AI Can Build the Workbook. It Cannot Own the Number.

The shift is from asking for an answer to asking for a usable object. An agent can gather material from files, email, chat and external systems; construct a workbook or report; apply a template; respond to comments; and leave the result in an editable format. That reduces the hand-off between an AI response and the software in which people complete the work. ¹

This matters because much knowledge work is not a single question. It is a chain of collection, calculation, formatting, challenge, correction and communication. Moving AI into the work surface can compress setup and iteration while preserving an object that colleagues already know how to review.

AWS has described the workflow rather than the application as an emerging unit of software delivery. That is a useful management frame: start with a business outcome and its complete workflow, then decide which steps an agent can draft, which steps a professional must inspect and which decision remains human. ⁵

SpreadsheetBench 2 tests 321 realistic generation, debugging and visualisation tasks built from business data, financial reports and corporate filings. Each task averages 11.8 worksheets and 593.5 cell changes. The paper reports that the best tested model achieved 34.89% overall task accuracy, while debugging accuracy was 12%. The dominant failures included insufficient inspection and selection of the wrong target cells. ³

WorkstreamBench reaches a similar conclusion in finance. The strongest agents produced more professional-looking outputs, but frequently fell short of professional standards as tasks extended beyond a few chained calculations. The benchmark evaluates accuracy, formulas and formatting because a finance artefact must be correct, readable and modifiable by other people. ⁴

These are preprint benchmarks, not a measure of every commercial product or configured workflow. They are strong enough to reject autonomous trust. They also point to a productive use: if expert review is already part of the process, a draft may create value before the system is reliable enough to own the outcome.

A reviewer needs more than the finished file. The test should preserve source links, assumptions, formula lineage, changes, agent actions and unresolved exceptions. It should make material uncertainty conspicuous and route decisions that exceed a threshold to the accountable professional.

The newest arXiv batch is moving in this direction. Quipu proposes governed, time-aware provenance for agent-written knowledge. Other new preprints examine coordination, long-running execution, policy-preserving action and the hidden cost of compressing an agent's context. They are not production evidence. Together they show why the artefact, history and operating path matter as much as the final answer. ⁶⁷⁸

Measure the reviewer as well as the agent. Correction time, missed dependencies, false confidence, exception demand and time to approval determine whether the draft reduces work or simply relocates it. Keep a clean prior process during the experiment so the comparison remains visible.

The draft–inspect–decide pattern works across markets. The content of inspection does not. Accounting conventions, professional standards, language, document structure, data-residency requirements, employment practices and sector rules change what a reviewer must verify.

Product availability also differs. Google describes US and EU data controls, client-side encryption and administration tools, with additional countries planned. Buyers should verify the exact models, actions, retention, processing location and audit features available in each country rather than assuming one global product name creates one operating condition. ²

In lower-connectivity or multilingual settings, a lighter workflow, local template or more human assembly may outperform the latest integrated feature.

Select a recurring, non-critical work product: a management workbook, operating report, planning pack or customer-review draft. Use real structure with protected data or an approved safe dataset. Name the owner, reviewer, baseline time, professional acceptance criteria and actions the agent is not allowed to take.

Run the complete workflow. Measure first-draft time, correction effort, error severity, reviewer confidence, time to approval and whether another professional can modify the result. Record where the agent lacked context and which failures were difficult to detect.

Expand when the draft reduces total effort and the owner can still explain the result. Redesign when review becomes the new bottleneck. Stop when a polished output hides dependencies the professional cannot efficiently reconstruct. Do not expand authority simply because the file looks finished.