Development Log: Verifiable Excel Workflows, from Preview to Finished Report

The latest Excel work was not about adding another button. It was about making long, practical jobs finish with a result a person can review and use.

GetSheetAI can now recognize a growing set of business workflows from an ordinary request, build a preview, ask for confirmation where the workbook will change, write the result, and verify what was written. This release also tightened the completion loop so an intermediate analysis sheet is not mistaken for the final deliverable.

Eleven business workflows are now available

The new workflow engine covers 11 production-ready workflow types:

• company and CRM data standardization;
• supplier invoice extraction and validation;
• deterministic data-cleaning review;
• supplier quotation comparison;
• inventory health and reorder analysis;
• forecast and budget variance;
• financial reconciliation;
• sales performance analysis;
• branch-report consolidation;
• version and formula-drift comparison;
• insurance policy extraction and comparison.

These are not fixed templates that force every workbook into the same shape. The agent uses the request and workbook structure to select a workflow, while the calculations that should be deterministic — totals, differences, ranks, margins, stock cover, and matching rules — remain deterministic.

Preview first, then controlled writeback

The main design rule is simple: analysis and workbook mutation are different stages.

Before a workflow writes results, it prepares a preview artifact and identifies exceptions. A supplier comparison can show unmatched items; an invoice run can flag duplicate invoice numbers or totals that do not reconcile; a CRM cleanup can surface ambiguous duplicates rather than silently merging them.

When a write is approved, GetSheetAI uses the existing checkpoint, paged writeback, verification, and recovery controls. Large result sets are written in bounded pages. The system reads the destination back and checks that the expected result exists. If a recoverable write fails, it can continue from the safe state instead of asking the user to repeat the whole task.

Finished means a usable deliverable

Multi-step spreadsheet work often creates temporary sheets. Earlier versions could treat one of those staging sheets as the result, even when the user had asked for a separate report.

The completion logic now distinguishes intermediate analysis from a user-facing deliverable. A task is not considered complete merely because data was calculated. If the request calls for a report, the report must be written, checked, and presented before the final summary appears. If the analysis sheet itself is the requested output, it still has to meet the formatting requirement.

The progress display follows the same principle. It reports the active tool and the real step in progress, while hiding raw function parameters and internal JSON. The final message summarizes the actual completed work instead of showing redundant recovery choices after success.

Capability boundaries are explicit

The available workflows operate inside the active workbook and supported reference files. They do not silently send email, schedule jobs, update a CRM, or write to an external database.

Those external actions require an authenticated connection, ownership checks, idempotency, and explicit user confirmation. Team assignment and remote approvals also remain disabled until a tenant-scoped backend is enabled. Insurance comparison extracts and compares policy information; it does not make underwriting, claims, or coverage decisions.

What this changes in daily work

A request can now be outcome-shaped rather than tool-shaped. For example:

“Compare these supplier quotations by landed cost and lead time, flag missing SKUs, and create a ranked report on a new sheet.”

The useful result is not a paragraph explaining how to compare quotations. It is a checked comparison table, with exceptions made visible and the source workbook preserved.

That is the direction of GetSheetAI's Excel development: more work completed end to end, with fewer hidden assumptions and a clearer boundary between analysis, approval, writeback, and verification.