Key Takeaways
- 91 percent report higher productivity: Among companies using generative AI, almost all of them notice a difference in day-to-day work.
- Adoption is not the bottleneck: 61 percent of German companies already use AI or plan to, yet only a minority can point to a measurable business result.
- Data and scaling are the real blockers: A missing data foundation and no scaling strategy are the two obstacles named most often.
- Mature AI beats the field on cost: Companies with mature AI operate roughly 4.5 percent more cost-efficiently than their competitors.
- Innovation is the underrated benefit: 76 percent of users report new ideas for products and services, not just faster paperwork.
- Relevance beats gadgets: Budget belongs on use cases with a revenue lever, not on the most impressive demo.
Introduction
Nearly every second company is discussing or planning artificial intelligence. Far fewer can name the process that got faster, the cost that went down or the revenue that went up. That gap between activity and result is the defining problem of this business year.
The uncomfortable part is that the technology usually works. Four studies published in recent weeks all point at the same causes: a weak data foundation, no plan for scaling beyond the pilot, and goals that were never tied to a business number in the first place.
This article walks through what the numbers actually say about AI productivity, why more budget does not automatically create more value, and which levers a management team can pull in the next quarter. My working principle stays the same: think first, then pick the tool.
The Productivity Promise Is Real
The OECD surveyed small and midsize companies on digitalisation and competitiveness, and the result is remarkably clear. Among those using generative AI, 91 percent report a noticeable gain in productivity. That is not a forecast from a vendor deck. It is the experience of companies that have already put the tools into daily use.
The second number is the one most executives overlook: 76 percent report more innovation. AI does not only make existing work faster, it changes what a team considers feasible. A proposal that used to take three days becomes a two-hour job, and suddenly the company bids on projects it would previously have skipped.
Two thirds of respondents also report a lower need for additional staff. That deserves an honest internal conversation rather than a footnote. In practice it rarely means layoffs in midsize companies. It usually means growth without proportional hiring, which is a very different message to send to your team.
Why Adoption and Value Are Two Different Things
Bitkom finds that 61 percent of German companies use AI in their operations or plan to do so. But 47 percent are still in discussion or planning. Planning is where many organisations quietly stall: a working group is formed, a tool is evaluated, a pilot runs in one department, and then nothing scales.
A measurable return needs a baseline. If nobody recorded how long a process took before the pilot started, no dashboard will prove an improvement afterwards. This sounds trivial, and it is the single most common reason a successful pilot cannot be defended in the next budget review.
Thirty-four percent of companies are already giving their own products an AI upgrade. That is the more ambitious path, and it only works when the customer benefit is obvious. An AI feature nobody asked for adds maintenance cost, not differentiation.
Build the Foundation: Data, Skills, Technology
PwC puts the sequence plainly in its AI performance study: get the foundation right before chasing the return. Foundation means three things — data you can actually use, people who know how to work with the tools, and technology that fits into the systems you already run.
Data quality is where most midsize companies lose the game before it starts. Offers, contracts and project notes sit in four different systems and three network drives. An assistant that cannot reach the right document will produce a fluent answer that is simply wrong, and trust is gone after two or three of those.
Skills are the second half. Handing out licences is not enablement. A short, practical session on how to brief an AI properly — role, context, task, format, examples — changes output quality more than any model upgrade. It is also far cheaper.
Spend the Budget Where the Revenue Is
The clearest recommendation from the PwC work is to point the budget at use cases with a genuine revenue lever instead of spreading it thinly across everything that looks interesting. Cost savings are easier to sell internally, but they are capped. Growth is not.
A useful test before approving anything: name the process, name the metric, name the person who owns the number. If any of the three is missing, the project is an experiment — which is fine, as long as everybody calls it that and funds it accordingly.
Benchmarking is the third piece of the advice, and it covers process, people and technology together. Looking only at the tool landscape explains very little. The interesting question is why a competitor closes a quote in two days while your team needs six.
