Insights · July 9th, 2026
A rare longitudinal study of AI inside a real corporate planning process finds that the executives asking “how much time will this save?” are asking the wrong question.
Published in Futures (July 2026), the study by René Rohrbeck (EDHEC Business School), Stephan Szuppa (Siemens Professional Education / SRH Berlin), and Julia Schmidt (Bluemorrow) tracks a single strategic foresight process across two complete cycles: fully human-led in 2023, AI-augmented in 2025. Read the study here.
Siemens Professional Education’s “Trend Radar” process launched in 2019 and generative AI tools were introduced incrementally from 2024, mapped against the five-phase Scope–Scan–Analyse–Assess–Act model.
The researchers drew on process documentation, trend profiles, effort logs, interviews, and two focus groups.
The headline: no uniform acceleration, no automation. What AI produced was task-level efficiency, a redistribution of human effort, and improved analytical breadth, structure, and completeness — concentrated in scanning, consolidation, and drafting.
The numbers that matter.
- Drafting time per trend profile: ~3.5 hours → ~1.25 hours
- Routine analytical labour in Scan and Analyse: roughly halved
- External reports analysed: 3 → 16
- Trends on the long-list: 73 → 87
- Prioritised initiatives: 4 → 8
- Overall cycle duration: unchanged
That last line is the one to sit with. The cycle is deliberately synchronised to Siemens’ annual planning rhythm; nobody set out to compress it. Time savings appeared inside tasks, not across the process. Meanwhile, external advisory and tooling costs went up moderately — configuration, governance, and validation are not free.
Where AI was allowed, and where it wasn’t. Source selection happened entirely before any AI touched anything. Reports and databases were curated by humans against three criteria: publisher credibility, coverage triangulation (a trend needed two independent sources to enter the long-list), and topical relevance. AI was never used to expand the source set or to judge source quality. It extracted, summarised, deduplicated, drafted, and proposed relationships.
The Scope phase used no AI at all. Practitioners rated its value there 1.9 out of 5 — the lowest of any phase — against 4.2 for Scan, 3.8 for Assess, 3.7 for Analyse, and 3.4 for Act. Overall: 3.8.
One genuine process change. AI was used to generate preliminary initiative hypotheses early, in the Scan phase, rather than after the interviews. Those hypotheses were then put in front of 57 expert interviewees to be challenged. Option generation moved upstream; authority over interpretation did not move at all.
Discussion Points: The Five Tensions
The paper’s most transferable contribution is not a metric but a taxonomy of trade-offs that participants reported repeatedly. These are not implementation failures. They are structural features of hybrid systems.
Speed versus legitimacy. AI-generated outputs were not accepted as legitimate inputs until experts had debated them. As one participant put it, being faster is worthless if people don’t trust the results or feel involved.
Breadth versus focus. AI reliably lowered omission risk — “the risk of forgetting something important is lower.” But broader coverage made prioritisation harder, not easier. Another interviewee: give people more options and everything starts to look relevant.
Standardisation versus context. Trend profiles became markedly more consistent. They also, in one participant’s words, don’t really know Siemens. Contextual intelligence stayed human.
Structured ideation versus radical novelty. AI helped surface alternative framings and handle “unknown unknowns.” It was described as recombining existing knowledge rather than producing genuinely discontinuous strategic imaginaries. The authors connect this to Doshi and Hauser’s finding that generative AI lifts individual creativity while reducing the collective diversity of output — a real problem for a discipline whose product is plural futures.
Augmentation versus accountability. AI could propose, summarise, and structure. It could not carry responsibility. The authors lean on Carnat’s argument that the mere presence of a human in the loop can shield an AI system from contestability rather than resolve accountability — a rebuke to the standard corporate governance answer.
Why It Matters
Three reasons this study should displace most of what executives currently read about enterprise AI.
It is longitudinal and internally controlled. Same organisation, same unit, same process architecture, same planning rhythm, same business context — run twice, two years apart. The authors are careful not to claim causal isolation, and they say so. That candour is itself a signal of quality.
