Bitkom 2026: 89 percent of DACH companies in on AI

Bitkom 2026: 41 percent run AI in production, 48 percent are planning, 11 percent are falling behind. What the top 20 percent are doing differently.

Key numbers at a glance

41 percent of German companies run AI in production in 2026, plus 48 percent in planning. That gives 89 percent in or arriving (Bitkom 2026).
11 percent are currently falling behind. That is the group that will not catch up in the next 18 months.
Plus 54 percent adoption growth in DACH mid-market year on year (Salesforce AI Index Mid-Market 2026).
74 percent of the economic AI value goes to the top 20 percent of companies according to PwC 2026. The rest will not catch up.
16 to 30 percent productivity gains for the top 20 percent according to McKinsey 2026, plus 31 to 45 percent quality improvement.
80 percent of companies currently see no measurable productivity gain from their AI initiatives according to PwC 2026. This is exactly where the wheat is separated from the chaff.

Bitkom published the AI study for Germany in February 2026, and the headline number is the doubling: 41 percent of companies run AI in production, up from 17 percent the year before. Plus 48 percent in planning. That gives 89 percent in or arriving. The remaining 11 percent are falling behind today and will not close the gap easily, because the top 20 percent are cementing their productivity advantages in 2026.

At Sentient Dynamics we have been working with DACH companies since 2025 on the jump from "we are planning" into "we run AI in production". And in a few engagements already on the jump into the top 20 performers group that McKinsey says realises 16 to 30 percent productivity gains. What we see is not spectacular. It is methodical.

In this post I read the Bitkom study from a practitioner perspective and name the three investments that actually move the needle in 2026.

What the Bitkom study is really saying

The study is reduced in most mainstream coverage to "many companies are using AI". That is the boring read. The interesting data points are elsewhere:

Doubling in twelve months. From 17 to 41 percent active usage in one year. This is not linear growth, it is a tipping point. From here it gets tight for laggards because the top performers are turning their productivity advantage into market position.

Top growth area is AI in software development. Bitkom lists AI agents and AI in software development as the two fastest growing application areas. That matches our engagement reality: 90 percent of the inquiries we get in 2026 are about coding agents and agentic AI in engineering, not marketing chatbots or customer service.

Main blockers are not tools, they are people. 53 percent of adopters fail according to Bitkom because of missing competence in the team, 44 percent on data protection and legal uncertainty, 39 percent on missing integration into existing processes. Tool availability is no longer a bottleneck in 2026.

41 percent expect cost savings as a value driver. Last year that number was 19.6 percent. CFO expectations have nearly doubled in twelve months. AI initiatives without hard outcome proof will not be funded in 2026.

Where the top 20 percent stand, and what they do differently

PwC published in parallel that 74 percent of the economic AI value goes to 20 percent of companies, McKinsey says 16 to 30 percent productivity gains at that peak. Both studies run independently, both arrive at structurally similar findings: AI adoption polarises the economy.

From the engagements we run at Sentient, plus from analysis of the Bitkom data, three patterns separate the top 20 from the middle and the bottom:

Pattern 1: They measure cycle time per size class, not lines of code

The middle measures lines of code, accepted suggestions or story points. All of these are AI-distorted. Top performers measure cycle time per ticket size class with a baseline from 12 to 18 months of ticket history. That eliminates story-point inflation and allows a clean 1.5x acceleration axis.

Pattern 2: They train hands-on on real tickets, not generic online courses

The middle buys a generic e-learning licence and ticks the box. Top performers run on-site workshops with pair programming on real backlog tickets. The difference is measurable: adoption rate from 10 to 70 plus percent in 90 days is achievable with hands-on, with pure e-learning it stalls at 15 to 25 percent.

Pattern 3: They segment their workforce by data into high performers, adopters, non-adopters

The middle sends all employees through the same training. Top performers work with individual ability-and-willingness scores and three workforce segments (high performer 20 percent, adopter 60 percent, non-adopter 20 percent). Coaching paths are segment-specific, which significantly increases the effectiveness per invested euro.

Where do you stand today? AI Readiness Check, 5 min, free →

Three investments with the biggest leverage in 2026

If you do not want to slip into the 11 percent falling-behind block in 2026 and instead move up into the top 20 percent, three investments are decisive based on our engagement experience:

Investment 1: Coding agents productive in the engineering team

Not "buy licences" but pull adoption rate to 70 plus percent. That does not happen through licence distribution but through structured 90-day hands-on programmes with phase 1 setup, phase 2 workshops, phase 3 evaluation.

Investment range: typically 50,000 to 200,000 euros for a 50-FTE team in the first 90 days, depending on tool choice and workshop depth.

