Basis puts the agent's reasoning on screen so accountants can review it

The journal entry is the interface, not the automation

Basis describes a concrete moment in its product that tells you where the real engineering went. When an accountant looks at a journal entry the system prepared, they don't just see the entry. They see an explanation of what data was used, why it was mapped that way, and how confident the system is in its recommendation.

That is a frontend decision as much as a model decision. The company built agents that "share context through a central layer, surfacing assumptions, data sources, and the logic behind each decision." In accounting, an output you can't inspect is worse than no output at all, and Basis states the principle plainly: automation is most useful if it's reviewable.

Everything we do depends on reasoning. That's why OpenAI's models, especially GPT‑5, are so critical. By scaling test-time compute well beyond what earlier models could support, while still exposing the model's reasoning, we can surface explanations that give customers visibility into and control over what is happening.Montana Labs

The important word there is exposing. A model that reasons well internally isn't enough for this product; the reasoning has to be extractable and renderable to a human reviewer. That constraint shapes what the interface can promise.

Confidence and provenance as first-class UI elements

Most agent demos show a result. Basis shows the reasoning path alongside it: the supporting materials reviewed, the data retrieved, the shared context and best practices referenced, and the sub-agents coordinated to produce the work. For the accountant, the decision surface is a chain of evidence rather than a verdict.

This is why Basis frames its trajectory as moving "beyond task automation into real workflow delegation." Delegation only works if the delegate can account for its work. The confidence signal attached to each recommendation lets a human decide where to spend review attention — which entries to accept quickly and which to scrutinize.

Function calling is what pushed this from proposal to action. Basis notes that it enabled agents "to complete multi-step processes like reconciliations and journal entries, not just propose them." The interface has to hold both states legibly: work the agent has done, and work awaiting a human's sign-off.

Explainability is a routing input, not an afterthought

Behind the interface, Basis runs a multi-agent architecture where a supervising agent — originally on o3, now migrated to GPT‑5 — routes steps to specialized sub-agents based on task, complexity, latency, and input type. Speed-critical interactions like mid-review clarifying questions go to GPT‑4.1; ambiguous classifications and month-end close go to GPT‑5.

What's notable for anyone building a review-driven UI: Basis benchmarks models on how clearly they explain their reasoning, not just accuracy. Those benchmarks decide "both which models to rely on for various tasks and when agents can safely take on new workflows." Explainability is treated as a measurable capability that gates deployment, which means the frontend's promise of reviewability is enforced upstream in model selection.

The company also reports GPT‑5 hitting a perfect 100% success rate on its parallel tool-calling benchmark with code interpreter and web search enabled. Parallel tool calls let a single workflow coordinate multiple structured actions — which the review surface then has to reassemble into one coherent, inspectable explanation for the accountant.

What the reviewable surface actually buys these firms

The reported outcome is up to 30% time savings on average across the large U.S. accounting firms Basis says it supports, with firms "continue expanding agent responsibilities as trust grows." That last clause is the mechanism. Scope expands as a function of demonstrated trust, and trust is a function of what the interface lets accountants verify.

The specific implication is that in high-stakes professional workflows, the product frontier isn't autonomy — it's the quality of the review surface. Basis didn't win by hiding the model; it won by rendering the model's reasoning, provenance, and confidence in a form an accountant can accept or reject. When the underlying model improves, the agents take on more, but only because the explanation layer makes each expansion auditable. For teams building agents into regulated or accountable work, that's the part worth copying: build the review interface first, and let it govern how much the agent is allowed to do.