Document Processing
Basis needed AI agents that could reliably finish long, detail-heavy accounting tasks with strong intent understanding, lower latency, and better task-path decisions than earlier models.
Reported impact
Basis cut completion time on complex tax work, improved internal evaluation scores, and reduced the effort needed to encode one-off rules for every edge case in accountant workflows.
Basis uses GPT-6 Astra inside tax-workflow agents designed to automate the repetitive, high-detail work accountants perform every day. The company benchmarked Astra against GPT-5.6 Sol on a 50-tab tax workbook and found Astra was able to complete the task in half the time while also showing a stronger understanding of user intent. That mattered because Basis cares not only about final answers but about how agents progress through long-running tasks: Astra makes better early decisions, corrects itself less often, and can adjust its reasoning effort dynamically as a job moves from simpler to harder steps without dropping cache state. Basis also evaluates whether agents follow templates, check their own work, and consult primary tax sources when appropriate, and Astra improved those internal evaluations significantly. Together, those gains let Basis reduce the amount of hard-coded rules needed for corner cases while increasing confidence that agents can generalize beyond what was explicitly covered in tests.
User instructions and workbook tasks enter Basis agents, which route the work through GPT-6 Astra for reasoning, task-path selection, and step-by-step completion. Internal evaluations then assess whether the agents followed templates, consulted primary sources, and checked their own work before results are returned.
"Basis saw about a 20% improvement in its internal evaluation scores with GPT‑6 Astra"
"GPT-6 Astra does a better job of really understanding the intent of the user and the problem."
"GPT‑6 Astra is able to complete that workbook in half the time that GPT‑5.6 Sol is able to."
"The model can make these adjustments while keeping its cache intact."
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
OpenAI Sample Oct 16, 2026 Benefit Street Partners Benefit Street Partners needed to answer investment-committee follow-up questions faster, compare lender marks across shared loans, and review 300-page credit agreements for clauses that could shift value away from… 10–12 hours total. BDC mark tracker · Codex · Credit agreement analyzer OpenAI Sample Oct 16, 2026 Itaú Itaú needed to make financial guidance simpler and more personalized for millions of customers while also scaling document-heavy internal workflows, all under the governance, observability, and economic controls… 100% of the customer base is the goal by the end of 2026 Customer DNA · IARA · OpenAI models OpenAI Sample Jul 31, 2026 Univé Univé needed to build organization-wide AI capability across insurance operations so employees could safely redesign their own work instead of waiting for centralized software projects or isolated AI pilots. 85% of licensed users active every week ChatGPT Enterprise · Custom GPTs · Workspace Agents 