An email in two minutes instead of ten, minutes at the push of a button: these gains are real. But they often vanish into the day. According to Gartner, only 21 % of the time saved with AIU goes into higher-value work. More than 1,800 employees were surveyed. Only 3 % of organisationsU exceed their cost-saving goals through productivity.
The reason is often simple: AI speeds up a step that was never the problem. If a quote waits three days for approval, a faster draft does not help much.
What the research suggests
In November 2026, Harvard Business Review recommends mapping workflows end to end, looking for the bottlenecks and redesigning the process around them instead of automating single tasks.M The article cites research with 515 companies: those that had seen examples of how others redesigned their work around AI found 44 % moreM use cases and completed 12 % more tasks. The summary does not say who ran the study or how. The figures are therefore a pointer, not proof.
A second lesson from the article: once one bottleneck is solved, a new one often appears elsewhere. Mapping is not a one-off exercise.
The workflow
- 01StartPick a workflow that happens often and involves several people
- 02PersonWrite down every step with everyone involved, including handovers and waiting times
- 03AIAssistant turns the notes into a table; participants correct it
- 04SystemFor one week, record handling time, waiting time and queries per step
- 05CheckIs the bottleneck in a step AI can take over or prepare?no: change rules, approvals or responsibilities first, then measure again
- 06PersonRedesign the step and measure again with the same method
- 07ResultAfter a quarter, look for the next bottleneck
The table
| Column | What you enter |
|---|---|
| Step | What happens, in a short sentence with a verb |
| Role | Who does the step, as a role, not a name |
| Tool | Where the step happens, such as Outlook, ERP or Excel |
| Handling time | How long someone actively works on it |
| Waiting time | How long the case sits until the next step starts |
| Queries and errors | How often the case goes back, and why |
The bottleneck is usually the step with the longest waiting time or the most queries. It is rarely the step that feels most tedious.
Here are my notes from a conversation about our workflow. Turn them into a table with the columns Step, Role, Tool, Handling time, Waiting time, Queries and errors. One step per row, in the order of the workflow. Do not invent values: if something is missing, write "open". At the end, list the handovers between roles and mark where I should ask follow-up questions.
An example from practice
For what this looks like, see our article on order entry at t3n. The bottleneck was not writing but transferring an order confirmation into tasks with subtasks and due dates, up to 40 minutesM per large booking. An agent now does that step and the team reviews it.
Limits and risks
- Memory deceives: Estimated times are often wrong. Measure for a week before you decide.
- Works council: Recording handling times per person can be a technical facility under Section 87(1) no. 6 BetrVG.R Record times per step, not per person, and involve the works council early.
- Data protection: Notes with customer names belong only in tools covered by a data processing agreement under Art. 28 GDPR.R
- Origin of the figures: Gartner is a survey; the HBR figures come from a summary. We have not checked the original studies.
How to start
- Pick a workflow that happens at least weekly and involves at least two roles.
- Book 60 minutes with everyone involved and write the workflow down together.
- For one week, measure handling time, waiting time and queries per step.
- Mark the bottleneck and decide on one measure, with or without AI.
- After four weeks, measure with the same method and compare.
Sources
- UGartner: HR Research Finds 77 Percent of Board Members Expect AI to Deliver Productivity and Cost Savings, 26 Oct 2026
- MHR Dive: Companies are not saving money with AI, 29 Oct 2026
- MHarvard Business Review: For Real Performance Gains with AI, Start by Mapping Your Workflows, 1 Nov 2026
- Mt3n: Blueprint, how we built an AI agent in plain language, 29 Oct 2026
- RGerman Works Constitution Act (BetrVG), Section 87
- RRegulation (EU) 2016/679 (GDPR), Art. 28