AI employees that understand your business and get work done.

Hire specialized agents, connect your company tools, and decide exactly how they work. Every action stays supervised, explainable, and on the record.

Live in minutes Human approval on every action Full audit trail
app.syntrum.io/dashboard
Syntrum workspace dashboard showing Alex's team activity, approvals, outcomes, and daily insights
1Observe2Understand3Decide4Act5Report
Inside an agent

Built like a hire, not a chatbot

Every Syntrum agent is a defined role, not a prompt. It carries its own identity, judgment rules, operating procedure, tunable settings, tools, and lasting memory, so you decide how it thinks and what it is allowed to touch, then edit it the way you would brief a new hire. Once it is live it watches the channels you connect and acts the moment work arrives, not only when you prompt it. Every event becomes a durable task with its own lease and attempt count, so even if a run is interrupted the work picks up again from a consistent state.

agents / atlas
Syntrum agent detail view for Atlas showing its role, current action, and activity
what the agent learned
What the agent learned: observed patterns, rules you taught, and system defaults with confidence scores
Context, not retrieval

Linked memory, so the right context shows up at the moment of action.

Syntrum does not rank lookalike messages by similarity. It connects who a person is, which thread this belongs to, what was decided and who decided it, and which document is current, then assembles that exact context the moment an agent acts.

Company knowledge stays indexed and linked, and the corrections you make to a draft become durable rules the agent keeps and applies to future work.

Human in the loop

You stay in the loop. Not in the way.

Consequential actions wait for your approval. Missing context waits for your answer. Nothing waits silently.

Approve the action, not the intention

Most tools ask permission, then let the agent decide what to do with it. Syntrum attaches the exact call: destination, recipient, content, every parameter. You see precisely what will happen. Approving executes it verbatim, and the agent never gets a second interpretation.

approvals
Approval · #INV-2214HIGH

Pay invoice — Brightpath Media · $4,820 · 50 min ago

New vendor above the $2,500 auto-pay limit — verify banking details before approving.
FieldValue
ACTIONQUICKBOOKS_SEND_PAYMENT
VENDORBrightpath Media
METHODACH
AMOUNT$4,820.00
APPROVE REJECT

Approving executes this verbatim · No second interpretation

When it's blocked, it asks

No guessing, no silent failure. The agent asks a real question where the work is, in the Slack thread or on the email chain, and parks until answered. Then it resumes at exactly the step it stopped on.

inbox / thread
Thread · Checkout latencyConfidence 60%
TimeFromMessageState
09:12YOUCan you take a look and tell me what you recommend?
09:18PULSEOn it — pulled the relevant records, here's what I found…
09:24PULSERecommendation ready with reasoning. Want me to proceed?ASK
09:30YOUYes, go ahead — but flag anything above policy.
09:36PULSEDone. Safe steps complete; one escalated for approval.HELD
RESUMED AT STEP 4 / 6NOTHING WAITS SILENTLY
Where it works

It lives in the tools you already use.

Any tool your team uses today, your agent can use too. Credentials are scoped per agent, not per workspace, so an agent sees only what a colleague in that role would see, and nothing more.

SlackSlack
GmailGmail
GitHubGitHub
GGoogle DriveGoogle Drive
CalendarCalendar
HubSpotHubSpot
SalesforceSalesforce
JiraJira
LLinearLinear
NotionNotion
Under the hood

Any frontier model. Or your own keys.

Choose the model your AI employees run on: Claude Fable 5, GPT 5.6, Google Gemini Pro, Kimi K3, Grok 4.5, DeepSeek V4, GLM 5.2. A reasoning model when the work needs judgment, a faster one when it needs volume.

Bring your own API keys and inference runs on your account, under your provider terms, with no data passing through us. Switch providers without rewriting a single agent.

models
Claude Fable 5Default
Claude Fable 5AnthropicDefault
GPT 5.6OpenAI
Google Gemini ProGoogle AI
Kimi K3Moonshot AI
Grok 4.5xAI
DeepSeek V4DeepSeek
GLM 5.2Z.ai
Governance and audit

Set the floor once. See everything that happens on it.