Investment 2: End-to-end skill and rules architecture

CLAUDE.md, Cursor Rules, Copilot Instructions are not "nice to have" but engineering standard 2026. Whoever sets up standards in CLAUDE.md plus a skill library plus custom commands cleanly multiplies the impact of every licence by factor 2 to 3. Whoever skips that buys expensive licences delivering only 30 percent of their potential.

Investment range: typically 20,000 to 50,000 euros for initial setup plus maintenance, then minimal follow-up effort.

Investment 3: KPI tracking and workforce segmentation

Without adoption rate, cycle time per size class and ability-and-willingness score you measure nothing in 2026. Boards and supervisory boards demand measurable ROI, AI Act Art. 4 requires competence proof, both need hard data.

Investment range: typically 10,000 to 30,000 euros for platform setup and analysis in the first year, then it scales.

ROI calculator: what would 1.5x mean for your team? →

What the 11 percent group is doing wrong

The 11 percent that according to Bitkom today neither run nor plan AI are mostly not "anti AI". They sit in industries where compliance or data protection is perceived as a show stopper (regulated healthcare, banks with outdated IT architecture, public sector), or in companies where the last top-down initiatives failed so loudly that nobody wants to touch the topic again.

In both cases the way out is not "another pilot" but a small, visible hands-on experience with a concrete result after 30 days. We pulled an AI-sceptical engineering team in a mid-market engagement in 2026 from "we will not do it" to "we want all Pro licences" in four weeks. Not through arguments but through two senior devs who closed a refactoring ticket in a live workshop together with Claude Code that had been planned as "three weeks, two devs". What used to be a month of work was done in a week on that workshop day.

This kind of experience cannot be transmitted by PowerPoint. But it tips 11 percent groups in 30 days.

What you should do next as CTO or CIO

If you are in the "we are planning" group (48 percent according to Bitkom), the step now is: a first structured 90-day pilot with measurable cycle time baseline and three devs as champions. Not "we will trial Copilot for two months" but a real methodology with pre and post KPIs.

If you are in the "we run AI in production" group (41 percent) but are not yet hitting top 20 performance, the step is: measure cycle time per size class, segment the workforce, set up an ability score per employee. We at Sentient build exactly that in 90 days.

If you are in the "we are waiting" group (11 percent), the step is: a 30-day demo setup with a senior-dev workshop. Low threshold, no large tool procurement, with a clear output.

Book a 30-minute assessment call, no sales pitch →

In each case: do not let the Bitkom headline reassure you that "almost everyone is in". The polarisation is happening within the 89 percent, not between the 89 and the 11. Whoever does not actively move into the top 20 performance group in the next 12 months drifts into the middle, and the middle is the most expensive position in 2026.

Pricing overview: Light from 425 euros per person, Pro success-based →

Frequently asked questions

What does "active AI usage" exactly mean in the Bitkom study? Bitkom defines "active" as verifiable productive use in at least one business function with documented impact. ChatGPT trial phases or pilot projects without KPIs do not count. The jump from 17 to 41 percent is therefore robust.

How do I measure whether my team is in the top 20 performance group? With three KPIs: adoption rate (70 plus percent is top tier), cycle time per size class (1.5x is a realistic top-quintile target in year one) and ability-and-willingness score distribution (at least 20 percent high performers in the team).

What does the jump from middle to top 20 performance cost? For a 50-FTE team we calculate 80,000 to 200,000 euros in the first year for a structured programme plus platform plus KPI tracking. With success-based Pro pricing, the fee is tied to identified savings, which smooths the cashflow.

What happens if our DACH competitors become top 20 and we do not? Concrete consequence: at the same headcount and the same engineering budget, top 20 competitors deliver 20 to 30 percent more output. Compounded over three years that gap turns into a structural disadvantage that nobody catches up without headcount reduction.

Is ChatGPT Enterprise enough for our devs? ChatGPT Enterprise is a good starting point for knowledge workers and simple coding tasks. For productive engineering workflows with multi-file refactoring, tool use and agentic behaviour you need specialised coding agents (Cursor, Claude Code, Copilot, Codex). ChatGPT alone typically delivers 10 to 15 percent productivity gain, specialised coding agents 30 to 80 percent.

We have a works council, can we still do KPI tracking per employee? Yes, with clear data protection and co-determination rules. We always set up KPI tracking with works council pre-clearance, with aggregated reports rather than individual-name reporting, and with clearly defined use cases. The workforce segmentation output is a recommendation data basis, not a personnel file.

Sources

About the author

Sebastian Lang is Co-Founder at Sentient Dynamics and leads the Agentic University programme. Before Sentient he ran AI workforce programmes in SAP's Strategy Practice with 15 plus years of engineering leadership experience. Sentient Dynamics works on success-based pricing and is in use at SHD and Bregal portfolio companies.

Subscribe to the newsletter | Sebastian on LinkedIn

Bitkom 2026: 89 percent of German companies are in on AI. Which group are you in?