The Role, Decoded

What Global Finance actually does

the company's Global Finance team is the operational backbone of one of the longest-standing crypto exchanges in the world. The team spans accounting, FP&A, treasury, tax, and finance operations across multiple regulated entities and 70+ countries. Day-to-day they produce and review:

• The close cycle — monthly/quarterly accounting close: book journal entries, reconcile subledger to general ledger, accrue, consolidate, sign off. At a public-style exchange this happens on a brutal 3-5 day cadence with SOX deadlines.
• Reconciliations — every dollar (and every coin) in custody, banks, exchanges, internal wallets has to tie to the ledger. Crypto adds a hard wrinkle: on-chain reality vs. internal books, across many blockchains.
• Treasury monitoring — cash and asset positions, sweeps, FX, money-market funds, custody balances, liquidity, counterparty exposure.
• Controls validation — does every transaction pass the controls (approval, segregation of duties, threshold reviews)? SOX requires you to prove this, not just do it.
• Audit prep & reporting — internal and external auditors ask for evidence. Finance produces packages. Regulators (SEC-style, PCAOB perspective on audit firms) ask for more.

All of it is document-heavy, spreadsheet-heavy, and tied to systems that don't talk to each other cleanly. Errors carry audit, SOX, security, and regulatory consequences — so manual review accretes faster than headcount does. That's the opening for an AI Solutions Architect.

What "AI-native finance operating system" means

The JD opens with a phrase worth chewing on:

"Operating system" is doing real work here. It is not a chatbot, a one-off automation, or a Zapier flow. It is a layer that:

• Sits across finance systems (NetSuite, BlackLine, Kyriba, Fireblocks, Lukka, banks, blockchains).
• Provides shared primitives — tool catalogs, identity, audit logging, prompts, evals, HITL gates — that any future finance automation reuses.
• Hosts agentic workflows that run end-to-end finance processes (close, reconciliation, treasury monitoring, controls validation) with humans gating the risky bits.
• Is auditable by design — every action attributable to a model version, a prompt, an input, an approver.

You're being asked to build a platform, not a script. In your interview, every answer should sound platform-shaped: reusable, governed, observable.

The two simultaneous horizons

This is the single most important sentence in the JD:

Translation: you must ship Q1 wins and not paint yourself into a corner. Concretely:

In interviews, if you only show horizon 1 you sound like a contractor. If you only show horizon 2 you sound like a consultant. Show both, switch between them deliberately, name which one you're in.

The five core responsibility areas

The JD lists five buckets. Memorize them — interviewers will probe each.

1. Assess and map Finance operations

Discovery work. Map how close, reconciliations, treasury, reporting, controls validation, and audit prep currently flow. Find the bottlenecks, the manual handoffs, the integration failure points across NetSuite / BlackLine / Kyriba / Fireblocks / Lukka. Output: a prioritized automation roadmap, scored by operational leverage and regulatory risk.

2. Architect the agentic finance platform

The platform-thinking horizon. Define how agents interact with financial infrastructure, data pipelines, controls, reporting. Evaluate the stack from first principles. Build the agentic layer alongside today's stack while defining the migration path.

3. Build and deploy production finance automation

The hands-on horizon. Ship agentic workflows for reconciliations, close, audit/reporting prep, treasury monitoring, financial analytics, and controls validation. Maintain human oversight where appropriate. Design for auditability, not just efficiency.

4. Own AI governance for Finance

The "no build goes to production without..." horizon. Risk tiering before work begins. Named Process Owner. Documented data flows. Access controls. Audit logging. Rollback. HITL checkpoints. Audit-ready design documentation. Represent Finance in the company's Automate Everything Center of Excellence.

5. Build the foundation for AI-native Finance

Capability multiplication. Reusable frameworks, templates, documentation. Train finance professionals. Build dashboards measuring efficiency gains, cost reduction, risk outcomes, ROI. Leadership reporting.

Risk tiering & the Automate Everything Center of Excellence

The JD twice mentions risk tiering before work begins. That's not a throwaway — it's the gating mechanism for everything you ship. A workable risk-tier framework:

The Automate Everything Center of Excellence is the company's cross-functional governance body. As Finance domain rep, you bring finance-specific risk standards (SOX, ICFR, segregation of duties) into a shared automation playbook so the same risk-tier discipline applies across Legal, HR, Compliance, etc.

"Process Owner" — what it actually means

A Process Owner is the human who owns the business outcome of the automated workflow — typically a Controller, Treasury Manager, or Accounting Lead. They are not you. You build it; they own the output the same way they'd own a junior accountant's. Their accountability:

• Sign off on the design (data flows, controls preserved/redesigned).
• Approve the risk tier classification.
• Be the named human in the SOX control narrative.
• Hold the keys to halt/rollback if the agent misbehaves.

