
   • Highlights a stale doc in the wiki that contradicts the current code.
   • Links to the test that covers the change.
   Permission boundaryGitHub repo membership and Jira project ACLs propagate at query time.

   Human in the loopThe engineer still writes the fix. The agent shows where to look.
4. 04

   IT helpdesk deflection

   VPN, MFA, and laptop-imaging questions answered before a ticket opens.

   • Employee types the question into the Slack agent.
   • Agent matches against the IT runbook and the last 90 days of solved tickets.
   • Returns the steps with a one-click "still broken" escalation.
   • If the runbook is older than the relevant config change, agent says so.
   • Logs the deflected ticket against the IT KPI dashboard.
   Permission boundaryReads runbooks the user's group already has Confluence access to.

   Human in the loopIT handles the escalations the agent could not close.
5. 05

   Sales enablement collateral lookup

   Pull the right one-pager, battlecard, and reference customer in 10 seconds.

   • AE asks: "What's our best reference for a UK retail bank evaluating us against Vendor Y?"
   • Agent matches industry, geo, and competitor tags across Highspot and Drive.
   • Returns the case study, the battlecard, and the latest pricing memo.
   • Cites the win-loss note that supports the competitor framing.
   • Refuses to quote NDA references and shows why.
   Permission boundarySalesforce sharing rules and Highspot groups gate every asset.

   Human in the loopThe AE makes the customer call. The agent prepares the kit.
6. 06

   Policy lookup for managers

   PTO carryover, parental leave, and travel rules per country and grade.

   • Manager asks: "How much PTO can my Singapore L4 carry into next year?"
   • Agent pulls the Singapore policy, the grade matrix, and the HRIS rule.
   • Returns the number with the policy clause and the HRIS field that proves it.
   • Notes the policy was updated 11 days ago and links to the changelog.
   • Refuses to answer if the user is not the employee's line manager.
   Permission boundaryWorkday manager-of relationship gates the query.

   Human in the loopHR owns policy. The agent surfaces it, with the version date.
7. 07

   Internal product FAQ

   Why was this feature built this way, and what was the trade-off.

   • PM asks: "Why did we deprecate the v2 webhook?"
   • Agent pulls the RFC, the deprecation memo, and the customer-impact email.
   • Returns the decision rationale with the names of the deciders and the date.
   • Surfaces the v3 migration guide and known caveats.
   • Shows the two customer tickets that triggered the change.
   Permission boundaryConfluence space and Notion teamspace ACLs.

   Human in the loopThe PM still decides what to build next.

Phase by phase.

The end-to-end path the agent runs. Every phase logs what it captured, what consent applies, and where a human gates the next step.

1. 01

   Connector onboarded

   For each source: SharePoint, Drive, Confluence, GitHub, Jira, Salesforce, NetSuite, Zendesk, Slack.

2. 02

   ACL graph extracted

   Group nesting, inheritance breaks, and per-document overrides imported from each source.

3. 03

   Chunking

   Documents chunked at paragraph and heading boundaries. Code chunked at function level.

4. 04

   Local embeddings

   Vectors generated and stored on the client server or in-region node. Nothing leaves the boundary.

5. 05

   Sensitive-field pass

   PII tagging and redaction. Restricted classes flagged for narrower retrieval.

6. 06

   Eval calibration

   First eval set runs against 200 golden questions per department to score grounding and refusal.

7. 07

   Rollout

   Behind a Slack and web client, with audit logging on every query and citation click.

8. 08

   Reindex and drift watch

   Per-source reindex schedule, freshness scores, weekly eval regression report.

The non-negotiables, spelled out.

Consent, security, accuracy, and residency. Explicit, auditable, and the same line items your CISO or GC already asks about.

What the numbers look like.

Figures we can point at. Every one carries a source you can verify.

What it replaces. What it augments.

What the Second Brain does

It learns your business, answers from it with citations, and stays inside the boundary you set.

The numbers, up front.

Per-headcount tier pricing matches where the value lands. The brain pays for itself in time recovered per employee, so a $4,500/month seat at 1,000 staff is $4.50 per head, well under the 2 to 4 hours per knowledge worker per week that platforms in this category report saving. The tiered cap protects mid-market budgets while letting enterprises with global custom needs negotiate residency, connector load, and SLA separately.