Receipt Claims Processing with Liquid AI Decision Models
Use Liquid AI decision models and its API to screen receipt claims, assess reimbursement policy and route exceptions. Explore our working receipt demo.
Executive brief
Automating receipt claims processing can reduce repetitive finance work, shorten the path to reimbursement and help teams apply company policy consistently. The opportunity is to check submissions as they arrive, prepare routine claims for approval and direct unclear or exceptional cases to the right reviewer.
This automated layer can improve efficiency and reduce manual work by preparing the checks and evidence before a person reviews the claim. The final decision can still rest with a human whenever needed, whether for an exception, an uncertain assessment or every claim under your approval policy. Reviewers receive a prepared assessment to help them decide.
A complete workflow combines receipt reading, decision models, business rules and connections to expense and accounting systems. Liquid AI's decision models can support the assessment step: identifying readable receipts, classifying expenses and flagging potential policy issues. Your business defines the policy, review requirements and authority to approve payment.
| Potential business benefit | How automation could help |
|---|---|
| Less manual handling | Extract receipt details and prepare assessments so finance staff spend less time entering data and repeating basic checks. |
| Faster reimbursements | Check completeness at submission and route ready claims to approvers, reducing avoidable back-and-forth. |
| More consistent policy checks | Apply the same approved rules and category definitions to each claim, while routing exceptions for review. |
| More focused reviewer effort | Surface unreadable receipts, conflicting amounts and unusual submissions with the evidence needed to investigate. |
| Better audit preparation | Keep the receipt, applicable policy, checks and approval history together in a traceable claim record. |
| Capacity to handle higher volumes | Automate routine preparation as submissions grow, with staffing based on the remaining review workload. |
These benefits depend on the complete workflow and the quality of its decisions. A useful pilot measures processing time, reviewer effort and incorrect outcomes against the current process. Start with one expense category and retain existing approval controls until the results justify expanding automation.
Our Receipts Decisions demo, introduced in Melvin's post, shows the receipt screening step using Liquid AI's d1 decision models and API. The sections below explain its five checks and how they could fit into a business claims workflow. Claimant records, accounting integration, approvals and payments would be additional implementation work.
Custom receipt automation for your business
AIBackends builds workflows connecting your receipts, business systems and approval process—with human decisions where needed.
Why use a decision model for receipt claims?
Liquid AI's decision models answer defined questions and return structured results with probabilities. For receipts, that could mean a yes/no assessment of readability or a category such as meals or transport.
The workflow can use these answers to prepare a claim and identify what needs attention. The assessments should be checked against claims your team has already reviewed before they drive automation.
Five checks in the receipt demo
The repository sends one receipt image and five typed questions in each API request:
| Check | What the model assesses |
|---|---|
| Valid receipt | Whether the image contains a purchase receipt with merchant, date, items and total. |
| Readable fields | Whether those required details can be read reliably. |
| Expense category | Meals, transport, lodging, office or other. |
| Within policy | Whether the purchase fits the demo's reimbursement rules. |
| Visible anomalies | Whether amounts appear overwritten, totals conflict or arithmetic looks inconsistent. |
The sample policy sets SGD limits of 50 for meals, 100 for transport, 250 for lodging and 100 for office supplies, with no alcohol or tobacco. These are demonstration rules, not a recommended company policy.
Screenshot from the original demo post. The displayed run uses unlabeled receipts, so it does not establish decision accuracy.
The local UI and CLI let developers inspect probabilities and save JSON reports. This provides a starting point for evaluating a receipt decision step, rather than evidence of a deployed claims service.
Two approaches: hosted API or local deployment
You can use a hosted decision API or run an open-weight decision model within your own infrastructure. Both approaches can feed assessments into the same business rules and human approval process. They use different model options, so evaluate each on your own receipts.
flowchart TD R[Receipt and claim context] --> A[Your claims application] A -- Approach 1 --> H[Hosted API<br/>Liquid AI runs the model] A -- Approach 2 --> L[Local deployment<br/>Your infrastructure runs the model] H --> D[Structured assessment] L --> D D --> B[Business rules and review] B --> F[Final decision<br/>Human approval when required]
| Approach | Where processing happens | What your team manages |
|---|---|---|
| Hosted API | Receipt images and context are sent to Liquid AI's hosted service. | API integration, data permissions, usage costs and approval rules. |
| Local deployment | An open-weight model runs on hardware within your infrastructure. | Model deployment, hardware capacity, maintenance and approval rules. |
Approach 1: use the hosted API
Your application sends a receipt image, the relevant context and defined questions to the Liquid AI API. It receives structured answers to prepare an assessment or flag a claim for review. This approach lets your team integrate the decision step without operating the model infrastructure.
Our demo uses the hosted API. Developers can find setup instructions in the demo repository and Liquid's API documentation.
Approach 2: run an open-weight model locally
Liquid AI's decision model documentation lists d1-2B, d1-omni-2B and d1-omni-600M as open-weight models under the LFM Open License.
Local deployment can keep receipt model processing within your infrastructure. This can suit businesses with internal data requirements, provided the rest of the workflow also follows those requirements. Your team operates the model and hardware, and should evaluate quality and total operating costs. Using this approach with our demo would require a separate integration.
From receipt screening to a claims workflow
For production, we would connect the decision step to verified claim data and explicit review rules:
flowchart TD A[Claim form and receipt] --> B[Extract and validate fields] B --> C[Liquid AI decision model] P[Applicable policy and claim context] --> C C --> D[Business rules and review gates] D -- Checks pass --> E[Approval queue] D -- Uncertain, conflicting or failed --> F[Human review]
The model helps interpret the receipt. Business rules check amounts, deadlines and allowances against trusted records. The workflow also needs claimant eligibility, the applicable policy and previous claims to check for duplicates.
A meal receipt can be readable and correctly categorized while exceeding the company's reimbursement limit. The workflow should flag that exception and present the receipt and policy to the approver. A human can retain the final decision for exceptions or for every claim, according to your approval policy.
Start with a focused pilot
Choose one expense category and compare the automated assessments with claims your team has already reviewed. Measure processing time, manual effort and incorrect outcomes. Keep unclear receipts, policy exceptions and failed checks in a review queue.
Visible anomaly checks can support investigation, but they cannot prove fraud. Reviewers should be able to see the original receipt, relevant records and assessment history. Agree which data may be sent to the hosted API before using live claims.
Include integration, storage and reviewer effort alongside API charges when assessing the business case. The goal is a more efficient claims process with dependable assessments and clear accountability.
Try the demo with your own workflow in mind
Explore the receipt decision repository to inspect the questions, sample policy and setup instructions. Use approved test receipts and reviewed labels to evaluate how the decisions behave on your expense categories.
For a related architecture that separates OCR from policy assessment, see our article on insurance receipt claims with Jev.
AIBackends builds and operates workflows connected to forms, documents, APIs and business systems. To explore receipt claims processing for your team, email us about your current process. Include the policy, systems involved, expected volume, sample receipt types, security constraints and approvals you need to retain.