# Evidence note: AI-assisted customer support in a small business

Last checked: **18 August 2026**. [Version française](tpe-customer-support-cases.fr.md).

This note supports the
[synthetic small-business case](../examples/en/tpe-customer-requests.md).
Its volumes, costs, thresholds, and results are fictional demonstration values.

## Useful anchors

- A Croatian SME case published by the
  [European Digital Innovation Hubs Network](https://european-digital-innovation-hubs.ec.europa.eu/knowledge-hub/success-stories/ai-powered-digital-assistant-customer-support)
  reports 50–60% less manual workload, faster standard replies, and fewer
  escalations. Definitions, raw data, and uncertainty intervals are not public,
  so these figures are a high scenario, not a benchmark.
- [Generative AI at Work](https://academic.oup.com/qje/article/140/2/889/7990658)
  studied 5,172 support agents and found 15% more issues resolved per hour on
  average. Less-experienced workers gained more; highly experienced workers had
  small speed gains and slight quality declines.
- A [maritime field study](https://arxiv.org/abs/2412.12732) found that drafts
  can help but often require substantial edits in a specialist, critical domain.
- [Dillon et al.](https://www.nber.org/papers/w33795) found less time spent on
  email among active users but no aggregate change in task quantity or mix.

The sources support an editable starting estimate for comparable drafting and
retrieval work. They do not justify automatic sending, job reduction, or an
annual return claim. Measure corrections, escalation, quality, active human
time, and all requests in the same denominator.

The worked case uses a 22% synthetic saving. It is intentionally between the
large controlled-trial average and the higher supplier-supported SME report.
The correction rate remains high enough to keep sending under human control.