The Challenge
Automating tax-form mapping with a self-validating AI agent
Every time a government agency issued a new tax form, an analyst at the company had to hand-map every field on that form to the right internal data source, a process that took multiple days per form and had to happen again, from scratch, the next time a form changed. That is not a process that scales with the number of agencies and forms a payroll platform needs to support, and it is exactly the kind of repetitive, high-stakes, rules-bound work that makes a strong case for an agent, provided the agent can be trusted to get it right.
The Approach
A two-stage pipeline for accurate tax-form field mapping
Solvd’s team built a multi-step AI agent that maps each field on a blank tax form to the correct internal data source, validates its own mapping decisions, and self-corrects until the completed form passes downstream validation on its own. A two-stage pipeline first groups raw fields semantically, employee tables, totals, company header information, and then maps each group to the right data-source templates, reusing parameters it has already resolved in earlier batches rather than starting blind each time. Domain-tuned tools let the agent search, look up, and create data sources directly, with a specialized tool handling amendment forms in a single pass. A self-validating loop rechecks every decision the agent makes, automatically remapping any field that fails and escalating to a human only the small residual it genuinely can’t resolve on its own. A custom evaluation framework checks output against golden datasets by comparing the referenced data source rather than exact text, which is a more honest test of whether the mapping is actually correct, and the whole pipeline runs through Langfuse for tracing, prompt management, and cost tracking.
The Outcome
Reduced manual effort to edge-case validation
What used to take multiple days of manual, analyst-driven configuration per form now runs as an end-to-end automated flow that produces production-ready configurations directly, reducing human effort down to reviewing only the small set of uncertain fields the agent itself flags as needing a second opinion. OpenAI and Anthropic models power the multi-agent pipeline, grouping fields, mapping them to data sources through tool use, and driving the self-validation and remapping loop that lets the system catch its own mistakes before a human ever sees them.
