How Payhawk uses AI

Last updated: May, 2026

This article provides comprehensive information about how Payhawk applies and uses AI in its products and services.

The included information covers what Payhawk AI features do, how they work, the controls and safeguards in place to ensure transparency and compliance, and how you can manage AI usage within your workspace.

Overview

Payhawk uses AI to automate routine financial tasks and improve efficiency across spend management, procurement, travel, accounts payable, and payment workflows.

These features are classified as non-high-risk under the EU AI Act because they:

Do not make autonomous decisions affecting employment, credit, or access to essential services.
Are designed to augment and accelerate financial operations while maintaining appropriate oversight.

All new Payhawk AI features undergo:

Legal, compliance and product sign-off to confirm non-high-risk classification.
Pre-release checks to block prohibited practices.

What Payhawk AI does

Payhawk AI is designed to automate routine spend management tasks on your behalf. It actively extracts information from documents, submits expenses automatically, reminds team members to submit outstanding items through AI agents, and provides troubleshooting assistance.

Payhawk AI performs tasks such as extracting text and amounts from expense documents, suggesting and applying expense categorization, drafting requests, providing travel options, automatically retrieving receipts and invoices online, and reviewing documents for compliance.

Payhawk AI cannot process payments outside approved workflows, execute unauthorized transactions, unblock cards, influence HR or disciplinary matters, use biometrics or facial recognition for employee monitoring, assign employee performance scores, or access user credentials.

Architecture and oversight

Payhawk AI processes information in the following stages:

Input stage - Receipt images, user prompts, such as “Summarize this expense report", and metadata from transactions.
Processing stage - AI models, such as Optical Character Recognition (OCR) and Large Language Models (LLM), analyze data and generate suggestions. Alongside OCR and LLM, Payhawk uses multimodal deep learning models when processing documents attached to expenses.
Output stage - Automated categorizations, text summaries, retrieved documents, or submitted expenses.
Human confirmation stage - Comprehensive logging ensures all AI actions are traceable. Users can review and modify outputs according to company workflows and risk thresholds.

Payhawk AI features are available in:

Payhawk Web Portal and Payhawk Mobile App.
Slack (coming soon: Microsoft Teams) via the Payhawk AI Agents.
ERP systems, such as Oracle NetSuite and Exact Online via automated exports with AI-enhanced metadata.

Payhawk AI operates within defined boundaries to ensure accuracy and compliance:

Payhawk AI automatically executes routine tasks such as extracting document information, submitting expenses, and applying categorization based on validated accuracy thresholds. For high-confidence actions verified through extensive testing, the AI operates autonomously without requiring manual review.
Payhawk maintains comprehensive audit logs of all actions taken, including timestamps, data sources, and confidence levels, ensuring full transparency and traceability.

To perform secure online retrieval of expense documents (invoices and receipts), the Agent Fetch functionality of the Financial Controller AI Agent operates in ephemeral, isolated Kubernetes containers that are created on demand and destroyed immediately after each task. No state persists between tasks, and containers cannot access other tenant workloads.

Agent Fetch exhibits the following key security features:

Users log in directly to the vendor website and the AI Agent never sees passwords.
The session profiles are encrypted and stored in Google Cloud Storage (GCS), keyed per user and provider.
Every HTTP request is Cloudflare Web Bot Auth cryptographically signed with Ed25519 keys.
The AI Agent retrieves only the document and does no scraping.

Human oversight and controls

Payhawk AI automates routine tasks autonomously while providing oversight mechanisms where needed:

Document OCR - AI extracts and processes data from receipts and invoices. For standard documents with high confidence scores, extraction and submission occur automatically. Employees can review according to company workflows, role-based permissions, and expense lifecycle stages when required.
- Expense owners can review extracted details before or after submission depending on company policy.
- Approvers and reviewers can check OCR-extracted information before approving and reviewing, respectively.
- Any extracted payment details are confirmed at the Pay step by clicking on the Confirm details button before moving to actual payment authorization.
Automated actions - AI Agents execute tasks such as expense submission, categorization, and document retrieval based on validated accuracy thresholds. Users retain the ability to review and modify outputs.
Requests - Each purchase must be explicitly approved through the request lifecycle. Payhawk AI does not execute purchases automatically.
Online document fetching - The Financial Controller AI Agent must first be prompted to activate its Agent Fetch functionality and to access a specific provider's website.
Permissions - Only users with appropriate roles can authorize payments. Payhawk AI cannot bypass role-based access controls or workflow requirements set by Payhawk Administrators at the company.
Undo and edit - You can undo, edit, or delete AI-processed content within defined time windows. For example, a request or expense note can be edited or deleted before the expense or request is reviewed. All system and manual changes are logged for audit purposes.

