# Upmo AI - Complete Product & Architecture Reference > Upmo AI gives your team dependable agents that work inside the software you already run. Every result is verified against the source, and anything that touches money, customers, or company records waits for your approval. Canonical Website: https://www.upmo.ai Repository: https://github.com/upmo-ai/upmo Company: Klassroom (KvK 95052917, Banning Cocqgracht 83, 1445 BW Purmerend, the Netherlands) Contacts: hello@upmo.ai | security@upmo.ai | privacy@upmo.ai --- ## Core Architecture ### Verification: Work is verified before it counts as done. Every task follows the same structured workflow: prepare, plan, work, verify. Upmo AI independently verifies the output against your source records, so a task only completes when the numbers and facts match. Highlights: Structured steps every time | Independent verification | Optional second reviewer ### Keys and approvals: Agents never hold your real credentials. Agents operate with stand-in placeholders. Upmo AI injects the real key only when a request goes to an approved service for that task. Anything that emails customers, moves money, or updates records waits for your sign-off. Highlights: Stand-in credentials | Approved services per task | Human sign-off before impact ### Models and costs: Predictable costs for every single task. Leading AI models are included, with no separate model accounts or API keys to manage. Pick the right model for each role, track exact cost per task, and set hard budgets that automatically pause spending. Switching models? Upmo AI benchmarks the new model on your past tasks first. Highlights: Models included | Exact cost per task | Hard budget limits ### Knowledge: Shared company knowledge with every source attached. Give agents your playbooks, house rules, policies, and past examples. They search across your team's documents, learn from corrections, and attach the original source and date to every fact so your team can inspect, edit, or delete it anytime. Highlights: Playbooks and SOPs | Sources and dates on every fact | Isolated per team --- ## Task Execution Lifecycle Every task moves through six consistent steps: 1. **Preparing**: Upmo AI opens a sealed workspace and connects only the tools required for the task. 2. **Planning**: The agent drafts an execution plan and reads relevant company knowledge and house rules. 3. **Working**: The agent performs work in connected tools while logging every read, change, and cost. 4. **Verifying**: Upmo AI independently verifies outputs against source records before completion. Optional second-reviewer auditing is supported. 5. **Needs your approval**: External sends, payments, or record changes pause and present exact payloads for human approval. 6. **Done**: Task completes with an output summary, full cost ledger, and tamper-evident audit log. --- ## Live Industry Workspaces - **All teams** (kade.upmo.ai): - **Finance agent** (Claude Sonnet 5.5, budget €0.60, sandbox: Exact Online · Outlook only): verifies "reminder amounts match Exact Online"; pauses for "Rule: ask before emailing customers" - **Support agent** (GPT-5 mini, budget €0.50, sandbox: Zendesk · Shopify · PostNL only): verifies "every reply has the current tracking status"; pauses for "Rule: ask before replying to customers" - **Sales agent** (Claude Haiku 5.5, budget €0.40, sandbox: HubSpot · Calendar · public web): verifies "every fact links to a source"; pauses for "Rule: ask before writing to HubSpot" - **Operations agent** (Mistral Large, budget €0.30, sandbox: Google Workspace · Slack only): verifies "no active sessions left"; pauses for "Rule: ask before suspending accounts" - **Contracts agent** (Claude Sonnet 5.5, budget €0.40, sandbox: Contracts folder, read only): verifies "dates match the signed PDFs"; pauses for "Rule: ask before posting in channels" - **Accounting firm** (jansen-accountants.upmo.ai): - **Bookkeeping agent** (Claude Sonnet 5.5, budget €1.50, sandbox: Exact Online · client inboxes): verifies "VAT and totals match each invoice"; pauses for "Rule: ask before booking entries" - **Client follow-up agent** (Mistral Small, budget €0.40, sandbox: Twinfield · Outlook only): verifies "each request lists only missing items"; pauses for "Rule: ask before emailing clients" - **VAT agent** (Claude Sonnet 5.5, budget €0.80, sandbox: Twinfield, read only): verifies "returns reconcile to the ledger"; pauses for "Rule: filing stays with your accountant" - **Law firm** (legal.upmo.ai): - **Contract review agent** (Claude Sonnet 5.5, budget €0.50, sandbox: Playbook · Word only): verifies "every flag cites a clause"; pauses for "Rule: ask before sending to the other party" - **Intake agent** (Mistral Small, local, budget