Scaling Is a Leadership Task
The World Economic Forum report on AI in action quantifies what maturity is worth. Companies with mature AI run about 4.5 percent more cost-efficiently than their competition, and early adopters of generative AI report up to 2.4 times the output at roughly 13 percent lower cost.
The same report names the two biggest stumbling blocks: a missing data foundation and no clear strategy for scaling. Both are leadership decisions, not technical ones. No department will consolidate company-wide data on its own initiative, and no pilot team can mandate its own rollout.
Practically, scaling means three unglamorous commitments: one owner with a budget, one documented process per use case, and a fixed review date. Companies that skip the review date end up with a landscape of half-used tools and no idea which of them earned their licence fee.
Data Protection Is a Guardrail, Not a Roadblock
Privacy remains the obstacle named most often in German companies, and it is a legitimate one. It is also frequently used as a reason to do nothing at all, which is the more expensive option in the long run.
Two guardrails carry most of the load: keep personal data out of the prompt whenever the task does not require it, and choose providers that process data within a compliant framework. A surprising share of everyday use cases — summarising a report, drafting a structure, rewriting a text — needs no personal data at all.
Where personal data is genuinely needed, compliant alternatives to the mainstream chatbots exist and are mature enough for production use. The decision belongs in a documented policy, not in the individual judgement of whoever happens to be drafting an email.
Your Next Steps
- Pick one process that costs real time and measure it for two weeks before you change anything.
- Name one owner who is accountable for the number, not for the tool.
- Fix the data the use case depends on — access, structure, quality — before rolling anything out.
- Train the team on briefing AI properly instead of only distributing licences.
- Write down the privacy rules once, so nobody has to improvise.
- Set a review date and decide honestly whether to scale, adjust or stop.
Mentioned Tools & Resources
- Bitkom, Artificial Intelligence in Germany 2026: Adoption figures for German companies. bitkom.org
- OECD D4SME Survey 2025: Productivity, innovation and staffing effects in small and midsize companies. oecd.org
- PwC AI Performance Study: Why the return follows the foundation of data, skills and technology. pwc.com
- World Economic Forum, AI in Action: What AI maturity is worth in cost efficiency and output. weforum.org
Frequently Asked Questions
How much productivity does generative AI actually deliver?
According to OECD data, 91 percent of companies using generative AI report a noticeable productivity gain, and 76 percent also report more innovation. The effect is real, but it shows up in individual processes rather than as a single company-wide number.
Why do so few AI projects show a measurable return?
Most projects start without a baseline measurement and without a metric tied to the business. When nobody recorded how long the process took before, no dashboard can prove an improvement afterwards. Weak data and a missing scaling plan do the rest.
Where should a midsize company start with AI?
Start with one process that visibly costs time and is repeated often — quotes, reports, customer correspondence. Measure it for two weeks, then change it. One finished use case convinces an organisation more than five parallel pilots.
What makes the difference between a pilot and real value?
A named owner with a budget, a documented process and a fixed review date. Pilots stall because nobody is accountable for the number the project was supposed to move, so nobody pushes the rollout.
Does AI reduce the need for staff?
Two thirds of surveyed companies report a lower need for additional staff. In midsize companies this usually shows up as growth without proportional hiring rather than as job cuts — but the topic deserves an open conversation with the team, not a footnote.
How do we use AI without breaking data protection rules?
Keep personal data out of the prompt wherever the task does not require it, and use providers that process data within a compliant framework. Write the rules down once as a policy so employees do not have to improvise case by case.
Conclusion
The studies agree on an unglamorous conclusion: AI productivity is available, and most companies do not collect it. Not because the models are too weak, but because the data is scattered, the goal was never a number, and nobody owns the rollout.
The fix is smaller than it sounds. One process, one metric, one owner, one review date — and the discipline to stop what does not work. Relevance beats gadgets, and that is as true for artificial intelligence as it was for every technology before it.
