The takeaway
A useful dataset lets a reviewer follow a real assignment through its context, actions and outcome.
What Turing announced on October 8
Company-reported research: Turing introduced CompanyBench using five years of one fintech firm’s operational history, spanning a data warehouse, documents, Slack, Jira, email and support systems. Assignments originate in historical work. Agents must produce practical outputs, including updates to workplace tools.
What the results do—and do not—show
Turing reports that its strongest tested model passed every required check on 44.4% of attempts across 64 hard tasks, with ten attempts per task. The underlying company data is private. These are results from a selected task set at one company; they do not establish a general enterprise success rate, licensing price or guaranteed improvement from buying data.
Our analysis: preserve the evidence around a task
For a business owner, the useful question is simple: could someone understand how this piece of work was completed? Keep the request, relevant policy, discussion, decision, action and result connected. For example, a support resolution becomes more informative when it links to the original issue and the approval that allowed a refund.
Our connection-manifest guide explains how to document those relationships. A small, representative sample can reveal missing links before a wider review.
Your next step
Choose a few completed workflows internally and list the systems involved. Note missing records and any restrictions on sharing. Then request a quality review to discuss connectivity, outcome evidence and sample scope. Do not send sensitive records with the initial request.
Sources & further reading
Sources checked October 12, 2026.