It refuses the efficiency frame. Prior research, the authors argue, has too often equated AI’s value with speed, scale, or automation. Those metrics are insufficient for foresight. The gains here were indirect: better breadth and structure improved the conditions for good sense-making, rather than producing better strategic insight automatically. AI outputs are not outcomes.
It gives you a map of the jagged edge. The paper invokes Dell’Acqua and colleagues’ field experiment with 758 consultants: inside AI’s capability frontier, users outperformed; outside it, AI use degraded performance. Foresight is not one task but a sequence of analytically heterogeneous activities, and practitioners cannot easily tell which side of the frontier each one sits on. This study locates the frontier empirically for one high-stakes process.
What This Means for CEOs
Stop underwriting AI programmes on cycle-time reduction. The cycle time here did not move — and the deployment was still clearly worth doing. If your business case rests on calendar compression, you have probably mis-specified the benefit and will declare failure at the wrong moment.
Budget for the cost line nobody forecasts. Routine labour fell; advisory and tooling spend rose. Governance, configuration, and validation are the price of admission, not overhead to be optimised away.
Put the human gate before the AI, not after it. Siemens curated every source by hand before any model saw it. Most enterprise deployments do the opposite: let the model gather, then have a human skim the output. That inverts the control point.
Demand output-level traceability — the paper admits Siemens didn’t have it. The authors name this as a limitation: no systematic link from each AI-extracted statement back to the source passage it came from. A reviewer could accept an AI-generated trend statement without knowing what it derived from. Build the traceability log before you scale the volume.
Understand what “human in the loop” does and doesn’t buy you. It can function as a liability shield rather than a control. Specify, in writing, which decisions are non-delegable, who signs, and what evidence must accompany a recommendation. Presence is not accountability.
Protect the front end. Scope was the one phase where AI was deliberately excluded, and practitioners rated it lowest for AI value. Framing, boundaries, and defining what counts as a useful answer are where executive judgment is irreplaceable — and where automation temptation is strongest.
Fund AI literacy, or don’t proceed. The authors are blunt: without skilled practitioners capable of critically engaging with AI outputs, integration may exacerbate epistemic risks rather than mitigate them. Good tools plus mediocre operators produce confident, well-structured, plausible errors.
The paper’s closing insight is the one worth carrying into a board meeting. The boundary between human and machine work in foresight is not set by what the models can currently do. It derives from organisational and epistemic requirements — legitimacy, contestability, responsibility. Better models will move the boundary’s details. They will not dissolve the logic.d about where these systems still break—will be positioned to lead the transition rather than absorb it.
This is an executive brief based on “Artificial intelligence in strategic foresight: Evidence from a longitudinal case at Siemens Professional Education” – read here
Read other articles in the series:
The CEO’s guide to AI: The Pope Has an AI Strategy Memo for Your Boardroom
The CEO’s guide to AI: State Media Control Influences Large Language Models
The CEO’s guide to AI: Social engineering is turbocharged by AI
The CEO’s guide to AI: Black-hat LLMs and Cyber-threats
The CEO’s guide to AI: We’re nowhere near AGI
The CEO’s guide to AI: Young workers as the canaries in the coalmine
The CEO’s guide to AI: The risk of vibe coding complacency
The CEO’s guide to AI: The Case for Superhuman Adaptable Intelligence (SAI)
About Nikolas Badminton
Nikolas Badminton is the Chief Futurist & Hope Engineer at futurist.com. He’s a world-renowned futurist keynote speaker, consultant, author, media producer, and executive advisor that has spoken to, and worked with, over 500 of the world’s most impactful organizations and governments.
Nikolas is an artificial intelligence expert and his 2026 keynote ‘The AI Leader: Create Incredible Productivity, Profit & Growth’ is the level up for the modern CEO and executive leader.
Please contact futurist speaker and consultant Nikolas Badminton to discuss your engagement.
