Time and cost saved
How much could the agent save on your project?
The agent writes the first draft of each project document. Your team reviews it and makes the calls. Move the sliders to see an estimate.
Human in the loop
The agent drafts. People decide.
Code review: the agent flags payment risks in the code, an engineer reviews and fixes them, and a person approves the merge.
Audit evidence: every answer is saved as an audit record. A reviewer checks it, attaches the evidence, and marks it verified, verified with changes, or rejected.
The agent never approves, deploys, moves money or signs off anything. It gives guidance, not legal advice.
Illustrative estimate based on typical hours for each document on one mid-size payment project. Your results will vary. A person still reviews every output. Token cost uses Anthropic list prices (Claude Haiku 5.5: $0.10 per million input tokens and $0.50 per million output tokens; Claude Sonnet 5.5: $2 and $10) for about 35,000 tokens read (mostly cached) and 1,500 tokens written per answer.
Real benefits
What the agent really changes
One connected thread, from idea to production
- Replace your compliance team, testers or sign-offs
- Guarantee rule details are current; every output needs review
- Touch real customer data, systems or money
These are the benefits the agent is designed to deliver. Measured results from real projects will be added here.
Deployment options
Where the agent can run
The AI model runs in the cloud. Choose the option that fits your data rules and budget.
Claude API
The website sends each question to Anthropic's Claude API and shows the answer.
- Data
- Sent to Anthropic over an encrypted connection
- Answer quality
- High
- Cost
- Pay per answer, a few dollars a year at this volume
- Setup
- Built and ready on this site
Claude in your own cloud
The same agent runs on Claude through AWS, Google Cloud or Microsoft Azure, inside your own cloud account.
- Data
- Stays in your cloud environment, under your controls and contracts
- Answer quality
- High, same model
- Cost
- Pay per answer, billed by your cloud provider
- Setup
- Weeks: connect the agent to your cloud account
Open-weight model on your servers
A free open model runs on GPU servers you own or rent.
- Data
- Never leaves your servers
- Answer quality
- Usually lower for complex payments and compliance work
- Cost
- GPU servers running all the time, often thousands a year or more
- Setup
- Months: hosting, security, updates and scaling are yours to run
I work with your bank's approved AI provider
| Provider | Models | How banks usually get it | With this agent |
|---|---|---|---|
| Anthropic | Claude (Haiku, Sonnet, Opus) | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry | Default. Tools and approvals work. |
| OpenAI | GPT models, the models behind ChatGPT | OpenAI API, or Microsoft Foundry on Azure | Supported: OpenAI or compatible endpoint |
| Gemini, the models behind the Gemini app | Gemini API, or Google Cloud | Supported: OpenAI-compatible endpoint | |
| Microsoft | GPT and Claude in Microsoft Foundry; Microsoft 365 Copilot for office work | Azure subscription; Microsoft 365 license | Supported through Foundry. Copilot is an office assistant, not a model service. |
| Amazon | Nova, plus Claude, Llama and Mistral on Bedrock | Amazon Bedrock on AWS | Supported: Bedrock |
| Open-weight | Llama, Mistral and others | Your own servers | Supported: compatible endpoint |
ChatGPT, the Gemini app and Copilot are chat products for people. The agent connects to the models behind them through each provider's API.
I keep the playbooks, payments instructions, human review and audit log the same. Only the model behind them changes, and I re-test the answers on real payment questions before any switch.
No fine-tuning is needed. The agent reads the playbooks on every answer, so it stays current when a guide changes. In every option, a person reviews and approves the output.
Log in to use the Payment FDE agent
This agent is open to invited users. Log in with the user ID and password SP gave you.
Need access? Contact SP.