Compliance, classification, and transparency

Payhawk AI is classified as non-high-risk under the EU AI Act. Article 50 transparency requirements are observed.

If AI is used, labeling is displayed to inform users unless the AI involvement is obvious from context. For example:

In-app indicators include a Generated with Payhawk AI - please verify banner on AI Agent replies, receipt extraction screens, and AI-processed outputs.
Files exported from Payhawk that contain AI-processed data include metadata tags or watermarks where applicable.
When labelling is not displayed, audit logs track AI usage across the platform and can be retrieved on request.

Accuracy and testing

Payhawk continuously measures and improves the accuracy of its AI features across core automation tasks. Performance is tracked internally through dedicated KPIs and regularly reviewed by the Product team.

Models are retested quarterly on representative datasets. Accuracy thresholds and known failure modes are reviewed after each major release to ensure reliable automation.

Known limitations and safe use

Payhawk AI may struggle in some specific cases, such as handwritten or faded receipts, non-Latin characters, multi-currency edge cases, uncommon invoice formats, and ambiguous expense categorization. CAPTCHA or two-factor authentication can also present a hindrance - in such cases, the AI Agent pauses and requests human assistance.

Always review AI outputs before confirming, especially for:
High-value transactions
Complex multi-line invoices
Tax-sensitive categorizations

User controls and data management

Payhawk Administrators can:

Enable or disable AI features for the entire workspace.
Revoke Financial Controller AI Agent’s access when fetching receipts and invoices online. If the AI Agent’s access for specific providers is needed, contact Payhawk Support.

Payhawk AI features process the following data:

Receipt and invoice images and metadata (amounts, dates, vendors).
User prompts (chat messages, queries).
Expense categorization and approval history.
Encrypted session profiles for the Financial Controller AI Agent (cookies, local storage, but not credentials).

Data retention:

Receipt images are retained as per Payhawk’s Privacy Policy and Data Processing Addendum.
Payhawk AI logs, such as prompts and outputs are retained for 30 to p60 days for audit purposes.
Financial Controller AI Agent’s session profiles are deleted on user revocation. Currently, this is a manual process and if you need user data to be removed, contact Payhawk Support.
Logs are stored securely and are tamper-evident.

User rights:

You can correct extracted expense data until the expense has been reviewed.
Access to Payhawk AI-processed data is controlled via role-based permissions.

Payhawk AI models are provided by vetted third-party vendors (sub-processors):

Google Gemini 3.1 Flash Lite Models or Google Gemini 2.5 Flash Models for AI Agents’ intelligence.
OCR and document processing services.

Payhawk AI features are covered by:

Legitimate interest (automation of manual tasks) and contract performance.
Access control as only authorized users can view Payhawk AI-processed data.
Retention schedules are defined per data type.

Specific privacy measures for the Financial Controller AI Agent:

Session profiles are AES-encrypted at rest in Google Cloud Storage.
Session profiles are decrypted only within ephemeral containers. Keys never leave Payhawk infrastructure.
No invoice content is logged, cached, or retained outside the expense record.
No data is shared with third parties, including the LLM provider.

For more information:

Security and incident handling

Payhawk's AI infrastructure is protected by:

ISO 27001 and SOC 2 compliance.
Encryption in transit (TLS) and at rest.
Regular security audits and penetration testing.
Cloudflare Web Bot Auth for identity verification by the Financial Controller AI Agent.

The Financial Controller AI Agent has the following security controls in place when performing online receipt or invoice fetching via its Agent Fetch functionality.

Control	Implementation
Credential storage	None. User logs in directly. The AI Agent never handles passwords.
Session encryption	AES-encrypted profiles in GCS, keyed per user and provider.
Container isolation	Ephemeral Kubernetes pods, destroyed after each task.
Network security	TLS internally. Cloudflare Web Bot Auth externally.
Agent identity	Cloudflare-signed Ed25519 requests. Listed in bots directory.
Scope limitation	Single-document retrieval. Prompt-enforced action boundaries.
Human escalation	Agent pauses for CAPTCHA, 2FA, and unresolvable obstacles.

Change management and updates

This article is refreshed:

After each major product release introducing new Payhawk AI features.
Quarterly (minimum) for accuracy metrics and model updates.
When regulations or classification change.