Cherry Finance Service Agent

An AI finance operations agent that handles supplier and payment queries, detects invoice exceptions, drives approvals and collections, and hands complex cases to humans.

Comment

Inspiration

Finance teams spend a surprising amount of time on repetitive operational questions:

  • Which bank transaction matches this invoice?
  • Why has this bill not been reconciled?
  • Does this transaction need approval?
  • Is there enough evidence to process it safely?
  • When should automation stop and involve a human?

We were inspired by a simple idea: AI should understand the messy context around finance, but it should not have unlimited authority over financial decisions.

That led us to build Cherry Finance Service Agent — an AI-powered finance operations agent that can understand documents and transactions, reconcile them when the evidence is strong, identify exceptions, route higher-risk cases for human approval, and maintain an auditable record of every decision.

Our goal was not to build another finance chatbot. We wanted to build an agent that can actually move finance work forward while remaining governed, explainable and human-controlled.

What it does

Cherry Finance Service Agent turns bills, receipts and bank-feed events into governed finance workflows.

The core flow is:

Understand → Categorise → Reconcile → Apply policy → Approve when needed → Create audit evidence

The agent can:

  • Extract structured information from invoices and receipts using Gemini
  • Suggest accounting categories and VAT treatment
  • Compare documents against candidate bank transactions
  • Rank reconciliation matches using supporting evidence
  • Automatically reconcile a transaction when confidence is high and policy permits it
  • Stop automation when amounts, currencies or evidence do not match
  • Route higher-value or uncertain transactions to a named human for approval
  • Keep a tamper-evident audit trail of workflow decisions
  • Generate an evidence pack containing the document data, matching results, policy decision and workflow history

Our demo demonstrates three important outcomes.

1. Autonomous workflow

The invoice and bank transaction agree on amount, supplier, date, reference and currency, and the value is within the configured policy limit.

Result: the transaction can be automatically reconciled.

2. Human approval

The supporting evidence is strong, but the transaction value crosses the configured financial threshold.

Result: automation pauses and asks an identified human to approve the workflow.

3. Exception handling

The supplier and reference may match, but the bank amount differs materially from the invoice.

Result: automation stops and requests further evidence instead of guessing.

This distinction is central to Cherry Finance Service Agent:

Gemini understands the financial context. Deterministic controls decide what authority the agent is allowed to exercise.

The agent never silently assumes approval.

How we built it

We built Cherry Finance Service Agent as a cloud-native agentic application.

Agent layer

We use Google Agent Development Kit (ADK) to orchestrate the finance workflow and Gemini on Vertex AI for reasoning and multimodal document understanding.

Gemini handles tasks where contextual understanding is important, such as extracting information from invoices and understanding the relationship between financial evidence.

Finance control layer

We deliberately separate AI reasoning from financial authority.

Deterministic controls evaluate conditions including:

  • Amount variance
  • Currency mismatch
  • Matching confidence
  • Document extraction confidence
  • Whether a transaction has already been reconciled
  • Financial approval thresholds
  • Availability of supporting evidence

The system then selects one of three governed outcomes:

Auto-reconcile | Human approval | Evidence exception

Application

The application and API are built with Python and FastAPI.

The browser interface allows the finance workflow to be demonstrated visually while the API provides the underlying workflow operations.

Google Cloud

We use:

  • Cloud Run for the application and API
  • Vertex AI / Gemini for AI reasoning and document understanding
  • Firestore for workflow state
  • Pub/Sub for event-driven workflow notifications
  • Cloud Storage for evidence packs
  • Artifact Registry and Cloud Build for container delivery

The application is containerised with Docker and the cloud infrastructure is reproducible using Terraform.

Auditability

Every material workflow transition is added to a SHA-256 hash chain.

An evidence package can then be generated containing the extracted financial information, reconciliation candidates, policy decision and complete workflow history.

This gives the agent something that is especially important in finance: traceability.

Challenges we ran into

Balancing AI autonomy with financial safety

The biggest challenge was deciding where AI reasoning should end and deterministic financial controls should begin.

Allowing an LLM to directly decide whether money-related actions are safe would make the system difficult to govern. We therefore designed the architecture so that Gemini can understand and recommend, while deterministic policy controls determine whether automation is permitted.

Building reliable human-in-the-loop behaviour

Human approval needed to be an actual control rather than just another generated response.

The system therefore requires an explicit approval for the specific workflow instead of allowing the agent to infer that somebody has approved it.

Handling imperfect financial evidence

Real invoices, receipts and bank descriptions rarely match perfectly. We had to distinguish between harmless differences and material mismatches while avoiding false automatic reconciliations.

Making AI decisions auditable

A finance agent cannot simply say, "the AI decided this."

We needed to preserve the evidence, matching results, policy checks and workflow history so that the reasoning behind an outcome could be reviewed afterwards.

Cloud deployment

Bringing together the agent runtime, Gemini, Cloud Run, Firestore, Pub/Sub, Cloud Storage, container deployment and the public application also required significant integration work during the hackathon.

Accomplishments that we're proud of

We are particularly proud that Cherry Finance Service Agent is more than a conversational prototype.

We built an end-to-end agentic finance workflow that demonstrates:

  • Real multimodal financial document understanding
  • Agent orchestration with Google ADK
  • Automated transaction reconciliation
  • Deterministic financial policy controls
  • Explicit human approval
  • Safe exception handling
  • Persistent workflow state
  • Cloud-native deployment
  • Tamper-evident audit history
  • Downloadable evidence packs

We are also proud of one design decision in particular:

When the system cannot safely establish the answer, it stops.

It does not convert uncertainty into fake confidence simply to make the demo look more autonomous.

What we learned

The biggest lesson was that good agentic finance is not about maximising autonomy — it is about maximising useful autonomy within clearly defined boundaries.

LLMs are extremely powerful at understanding unstructured financial information, but financial systems also require deterministic controls, evidence and accountability.

We learned that the strongest architecture combines both:

AI for understanding + deterministic policy for authority + humans for consequential decisions.

We also learned that the human handoff is not a failure of the agent. Knowing when to stop, explain the problem and involve the right person is itself an important agent capability.

What's next for Cherry Finance Service Agent

Our next goal is to evolve the prototype into a broader finance operations agent.

We plan to add:

  • Live Open Banking transaction feeds
  • Supplier and payment-status enquiries
  • Automated invoice and receipt ingestion
  • Email-based finance requests
  • Configurable approval policies
  • Approval notifications and workflows
  • Accounts receivable and collections assistance
  • Accounting-platform integrations
  • Role-based access controls
  • More advanced fraud and anomaly detection
  • Production monitoring and observability
  • Richer agent-to-human explanations

Longer term, we see Cherry Finance Service Agent becoming the intelligent operations layer behind Cherry Money — allowing small businesses and finance teams to spend less time chasing transactions and exceptions while retaining human control over important financial decisions.

Smarter finance operations. Less chasing. Better evidence. Humans in control.

Built With

Share this project:

Updates

Submission history