local, sandbox: No internet · local model): verifies "every hit links to a matter"; pauses for "Rule: ask before sharing conflict results" - **Billing agent** (GPT-5, budget €0.60, sandbox: Time and billing only): verifies "hours match the timesheets"; pauses for "Rule: ask before creating invoices" - **Online store** (store.upmo.ai): - **Returns agent** (GPT-5 mini, budget €0.40, sandbox: Shopify · Gorgias · Sendcloud): verifies "every approval matches the policy"; pauses for "Rule: ask before sending return labels" - **Catalogue agent** (Claude Sonnet 5.5, budget €1.20, sandbox: Shopify · supplier sheets): verifies "every spec matches the supplier sheet"; pauses for "Rule: ask before publishing products" - **Stock agent** (Mistral Small, budget €0.30, sandbox: Shopify, read only · email): verifies "quantities match sales data"; pauses for "Rule: ask before ordering from suppliers" Additional quick hand-off workflows: - Summarise this week's customer calls (Call notes · Slack, budget €0.40) - Draft the monthly management report (Exact Online · HubSpot, budget €0.60) - Clean up duplicate contacts in HubSpot (HubSpot, budget €0.30) --- ## Ready-Made Marketplace Agents ### AC-01: Purchase invoice processing (Accounting) - Package: `upmo/purchase-invoices@1.0` - Budget: €1.50 per run - Connected tools: Exact Online, Twinfield, Outlook - Summary: Reads purchase invoices from your clients' inboxes, suggests the ledger account and VAT code, and books them in Exact Online or Twinfield. Anything it isn't sure about goes on your review list. - House rule: Never book an invoice if the supplier's bank account differs from the one on file. Flag it as possible fraud. - Verification rule: VAT and totals match each invoice - Guardrail: Rule: ask before booking entries - Steps: 01. Collect new invoices from client inboxes and upload folders 02. Read supplier, amounts, VAT and dates 03. Suggest the ledger account and VAT code from past bookings 04. Ask a bookkeeper before anything is booked ### LG-02: NDA and contract review (Legal) - Package: `upmo/contract-review@1.0` - Budget: €0.50 per run - Connected tools: Word, SharePoint, Your playbook - Summary: Checks incoming NDAs and supplier contracts against your own playbook, marks every deviation, and drafts a redline with your preferred wording. - House rule: Governing law must be Dutch, with disputes going to the court in Amsterdam. Flag anything else. - Verification rule: Every flag cites a clause and a playbook rule - Guardrail: Rule: ask before sending to the other party - Steps: 01. Read the contract and identify its type 02. Compare each clause with your playbook 03. Draft a tracked-changes redline and a short summary 04. Ask the responsible lawyer before sending ### RT-03: Returns desk (Online retail) - Package: `upmo/returns-desk@1.0` - Budget: €0.40 per run - Connected tools: Shopify, Gorgias, Sendcloud - Summary: Handles return requests in your helpdesk. It checks the order against your return policy, sends a label for eligible returns, and passes the rest to your team. - House rule: Returns after 30 days go to a person, even for regular customers. - Verification rule: Every approved return matches the policy - Guardrail: Rule: ask before sending return labels - Steps: 01. Read new return requests from the helpdesk 02. Check order date, item and reason against your policy 03. Prepare return labels for eligible orders 04. Ask before labels go out; pass exceptions to the team ### SU-04: First-line support replies (Customer support) - Package: `upmo/support-replies@1.0` - Budget: €0.50 per run - Connected tools: Zendesk, Intercom, Help centre - Summary: Reads new tickets, finds the answer in your help centre and past replies, and drafts a response. Urgent or unclear tickets go straight to a person. - House rule: Never promise a delivery or fix date the customer hasn't already been given in writing. - Verification rule: Every reply cites a help article or past reply - Guardrail: Rule: ask before replying to customers - Steps: 01. Read new tickets and sort out urgent ones 02. Find the answer in your help centre and past replies 03. Draft a reply in the customer's language 04. Ask before sending; hand unclear tickets to a person ### SA-05: Meeting prep briefs (Sales) - Package: `upmo/meeting-briefs@1.0` - Budget: €0.40 per run - Connected tools: HubSpot, Google Calendar, Company websites - Summary: Every morning, writes a one-page brief for each sales meeting in your calendar: who you're meeting, what was said last time, open deals and recent company news. - House rule: Only use news from the last 90 days, and link every fact to its source. - Verification rule: Every fact links to a source - Guardrail: