Document processing
Invoices, attachments, and PDFs parsed into the fields finance or operations already use, then checked before anything posts.
AI agents
Put AI to Work Inside Your Business.
AI agents that process documents, execute workflows, interact with existing systems, and handle repetitive operational work — with rules, audit trails, and a person on the cases that actually need one.
Where an agent earns its keep
High-volume, rule-heavy work with a clear exception path. Not a chatbot bolted onto a broken process.
Workflow execution
Coding, routing, keep-versus-return, and similar decisions that currently wait on a person to copy values between systems.
Grounded operational replies
Drafts and recommendations tied to current order context and retrieved documentation, with a verifier before anything auto-sends.
Evidence from production
These are systems I built and ran, not vendor case studies.
Finance
AI-powered accounts payable
LLM workflow for invoice intake, business rules, duplicate checks, and approval routing. Estimated savings of about $150K a year in avoided manual AP work. Automotive, QuickBooks.
$150Kestimated annual savings
Repair operations
Labor-coding engine
Prediction API for repair labor coding, with an automated weekly feedback loop against actuals so accuracy keeps moving. 20% faster coding and payouts.
20%faster coding and payouts
Finance operations
Invoice matching and reconciliation
A separate system from AP automation: AI-assisted matching of OEM invoice line items to orders. The model proposes matches and flags; finance confirms before anything becomes a dispute. Built as a Next.js reconciliation app on Supabase and MotherDuck.
Technician support
Shop-message copilot
Async service that decides whether a shop message needs a reply, which workflow actions to propose, and whether to auto-send or draft for review. Courtesy messages bypass the model. Other drafts run through a grounding verifier against retrieved repair docs. Currently in beta on the CRM.
How I treat the model
Deterministic rules first. Retrieval and grounding when the answer depends on documents. A hold-for-review path when the verifier cannot support the claim. One internal system — an MCP server — was the security-sensitive build that needed careful planning before it went near production data.
- Known-safe patterns can auto-send; everything else drafts
- Business rules and already-taken actions are merged so the agent does not duplicate work
- Predictions persist so a retry does not invent a second draft
- Claude is in the stack for structured extraction and document work where that is the right tool
Common questions
Why not just use ChatGPT ourselves?
The prompt is the easy part. The hard part is the workflow around it: which cases skip the model, which systems it can write to, how exceptions are audited, and who owns the misses. That is the work I build.
Do we need to replace our existing software?
Not always. Many agents sit on top of the systems you already have. Replacement is worth it when the current tool is a poor fit and the replacement is a bounded application with a simple data model, limited integrations, and maintenance you can own. I do not claim every tool can or should be replaced.
Is this strategy or implementation?
Implementation. Scoping is part of the work. A paid discovery exists for projects where the workflow is still unclear. It is not the default pitch.
Have repetitive operational work that still needs a person?
If the volume is real and the rules can be written down, an agent is often the cheaper path than another hire or another SaaS seat.
Discuss a Project