First published on LinkedIn in May 2026. Republished here with light edits.
Over the past months I've had the same conversation with dozens of CFOs and finance directors across the Nordics and Baltics. It goes roughly like this: they've watched Anthropic ship Skills, Plugins, Artifacts, Cowork, Scheduled Tasks, Managed Agents — a release cadence that's hard to keep up with even if it's your job — and they have a sense that something important is happening for finance work. What they don't have is a map. Which of these things do I actually use? For what? Where does the recurring work live? What can I put in front of an auditor?
The confusion is understandable, because the mental model most people carry — "Claude is a chatbot, and these are chatbot features" — is wrong. Claude today is not one product with one interface. It is four distinct surfaces, designed for four different kinds of work. And the one that matters most for running a finance function doesn't live in Claude.ai at all.
Here is the map.
Surface 1: Claude.ai Chat — where you supervise
The one everyone knows. A conversation: you type, Claude responds. You attach files, connect tools — Google Drive, Gmail, Slack, and through MCP, your ERP or ledger — and work through a problem together.
The defining property of Chat is that it runs only while you're in it. It's synchronous and supervised by construction. That makes it the right surface for exactly the work where you want a human in the loop on every step: investigating a variance, drafting a board commentary, asking "why did software spend jump 40% in March?", reviewing a reconciliation exception.
What Chat is not is an automation surface. Nothing happens when you close the tab. If your mental model of "using Claude for finance" is a person pasting things into a chat window, you've seen one quarter of the product — and the least leveraged quarter.
Surface 2: Claude.ai Cowork — where you delegate
Cowork is the autonomous mode inside Claude.ai. The interaction inverts: instead of driving turn by turn, you describe the outcome — "process this folder of supplier invoices and prepare the entries" — and Claude plans the steps, executes them, and comes back with results. You review the output, not every keystroke.
Cowork is also where Plugins are hosted and where Scheduled Tasks live — recurring jobs that re-run on a cadence. That's a real step toward automation, with one honest caveat that matters for anyone planning operations around it: the work runs in your session. A scheduled task is a delegated routine, not deployed infrastructure.
For finance, Cowork is the surface for multi-step work where a human stays nearby: a month of bank statements imported and classified in one pass, a first-draft FP&A scenario, a batch of expense reimbursements prepared for approval.
Surface 3: Plugins — the desk, set up before you arrive
Plugins are pre-configured bundles: Skills, connectors, and sub-agents packaged so a particular domain of work starts from posture instead of from zero.
The analogy I keep using: a Plugin is a desk setup, not a finished deliverable. When a new accountant joins your team, you give them a configured workstation — the chart of accounts, the posting conventions, access to the bank feeds, your firm's closing checklist. The desk doesn't do their job. It removes every piece of friction between them and the job.
That's what installing a finance Plugin does for Claude. It doesn't "automate your accounting" on day one — anyone promising that is selling something else. It gives Claude the domain posture: what a payroll run involves in your jurisdiction, how a VAT declaration is assembled, which accounts a depreciation entry touches. The work still happens in Chat, Cowork, or — the interesting part — on the fourth surface.
One property of well-built Plugins that finance people consistently underrate: they're readable. A Plugin's logic is plain text. Your accountant can open the payroll Plugin and audit the actual rules — income tax 22%, social tax 33%, unemployment insurance — the way they'd review a junior's working papers. Try doing that with the payroll logic inside a SaaS product.
Surface 4: Managed Agents — the one that isn't in Claude.ai
Here is the part most finance leaders haven't seen, because it doesn't live where they've been looking.
Managed Agents run on the Claude Platform — platform.claude.com, the developer-facing surface — not in the Claude.ai app. They run autonomously for extended periods, on schedule, with no session and no laptop open. And critically, they come with the operational shell that compliance-heavy work actually requires:
- Audit logs — every action recorded with actor, timestamp, and payload
- Permission controls — what the agent may read, what it may write, what requires a human
- Credential vaults — bank and ERP credentials held properly, not pasted into a chat
- Monitoring — success rates, failures, escalations
Read that list again as a controller. It's the same list you'd demand from any system that touches your ledger. That is not an accident: this is the surface Anthropic built for work that has to survive an audit.
For a finance function, Managed Agents are where the recurring, deadline-driven work belongs: nightly bank reconciliation, bill processing from the AP inbox, payment proposal generation, anomaly detection on the GL, the month-end close checklist that starts on day one whether or not anyone remembered. The work that must happen, on time, with evidence — whether or not a human felt like opening a chat that day.
The decision framework
When someone asks "should this be a chat, a plugin, or an agent?", the answer falls out of two questions: is a human present? and does it recur?
Most finance functions will end up using all four — Chat for investigation, Cowork for batch work, Plugins to carry the accounting logic, Managed Agents for everything that runs on a calendar instead of on curiosity.
Why the map matters more than the capability
Here's my actual thesis, and it's not about Anthropic's product line.
The capability already exists. The models are good enough; the surfaces are shipped; the controls on the fourth surface are stronger than the controls in most mid-market finance stacks today. What hasn't caught up is the mental model. Finance leaders are evaluating "Claude" as if it were one tool — usually by trying Chat for a week — and concluding it's impressive but not operational. They're judging the whole product by its most supervised surface.
Meanwhile the actual operating pattern — judgment in Chat, batch work in Cowork, domain logic in Plugins, recurring execution in Managed Agents, all posting to one audit-trailed ledger underneath — is a complete finance function. Not a demo. A playbook.
That gap between what's possible and what's understood is the single biggest arbitrage in finance operations right now. The teams that close it first won't be the ones with the biggest AI budgets. They'll be the ones with the clearest map.
This map is the foundation of how we built Artifi — one ledger underneath all four surfaces, with the accounting logic shipped as open, readable Plugins and the recurring work running as managed agents with full audit trails. If you want to see the four surfaces running real finance workflows — reconciliation, reporting, payroll, VAT, close — the walkthroughs show actual runs, screenshots and all.