In your operations layer. On your infrastructure. With your team. We measure the maturity, build the agents, hand over the controls. Your data and models stay yours.
Operations are becoming autonomous. The shift is under way. The only open question is who shapes it for the Mittelstand. Today three paths stand open, each with a price.
SPAX sits in the middle: methodologically leading, technically deep, regulatorily at home.
We build and run the agents ourselves, along a verifiable maturity method, model-neutral and with documented exit terms. EU AI Act, GDPR and the works council are part of the engagement, not an afterthought. What we recommend, we deliver.
Anyone who understands Full Self-Driving knows: at the end, the human is still in the car. What changes is the use of their hour. The presence stays. The same movement is now running through operations. Execution gets handed over. The person stays.
Whoever handles thirty complaints a day today designs the agent that handles three thousand tomorrow. From the execution layer into the orchestration layer. Promotion with leverage.
After the promotion
Writes the system prompt and tool boundaries. Sets the bounded context: which data, which actions, which refusals. Decides which calls the agent makes alone and which it hands up.
Builds test sets, replays historical cases in simulation, curates edge cases. Tracks pass rate, hallucination rate, escalation rate. Every new version is re-eval'd against the same bar.
Defines when the agent stops, whether via confidence threshold, domain boundary, or compliance flag. Who it hands to, at what latency, with what context and reasoning trace. Escalation is an auditable action, not a fallback.
Takes the cases the agent doesn't carry: edge cases, sensitive contexts, customer escalations. Every handling produces a data point: a new eval sample, a prompt update, a tool adjustment. The agent grows through the exception, not despite it.
Operates the dashboard: throughput, cost per case, escalation rate, customer view. Decides when the next level step makes sense, and when the agent is pulled back. Steering in steady state, not just in crisis.
Methodologically reachable. Across six maturity stages.
Four steps, each bookable on its own and a gateway to the next. You start small and scale only when the diagnosis holds.
Maturity diagnosis per process. We prioritise the candidates, quantify the effort and lay the path open. Fixed price, clearly scoped.
Target level per process, architecture decision, compliance path. Our own platform is assessed transparently against existing solutions, even when the recommendation goes against it.
Custom agents, deployed and running. No template sale. Your team takes over the controls from day one.
Running operation of the deployed agents: monitoring, drift detection, incident response, escalation paths and periodic re-assessment against the target level.
Compliance, workforce and governance we deliver through a certified partner network. SPAX stays your single point of contact throughout.
Six maturity stages from L0 (Manual) to L5 (Full): from the case handled by hand to end-to-end execution by the agent, with the human in a strategy and audit role. Analogous to SAE J3016 for autonomous driving, transferred to operational business processes. OAL makes verifiable where a process stands today, where it is heading, and which architecture structurally carries it.
Three things in one language: diagnosis, roadmap, and shared vocabulary for all stakeholders. OAL is the language we run the diagnosis in. It stays when we leave. Your team carries it on.
Six maturity levels. From operator to architect.
Orders come in by phone, complaints get handled via email, status updates pass by word of mouth. Excel is the only system more than one person sees. Knowledge sits with experienced staff, not in documented processes.
First dashboards show volume and handling times. Individual steps get automated: order confirmations, simple lookups. Every decision still runs through a person, every case is touched manually.
Systems classify inbound work, pull relevant customer data, propose response templates. The human sees the context preassembled and decides faster, but confirms every action themselves. Augmentation. The decision stays human.
Standard cases run through automatically: order intake at clear parameters, status updates without human touch, routine reports generated on demand. The human reviews by sampling and takes over anything outside the rule boundary.
The agent independently negotiates complaints, closes tickets, coordinates with sibling systems. When it spots an edge case or a compliance-relevant constellation, it hands over to the human with full reasoning trace. The default case runs without a human.
The agent runs the whole process end-to-end without intervention. The human moves into a strategy and audit role: sets goals, reviews audit trails, decides when the agent advances. Strategic reviews once a quarter, operational interventions belong to the exception.
The agent works inside a corridor you define. At each level it takes on more, and control stays traceable.
At each OAL level, the work time of an employee who stays shifts from Execution to Design to Strategy. Same person, different layer. That is the stack ascent.
Layers
Setup. Industrial company. Operational routine layer with daily four-digit case frequency, maturity L1.
Path. Four weeks of on-site analysis → multi-vendor architecture without lock-in → pilot on 10% of volume → gradual roll-out.
Handover. Three operations staff become Agent Trainers and Supervisors. They run the agent themselves from day one.
What remains. Scaled case frequency at unchanged headcount. Audit trail. Internal team in the new role class.
Setup. Specialized wholesaler. Moderate case volume, but every escalation is regulatorily or contractually dense.
Path. Document escalation paths → conditional-autonomy agent with threshold routing → audit trail per escalation.
Handover. Agent Supervisor role is introduced in the operations team, with clear handover logic to operations lead.
What remains. Escalation response time halved. Documented audit trail for every regulatory review.
Setup. Regulated service provider. Every case must be EU AI Act compliant.
Path. Compliance mapping per case type → hybrid architecture (autonomous run only under documented conditions) → audit-trail-first deployment.
Handover. Agent Governor role as key position for ongoing compliance responsibility.
What remains. Audit readiness at all times. Scalable compliance without linear personnel growth.