Rule: ask before writing to HubSpot - Steps: 01. Read today's customer meetings from the calendar 02. Pull recent emails, open deals and company news 03. Write a one-page brief per meeting 04. Ask before adding briefs to HubSpot ### OP-06: Employee offboarding (Operations & HR) - Package: `upmo/offboarding@1.0` - Budget: €0.30 per run - Connected tools: Google Workspace, Microsoft 365, Slack - Summary: When someone leaves, it finds every account they have, moves their files to their manager, and suspends access on their last day. Nothing gets deleted. - House rule: Suspend, never delete. Keep accounts for 90 days in case files are needed. - Verification rule: No active sessions left - Guardrail: Rule: ask before suspending accounts - Steps: 01. Find every account and shared file the person owns 02. Move files and calendars to their manager 03. Schedule suspension for the last working day 04. Ask before any account is suspended --- ## Zero Trust Security Model ### Three Principles 01. **Verify every request**: Each action is verified against the task's rules and allowed destinations. Nothing passes just because an earlier request did. 02. **Grant the minimum access**: Each task receives only the specific tools it needs, a stand-in for your credentials, and a strict budget. 03. **Contain every risk**: Even if an agent reads a malicious document, the blast radius stops at the sealed workspace wall and a person decides. ### Built-In Security Controls - **Network (Agents only reach sites you allow)**: Each task runs in a sealed workspace with no internet access. It can only reach the tools on its own list. ([Read more](https://www.upmo.ai/zero-trust-network.md)) - **Keys (Agents never hold your keys)**: Agents work with stand-in keys. Upmo AI adds the real one on the way out, only for allowed sites. ([Read more](https://www.upmo.ai/zero-trust-keys.md)) - **Approvals (Risky actions wait for a person)**: Sending data out, moving money or changing records stops and asks. You see exactly what will happen. ([Read more](https://www.upmo.ai/zero-trust-approvals.md)) - **Data (Each team only sees its own data)**: Access rules are enforced by the database itself, so a bug in the app can't show one team another's records. ([Read more](https://www.upmo.ai/zero-trust-data.md)) - **Sign-in (Sign in with a passkey or SSO, and confirm again for sensitive changes)**: Use your fingerprint or face, or your company's SAML single sign-on. Changing roles, keys or security rules asks you to confirm again. ([Read more](https://www.upmo.ai/zero-trust-sign-in.md)) - **Records (Every action and every memory has a source)**: Each step is logged, and each fact an agent remembers shows where it came from. Correct or delete it at any time. ([Read more](https://www.upmo.ai/zero-trust-records.md)) Certification note: Upmo AI is not SOC 2 or ISO 27001 certified yet. The codebase is open source at https://github.com/upmo-ai/upmo so security teams can audit every control directly. --- ## Deployment Plans ### Cloud (upmo.ai) - Coming soon For teams that want agents working without running any servers. - We run and update everything for you - Models included, no separate AI account needed - SAML single sign-on (SSO) with your identity provider - Ready-made agents from the marketplace - Scheduled tasks, and agents that suggest work - Cost per task, team budgets and spend limits ### Enterprise (Cloud or your own servers) - Coming soon For companies that need their own servers, a private network, or no internet access at all. - Runs on your own servers or private network - Works without internet access, with local models - Audit logs sent to your own security tools - Access rules per team and per person - SAML single sign-on (SSO) - Help from our engineers during setup --- ## Frequently Asked Questions ### What happens if I deny a request? Nothing is sent or changed. The agent records your answer and skips that step, or stops the task if it can't continue without it. ### Can an agent go over budget? No. When an agent reaches its budget it pauses and asks you whether to continue. ### Which tools can I connect? Accounting, CRM, helpdesk, webshop, email and chat tools. If a tool has an API, Upmo AI can connect to it. ### When can I start using Upmo AI? We're still building it and it isn't open for sign-ups yet. Get in touch at hello@upmo.ai and we'll tell you when it is. ### Can my team sign in with our company accounts? Yes. Upmo AI supports SAML single sign-on through your identity provider, such as Okta, Microsoft Entra ID or Google Workspace. Passkeys work too. ### Is Upmo AI SOC 2 or ISO 27001 certified? Not yet. The code is open source, so your security team can read exactly how it works.