productivityatlas.

Research / Chapter 09

Try it. Measure it. Then decide.

A repeatable way to test, adopt, restrict or reject a tool.

Research edition · 14 September 2026

The report’s observations, prices and forecasts are a dated snapshot. Examples of savings are estimates unless explicitly identified as study results. This is not a fresh verification of every claim.

Common pitfalls

  • Buying several overlapping assistants, note apps, meeting bots, and writing tools.
  • Confusing a broad feature list with a complete workflow.
  • Treating downloads, user enthusiasm, or a vendor ROI calculator as productivity evidence.
  • Building a critical process on a beta feature or an unstable product name.
  • Ignoring usage allowances, credits, per-resolution fees, token pricing, or auto-purchase settings.
  • Allowing agents to write directly to systems of record before measuring false positives and duplicates.
  • Putting confidential information into a consumer account because the interface is convenient.
  • Assuming a citation means the answer is supported; citations can be irrelevant, stale, or fabricated.
  • Letting AI-generated notes become an unowned archive rather than accountable tasks.
  • Automating email, calendar changes, payments, permissions, hiring, or customer commitments without approval.
  • Measuring time saved while ignoring review time, rework, context switching, surveillance, and workload pressure.
  • Failing to export data and document a replacement before a price or feature change.

Test / adopt / restrict / avoid rubric

Decision Use when Minimum evidence
Test Potentially useful but uncertain Defined task, safe sample data, baseline, owner, budget cap, and stop criteria
Adopt Benefit is repeatable and reviewable Measured quality/time improvement, acceptable error rate, trained users, documented policy, export path
Restrict Useful only with narrower permissions or review High value but material privacy, action, or regulatory risk; use sandbox, approval, and limited scope
Avoid No clear job, no evidence, unacceptable data handling, or unbounded cost Written rationale and a revisit trigger if the product changes materially