AI to build solutions.
AI to perform finance work.
We use AI to analyse and build solutions. Where it adds value, we also put specialised agents inside the finished workflow.
AI to analyse and build.
We use coding agents with access to agreed files, systems and development tools. They investigate existing processes, write and run code, and test results under our direction and review.
During development, agents can attempt real work. We review the method and turn useful steps into reusable tools, checks and workflows. The finished solution may be conventional software, an AI-enabled workflow, or both.
AI within the solution.
Specialised agents perform or coordinate defined finance work using approved tools and access. They gather evidence, investigate exceptions and prepare outputs. Tested code handles calculations and validations; finance owns policy, judgement and approvals.
Software around the model manages tasks, tool use, execution records and when to stop or ask for help. Existing finance systems remain the systems of record.
Technology choices follow the workflow.
An AI-enabled finance workflow is more than an LLM. We connect an approved model provider to a controlled harness, workflow-specific tools, deterministic software and human review.
For client data, we normally use the provider and account agreed for the engagement—often the client’s approved provider or environment. Where the client has no suitable environment—or explicitly approves another route—we can use an approved provider of our own, subject to agreed data-processing terms and controls.
There is no fixed technology stack. We choose the approach around the workflow, data requirements, existing IT environment and risk.
| Technology decision | Usual starting point |
|---|---|
| AI provider and account | Use the provider and account agreed for the engagement—normally the client’s approved provider or environment for client data, or an approved provider of our own where appropriate. Consider data sensitivity, client policy, residency, retention, training settings, security and contractual terms. |
| Model strategy | Use the simplest capable model for each task. Route a mix of models only where complexity or volume justifies it, based on task complexity, evaluation results, cost and response time. |
| Agent harness | Use Pi coding agent for our own AI-assisted build work: minimal, inspectable and customisable. For the finished workflow, use the simplest client-approved or workflow-specific runtime with task scope, tools, state, logs and stop/escalation rules. |
| Tools and system access | Start with controlled read and draft interfaces. Add approved action tools, APIs, imports, exports or workflow integrations where needed, based on permission risk, approval requirements and the recovery path. |
| Exact finance work | Use deterministic code and tested tools, or existing finance-system logic and specialist services where suitable. Prioritise reproducibility, control evidence and calculation accuracy. |
| Systems of record | Core ERP, TMS, reporting and approval systems remain authoritative. Allow controlled write-back or submission only where specifically designed and approved, with system ownership, auditability and correction requirements clear. |
Bring one process that should work better.
Describe the work, the difficulty and the result you want.