Databricks puts GPT-5.5 in charge of enterprise document agents

What OfficeQA Pro actually measures

Databricks built OfficeQA Pro to test the unglamorous parts of enterprise work: parsing, retrieval, and grounded reasoning across scanned PDFs, legacy files, and long-context documents. These are the tasks Databricks says "frequently break production agent systems."

That framing matters. The benchmark is not asking whether a model can reason well in the abstract — it is asking whether a model can extract the right digit from a scanned page and carry it correctly through a multi-step workflow. On this test, GPT-5.5 became the first model to pass 50% accuracy and reduced errors by 46% compared to GPT-5.4.

A benchmark centered on scanned and legacy documents is a tell about where Databricks' customers actually operate: archives of imperfect files, not clean structured data.

Why a single mis-read digit cascades

The most concrete detail in the announcement is about error propagation. Research Engineer Arnav Singhvi describes how small extraction mistakes compound across an agent's trajectory.

Once you can't extract a certain digit or number, that changes the entire trajectory of what the agent works with.Montana Labs

This is the core insight behind the 46% error reduction. In a chatbot, a bad parse produces one wrong answer. In an agent, that bad parse feeds every subsequent step, so parsing accuracy is a multiplier on total workflow reliability. Databricks reports its largest gains from GPT-5.5 came specifically in these parsing-heavy workflows, which Singhvi calls a "step-function lift in parsing older documents and scanned PDFs."

Fewer search detours, less supervision

Beyond parsing, Databricks points to orchestration behavior. Singhvi says GPT-5.4 would sometimes take "unnecessary search detours" that produced inefficient trajectories, and that GPT-5.5 was more reliable at retrieving relevant context and finishing complex tasks without added supervision.

That reads as an efficiency and cost claim as much as an accuracy one. Wasted retrieval steps cost tokens and latency, and every step that needs human oversight erodes the case for automation. Databricks is describing a model that stays on task.

The model is being deployed as a supervisor, not a worker

The deployment path is the most specific part of this announcement. GPT-5.5 is available through Databricks' AI Unity Gateway, and it runs inside workflows built with AgentBricks and the Agent Supervisor API. In those systems, Databricks says GPT-5.5 "orchestrates parsing, retrieval, and execution across specialized agents."

So GPT-5.5 is not just answering document questions — it is being placed at the coordination layer above other agents. The implication is that Databricks is betting the reliability gains it measured on parsing and orchestration are exactly the qualities needed for a model to safely direct other components rather than merely execute one task.

Having GPT-5.5 supervise these workflows is really exciting.Montana Labs

For teams building enterprise agents on legacy document stores, the practical takeaway from this specific rollout is narrow and useful: measure your model where extraction failures cascade, and treat the supervisor role as a distinct reliability requirement from the task-execution role.