The platform, layer by layer.

Artifi is ten layers that work as one: the conversation you work in, the protocol that connects it, the workflows that govern every write, the tasks where people decide, a ledger that checks itself, the panel where the team works, the agents that do the volume work, the email channel they work from, the integrations that bring your systems in, and the access model that scopes all of it.

Each layer below is shown with a real screen from that layer, taken in the demo organisation.

Layer 1 of 10Claude

You say what you need.

Work starts where you already are: Claude.ai, Claude Desktop, Excel, or an IDE. You describe the job in plain language. Claude reads your financial context, prepares the work, and shows you what it is about to do before anything is written.

  • No forms to learn; the conversation is the interface
  • Results come back as tables, reports and Artifacts you can keep
Claude · one command runs the VAT pre-close gate and prepares the return; the files are ready in the sidebar

Layer 2 of 10MCP

The connection is a standard, not an integration.

Artifi is an MCP server. Any client that speaks the Model Context Protocol connects the same way and receives the same tools and finance skills: Claude, Cursor, Windsurf, VS Code, or your own. Access is by OAuth or an API key, and every call carries the identity and entity scope of the person or agent behind it.

  • Two transports, one tool registry, one permission model
  • Skills ship to IDE clients as standard skill files
Claude · Artifi as a standard connector: its tools, and the permission you give each one

Layer 3 of 10Workflows

Behind every write there is a definition.

Nothing posts directly. Every operation — posting a bill, creating a vendor, releasing a payment — goes through a workflow definition that names its risk, its steps and who has to approve. Low-risk work runs straight through; anything unusual waits for the reviewer your policy requires. An operation without a definition is refused, never executed.

  • Green lane: auto-approved · Yellow: one approver · Red: several
  • Definitions are cloned from a template and edited per organisation
Admin app · 312 workflow definitions by category, each with its risk level and steps

Layer 4 of 10Tasks

When a decision needs a person, it becomes a task.

A workflow step that needs judgement, an agent that cannot resolve something on its own, an unusual amount your policy flags: each arrives in one queue, addressed to the right person. A task shows the proposed record exactly as it will be written, the workflow, step and lane it came through, and one decision to make. When an agent asked the question, your answer goes back to the agent and it acts on it — so a review is never just a comment left somewhere.

  • Workflow approvals and agent questions in the same inbox, one row per task
  • Every decision recorded with who took it, when, and any comment
Admin app · one task: the proposed vendor record, its workflow and lane, and the reviewer's decision

Layer 5 of 10Ledger

The database checks itself.

Every financial event — a vendor bill, a customer invoice, a payment, a journal, a payroll run — is a typed transaction in one ledger, with balanced journal lines. The ledger enforces its own rules at commit: debits equal credits, posted history cannot be edited, closed periods stay sealed, document numbers are gapless, and every change leaves an activity row written by the database itself, attributed to the person, agent or connector that made it.

  • 61 transaction types, one table, one schema per organisation
  • A nightly self-check re-verifies every organisation and mails its admins
Admin app · a posted transaction with its journal, balanced, source: agent

Layer 6 of 10Admin panel

Where people review, approve and configure.

The admin panel has no entry forms. It is where your team sees what is waiting on them, approves or returns it, reads the audit trail, and configures the rules everything else follows: workflows, roles, entities, connectors, agents.

  • Every submission listed with who, when, and which lane it took
  • Tasks, approvals and agent requests in one queue
Admin app · the ledger, with every record type one menu away

Layer 7 of 10Agents

Workers with job descriptions.

Agents do the volume work: reading bills from email, classifying bank lines, building payment batches, chasing missing documents. Each one runs in one of three modes — a language model, deterministic code, or a hybrid pipeline of both — and writes through the same gateway and lanes as a person. They hand work to each other, and what they cannot resolve becomes a task for the right person.

  • Describe a new agent to Claude in plain language; it gets its own triggers, steps and limits
  • Every run is recorded with its steps, duration and model cost
Admin app · the agent roster: mode, origin, runs
One run · an emailed invoice posted, its receipt correctly left alone: 29s, two steps, $0.0151

Layer 8 of 10Communications

Email is a channel, not an inbox to empty.

Every entity gets its own addresses for bills, invoices and receipts. Whatever arrives is classified before any agent sees it: spam is quarantined rather than dropped, uncertain messages are flagged for a person, and the rest becomes work for the right agent, with attachments stored and read on demand. Outbound mail — bill chases, invoice deliveries, receipt requests, payment confirmations — lands in the same view, so both directions of every conversation are on one auditable record.

  • Addresses per entity and purpose, or your own domain once it is verified
  • Every message is one row with its classification, status and outcome
Admin app · the communications view: one folder per purpose, one row per message with its outcome

Layer 9 of 10Integrations

Three ways in.

Native connectors bring in bank statements, card activity, accounting data and documents from the systems we support directly. Beyond those, anything with an MCP server can be reached the same way you reach Artifi — through Claude or through the platform — so your ERP, billing tool or data warehouse can take part in a workflow without a bespoke build. Custom connectors, for systems with neither, are in design.

  • Credentials encrypted at rest; agents receive scoped, expiring access
  • Connected systems keep their own permissions and boundaries
Admin app · adding a connection in six steps; nothing syncs until the test passes

Layer 10 of 10Users and roles

Two layers of access.

Membership decides what a person can do in the admin panel. Permission roles decide which accounting operations they can perform, through Claude or anywhere else, and both are scoped to the legal entities they may touch. Agents and API keys are subject to the same model, so a UK billing agent never sees the EU payroll ledger.

  • System roles out of the box, custom roles from the same permission groups
  • Entity scoping enforced on reads as well as writes
Admin app · roles and the permissions each one carries

Go deeper on any layer:

Your processes, your rules

See the platform run
on your own books.

Bring one real process. In a working session we run it through every layer with your approvers and your thresholds.

hello@ar-ti-fi.com