One workspace policy decides what agents may do without asking. One immutable log records every action they took, every decision you made, and why each one was allowed.

A workspace floor that agents cannot lower

Pick an independence level and every agent inherits it: ask before every external action, act on routine and ask on risk, or act unless something is genuinely dangerous. Payments and password resets stay blocked at every level. Policies are enforced server-side on every action, not suggested in a prompt, and changes hot-sync to running agents within seconds.

Fine-tune anything underneath: which actions need approval, which are sensitive, what agents are allowed to learn, and which safeguards trip automatically.

app.syntrum.io/safety
IndependenceFloor
STRICTASK EVERY ACTION
BALANCED✓ DEFAULTASK ON RISK
AUTONOMOUSACT UNLESS DANGEROUS
RuleState
Routine repliesAUTO
New-contact outreachHOLD
Destructive operationsHOLD
Payments · password resetsBLOCKED

Enforced server-side · Block never loosens

Every action, decision and change, permanently recorded

Nothing an agent does is off the record. Filter by AI actions, human decisions, learning events or security events, search by action, target or agent, and export the whole thing to CSV for your auditors.

Each entry carries the agent, the integration, the outcome, a risk rating, and the reason the action was permitted, so you can reconstruct any task months later without asking anyone what happened.

app.syntrum.io/audit
Audit log LiveImmutable · CSV export
TimeAgentActionRisk
09:41PULSERollback recommendation → CheckoutHIGH
09:42PRIYAApproved refund → Order #48120LOW
09:44SENTINELAdmin access grant → GitHub orgCRIT
09:47IRISIssued refund → Acme CorpLOW
09:52MUSEDraft published → Help centreLOW
10 EVENTS TODAY0 UNEXPLAINED
Immutable logCSV exportRisk rating on every eventServer-side enforcementPer-workspace isolationCredentials scoped per agentSOC 2 Type IIGDPRNo training on your dataSelf-hosted available
Role templates

Every role, a complete pipeline.

Not a prompt library. Each template is a working agent with a mission, an operating procedure, scoped tool access, and a definition of done it has to prove.

Meeting booked

CRM record + prep brief

Verified

Sales Manager

Turn a raw lead into a qualified meeting on the AE's calendar.

findenrichqualifydraftsendfollow uprepliesbookrecord
findenrichqualifydraftsendfollow uprepliesbookrecord
  • Enriches and scores every inbound lead against your ICP within minutes of arrival.
  • Runs the full cadence and stops the moment a reply lands.
  • Negotiates a time by email, sends the invite, and briefs the AE before the call.

Done means: a meeting on the calendar, a logged CRM record, and a prep brief, all three verified to exist.

Delivery report

Every line has an artifact

Verified

Project Manager

Turn a plan into tracked work, keep it moving, and close it out honestly.

ingest planset uptrack dailyunblockmanage changereportclose out
ingest planset uptrack dailyunblockmanage changereportclose out
  • Builds the work breakdown and dependency graph from your kickoff doc.
  • Flags tickets marked done with no linked PR behind them.
  • Drafts the re-plan when scope moves, for the project lead to approve.

Done means: a delivery report where every shipped line is backed by an artifact, and no orphaned work.

Ticket resolved

Confirmed by requester

Verified

Support Manager

Own the queue end to end and make the knowledge base smarter with every ticket.

capturetriageresolveaskescalatechaseconfirmlearn
capturetriageresolveaskescalatechaseconfirmlearn
  • Answers in thread from past resolutions, citing the source.
  • Escalates hard cases with a full brief of what was already tried.
  • Turns recurring questions into knowledge base articles.

Done means: every ticket closed on a confirmed resolution, never on the agent's say so.

Month closed

Board pack traced to ledger

Verified

Finance Manager

Run the monthly cycle from raw ledger to distributed board pack.

kick off closereconcilechaseverifyanalyzewatch cashassembledistribute
kick off closereconcilechaseverifyanalyzewatch cashassembledistribute
  • Works the close checklist in dependency order, P&L blockers first.
  • Asks each budget owner to explain variances and remembers the answer.
  • Alerts you when runway or a covenant threshold is at risk.