This is huge for your interview answers. Whenever you describe a build, say: "Process Owner is the Controller. They sign off on tier, approve quarterly performance review, and hold the kill switch." That single sentence tells the interviewer you understand the governance model.

The tech stack — what each piece is for

Claude API / Anthropic SDK

Anthropic's API for Claude models. Cold-knowledge requirements:

• Models: Opus (most capable, deepest reasoning), Sonnet (balanced workhorse), Haiku (fast/cheap, screening & classification). Pick by task and by token cost.
• Tool use / function calling: the model decides which tool to call with which arguments; your code executes; you return the result; the model continues.
• Prompt caching: cache the long stable preamble (system prompt, chart of accounts, controls policy, GL schemas) for ~90% discount on cache hits. In finance this is huge — most reconciliation calls reuse the same context.
• Structured output: JSON-shaped responses you can validate against a schema before passing to downstream systems.
• Long context: 200K-1M tokens. Useful for entire bank statements, full trial balance, multi-period comparatives.

Covered in 03-core-fundamentals and 04-deep-dive-primary.

n8n

Self-hostable workflow automation. Visual editor, with native AI/agent nodes, queue-backed execution, and webhooks. In finance contexts:

• Orchestration layer — workflows, retries, branching, schedules, error trapping.
• HITL approval steps (Slack/email approve-or-reject node).
• Lets non-engineering Finance team members read the workflow.
• Self-host = data residency & SOX-friendly.

The pattern: n8n is the orchestrator and human-gate; heavy reasoning lives in Python + Claude API; integrations live in MCP servers.

Python

The glue, analytics, and integration layer:

• API clients to NetSuite, BlackLine, Kyriba, Fireblocks, Lukka, banks.
• Pandas / DuckDB / SQL for reconciliation logic and analytics.
• Decimal arithmetic (not floats!) for money.
• Custom MCP servers (Python SDK exists).
• Eval harnesses, retry libraries, async I/O.

MCP (Model Context Protocol)

Open protocol that standardizes how agents access tools, data, and prompts. Primitives: tools (callable functions), resources (read-only data), prompts (reusable templates). In finance, MCP is how an agent reaches into NetSuite, BlackLine, or your custody system without each agent re-implementing those integrations.

Covered in depth in 04-deep-dive-primary.

SOX constraints that shape every architecture answer

Sarbanes-Oxley (SOX) sections 302 and 404 require management to attest that internal controls over financial reporting (ICFR) are effective. For your role, that means:

• Every financially material change needs a control. If an agent posts to the GL, the control is "who approved this and how is that proven?"
• Segregation of duties. The same agent (or human) shouldn't create and approve a journal entry.
• Audit trail. Who/what/when/why for every action — including which model version, which prompt, which input data.
• Change control. You can't push a prompt change to a production agent posting to the GL without going through SOX change management.
• Walkthroughs. External auditors will walk through your automated control quarterly. Your design must survive a literal walkthrough conversation.

You won't be quizzed on SOX statute. But sentences like "This agent runs in Tier 2 — write-to-staging — so the SOX control is the human approver, and the audit log is the evidence" separate you from a coder who's never shipped in a regulated environment.

Soft signals in the JD

• "Production environments used by others, not just experimentation or internal demos" — they have been burned by demo-grade builds. Anchor your stories in things real users depended on.
• "From discovery to design to build to deployment to optimization without relying on external consultants" — they want a builder, not an architect-who-doesn't-ship.
• "Translate finance and business requirements into technical roadmaps" — practice explaining a recon to a non-accountant in 30 seconds, and an MCP server to a Controller in 30 seconds.
• "Halt or redesign solutions posing regulatory risk" — they want someone willing to say no. Have at least one story where you said "this shouldn't ship."
• "High-velocity or resource-constrained environments" — they're not handing you a team. You're the one.

What to ask them

Role-fit questions to have ready:

1. "What's the highest-stakes finance workflow currently being considered for agentic automation? What gates are non-negotiable?"
2. "How is Finance currently evaluating whether an AI-drafted reconciliation or accrual is good enough before a human signs it?" (probes eval maturity)
3. "When external auditors walk through an automated control, what does the replay/evidence story look like today?"
4. "Where do n8n vs. custom Python vs. MCP cleanly divide responsibilities in your existing builds, and where is the boundary still fuzzy?"
5. "What's the biggest unsolved problem in close right now? Where would a new architect have the most leverage in their first 90 days?"
6. "How does the Automate Everything CoE actually govern? Who has veto power on a Finance build?"

These show you're thinking like an architect, not a coder.