Business logic was never the hard part.

Running it inside a real company is: your systems, your data, your rules, every day. This page is how Vermoon does it, part by part.

Database
EU, Ireland
Encryption
TLS in transit, AES-256 at rest
Models
Claude, GPT, Gemini, Grok and more
Compliance
GDPR and EU AI Act programme

Connects to what you already run.

The tools your team uses every day connect with one login, and each person grants only the permissions they choose. Anything with no connector, from your ERP and your SQL database to a carrier portal or bank files, we connect for you as part of the build.

One login
Gmail, Outlook, Drive, WhatsApp, Instagram, Facebook, LinkedIn, X
Connected by us
ERP, SQL databases, carrier portals, bank files (Norma 43)
Permissions
Per person, per account, revocable at any time

Rules where it must be exact.

Ledgers, tax, matching and pricing run on deterministic rules: same input, same answer, every time, and never a model guessing a number. Agents take the parts that need judgment and check the result of the rule before it goes anywhere.

Invoice 2026-0412, booked by rule
430 Customers€1,210.00 debit
700 Sales€1,000.00 credit
477 VAT charged€210.00 credit
Debits equal creditsBalanced
Agent checked: it matches the delivery note
Runs on rules
Booking, tax, matching, costing, tracking
Runs on judgment
Reading, deciding, writing, the exceptions
Every posting
Balanced, recorded, reversible

A person stays in control.

Every step is classed by what it does: read, draft, send, spend, or something that cannot be undone. Reading and drafting can run alone. Sending waits for a person unless the company allows it. Spending and anything irreversible always wait.

Mode
Read the inboxAlone
Draft 14 repliesAlone
Send the repliesAsks you
Pay the carrierAsks you
Three modes
Manual, Supervised, Automatic: one for the company, one per person
Always waits
Spending, negotiating, anything irreversible
Trust
Earned per kind of task, reset by one rejection

Model-agnostic.

Claude, GPT, Gemini, Grok, Llama, Mistral, DeepSeek. Each step in a process runs on the model that does it best, chosen per company, per agent and per step. Moving a step to another model changes one setting, not the agent, and adding a provider is one adapter, so a better or cheaper model is adopted the week it appears.

Wired today
Anthropic, OpenAI, Google
Next
xAI, Meta, Mistral, DeepSeek: one adapter each
Chosen per
Company, agent, step
Processing
A request may be processed in the US by the provider
Read a scanned receipt
Negotiate with a supplier
Summarise the inbox
Sort 400 support tickets

Changes are tested before they reach you.

Every night the Doctor checks each step for cost, speed and failures. When it finds a better setting it does not just apply it: it tries it on past work, releases it to a few, watches it, and undoes it by itself if results drop. Nothing changes in your agent without passing that.

25past requests it is tried on first
20%of new requests get it next
7 dayswatched, then undone by itself if results drop
08:02Listed 14 late invoices€0.04
08:03Drafted 14 reminders€0.18
08:31Marta approved the 14 sends€0.00
08:31Sent 14 reminders€0.02

Every step is recorded.

What each agent did, when, on which data, who approved it and what it cost, readable by the people you choose. A task's cost is shown before it runs and after it ends, so nobody is surprised by the bill.

Per task
Agent, time, data touched, approver, cost
Per person
Own chats, own files, own connected accounts
Per month
Model and tool costs against the limit you set

Your data lives in the EU.

The production database runs in AWS eu-west-1, Ireland. Data is encrypted in transit with TLS and at rest with AES-256, and the tokens for connected accounts are encrypted per person. Model providers may process a request in the US; we say so rather than hide it.

DatabaseEU · Ireland
In transitTLS
At restAES-256
Connected accountsEncrypted, per person
Deletion on requestWithin 30 days
Database
AWS eu-west-1, Ireland
Encryption
TLS in transit, AES-256 at rest
Deletion
On request, within 30 days

GDPR and the EU AI Act.

Every agent is classified by risk under the AI Act. Recruiting is the only high-risk use, and there a person makes every decision. Content and support agents carry transparency duties; finance and operations are minimal risk. The programme runs with external counsel and is not yet certified.

RecruitingHigh risk · a person decides
Social contentTransparency
Support chatTransparency
Finance, logistics, the restMinimal risk
High risk
Recruiting: a person decides
Transparency
Social content, support chat
Counsel
External; programme in progress, not certified

One team with yours.

You are never handed software and left alone. We build it, run it and keep improving it.

Your team

  • Shows us how the work is done today
  • Tests the agent on real cases
  • Approves what matters, every day

Vermoon

  • Builds the agent and connects your systems
  • Runs it, watches it and fixes it
  • Reviews it with you and improves it every month

Put your first agent to work.

Tell us one process that eats your team's time. We show you the agent that does it.

Technology · Vermoon