Done means: a closed month where every item has its artifact and a board pack whose numbers trace to the ledger.

Payables paid

AR aging matches ledger

Verified

Accounting Manager

Own both sides of the money pipeline, invoice in to cash collected.

capturematchschedule paymentissuedunapplyverifyreport
capturematchschedule paymentissuedunapplyverifyreport
  • Extracts vendor, amount, PO and due date from invoices in any format.
  • Books clean three way matches and sends mismatches to you with the discrepancy named.
  • Runs the dunning sequence and matches incoming payments to open invoices.

Done means: payables paid on schedule, receivables collected or actively chased, aging that matches the ledger.

Campaign launched

Report matches source

Verified

Marketing Manager

Take a campaign from a one line brief to launched, measured and reported.

briefproducereviseassemblelaunchnurturemeasurereport
briefproducereviseassemblelaunchnurturemeasurereport
  • Drafts every asset in the brand voice learned from your past edits.
  • Builds the launch as a dependency chain and holds the send for approval.
  • Routes hot responders to the sales agent with full context.

Done means: a launched campaign, routed leads, and a results report whose numbers match the source systems.

Hire live in HRIS

Accounts + equipment + sign-offs

Verified

HR Manager

Run recruiting and onboarding as one pipeline, headcount approved to productive hire.

openscreenschedulechase feedbackadvanceofferonboardconfirm
openscreenschedulechase feedbackadvanceofferonboardconfirm
  • Runs all interview logistics including reschedules and panel calendars.
  • Chases overdue scorecards and surfaces candidates before they go cold.
  • Verifies onboarding against the real systems, not a checkbox.

Done means: a hire in the HRIS with accounts live, equipment delivered and sign offs on file.

Renewal on file

Health score current

Verified

Customer Success Manager

Own the post sale lifecycle, activation through renewal.

kick offactivatewatch healthintervenegrowrun QBRsrenewrecord
kick offactivatewatch healthintervenegrowrun QBRsrenewrecord
  • Takes the closed won handoff with full deal context and books the kickoff.
  • Opens a save play on usage decline or a departed champion.
  • Starts the renewal motion 90 days out with a usage backed value narrative.

Done means: a renewed contract on file, a current health score, and a CRM that tells the whole true story.

PO received, vendor paid

Matched invoice on file

Verified

Operations Manager

Run procurement from a Slack request to a delivered order and a paid vendor.

intakesourceapproveordertrackreceivepayfile
intakesourceapproveordertrackreceivepayfile
  • Checks every request against policy and budget before sourcing.
  • Assembles a vendor quote comparison, or reuses the approved vendor from memory.
  • Chases shipping slippage and keeps the requester informed unprompted.

Done means: goods received, vendor paid against a matched invoice, every document filed with renewals tracked.

Calendar defended

Every commitment chased

Verified

Executive Assistant

Run the operational layer around an executive, inbox to follow through.

triagedefend calendarpreparecapturechasebrief dailydigest weeklycoordinate
triagedefend calendarpreparecapturechasebrief dailydigest weeklycoordinate
  • Surfaces only what needs you and routes the rest to the right specialist agent.
  • Builds a one page brief before every meeting with verified status on last time's actions.
  • Chases every commitment in both directions.

Done means: a defended calendar, briefed meetings, and no commitment falling through.

Your definition of done

You decide what it has to prove

Build your own

Start from a blank agent and shape it into any role your team runs.

  • Write the identity, judgment rules, and operating procedure.
  • Scope exactly which tools it can touch.
  • Define the artifact it has to produce to be done.

Done means: whatever you say it does, verified the same way every template is.

Or write your own. Every template is just files you can edit: identity, judgment rules, operating procedure, config, tools, memory. Build a custom agent

Questions

What teams ask before they start.

A clear starting point for bringing your first AI employee into the team.

Start small

Put one agent to work this week.

Start with one repeatable job. Keep approval over every action, see the results, and expand when you are ready.