Thomas Daiber built Hubject into one of the world’s leading e-mobility platforms as its CEO: the eRoaming joint venture of BMW, Bosch, Daimler, EnBW, RWE and Siemens that made charging across providers possible in the first place. Before that, charging infrastructure at EnBW, e-mobility consulting at Porsche Consulting, fuel-cell development at Daimler.
In 2019 he founded Cosmic Cat Group in Berlin, a strategy and venture house for new mobility, from e-mobility through MaaS to autonomous driving. As a founding investor he backs several startups. He does not talk about the future of mobility, he builds and funds the companies that carry it.
Mehmet Emin Tok has spent 17 years across DACH enterprises: insurance, logistics and maritime, energy and grid operations, retail and e-commerce. Industries seen at the point where IT digitized the operational layer, often owning the rebuild himself.
At a major insurer, drove the move from monoliths to event-driven microservices across 30 development teams, with governance up to CIO level, and brought the organization from waterfall to domain-driven work. At one of the largest container shipping lines, ran the sales systems inside a program of over 100 people. In the energy sector, brought disruptive e-mobility concepts to market.
Mehmet and Thomas have run the same wave from two directions: public charging infrastructure put on the road at scale, a platform business scaled across corporate boundaries. Around the two of them, a core of engineers and strategists with hands-on build experience.
The same shift is now running through operations, at a different magnitude. We know how a company in the DACH region must be built for it. We have run the last wave already, in large corporations and in our own companies.
Hamburg · The Human Ascent GmbH
Data, models, process logic and knowledge stay owned by the company that created them. Model-neutral, with documented exit conditions. We don't do lock-in.
A clear language, a clear framework and a traceable path are worth more in the mid-market than any collection of buzzwords. Anyone promising leaps is lying.
Our own operations run on agents, from lead intake to deployment. If we claim operations can become autonomous, ours have to prove it first.
Eleven examples across ten industries.
Zalando
GenAI and AI model digital twins generate editorial and campaign imagery. Production costs fall by 90 %, and image production drops from 6 to 8 weeks to 3 to 4 days. In Q4 2024 around 70 % of editorial images were AI-generated.
Source. Matthias Haase (VP Content Solutions), speaking to Reuters · 2025
Allianz
Project Nemo applies seven specialized Agents to simple, high-frequency claims. Handling and settlement time: minus 80 %, from roughly seven days to under one day. It rolled out in under 100 days.
Source. Maria Janssen, Chief Transformation Officer, Allianz Services (Allianz Newsroom) · 03.11.2025
BMW Group
AI agents correct robots on the assembly line, computer vision (AIQX) handles quality inspection. A single application saves over 1 million $ per year, six employees have moved into higher-value roles, and quality inspection reaches five times what was thought possible.
Context. The 1 million $ refers to one specific application, not the plant as a whole.
Source. Curtis Tingle (BMW Group Manager) speaking to CNBC · July 2023
Bosch
Computer vision quality inspection, predictive maintenance and AI energy management deliver six- to seven-figure savings per year and plant. At the Bursa plant, water dropped 30 %, energy 6 % and scrap 9 %.
Context. Plant-level savings should be read as a range (six- to seven-figure). The Bursa percentages are externally validated by the WEF.
Source. Bosch Manufacturing Leadership; WEF Lighthouse recognition · 2023 to 2025
Otto Group
The Movex Virtual Content Creator generates GenAI images from a single product photo as input. This yields 5x more content per day, with new collections ready in hours instead of weeks.
Context. Productivity and speed metric, not a euro figure.
Source. Marc Opelt (CEO Otto), CIO.de · 2025
Deutsche Telekom
The AI assistant "Frag Magenta" (chat and voice, LMOS platform) handled over 5 million customer dialogues in 2023 across around 400 topics, with more than a third resolved immediately. The mail AI "Sherloq" recognizes 80 % of mail requests.
Context. Operational metrics: solid volume and automation, no euro savings are reported.
Source. Deutsche Telekom Corporate · 2023 to 2024
Commerzbank
Enterprise AI "Momentum 2030" with the "Ava" assistant, the "CobaGPT" LLM and agentic AI for complaints, KYC and contracts. More than 10 million € per year are already realised today. "Ava" handles around 30,000 requests per month, over 70 % of them autonomously.
Context. The cited roughly 350 million € in cost savings and 500 million € in profit contribution per year are a target by 2030, not an audited figure. What is realised today is the more than 10 million € per year.
Source. Bettina Orlopp (CEO), Annual General Meeting; Bloomberg · 20.05.2026
Klarna
The AI service assistant handles Tier-1 chats end-to-end with backend access. In the first month 2.3 million conversations ran through it (around two thirds of all chats), and resolution time dropped from 11 to under 2 minutes. It was set against the workload of 700 full-time staff and a profit effect of around 40 million $ for 2024.
Context. Klarna walked back the 700 full-time staff framing in 2025 and rebuilt human support for complex cases. The 40 million figure should only be read with this context.
Source. Klarna press release · 27.02.2024
Siemens
Senseye Predictive Maintenance combines AI and IoT to flag machine failures before they cause downtime. Reported figures are up to minus 50 % unplanned downtime, up to plus 55 % maintenance efficiency and an ROI of around 3 months.
Context. Vendor-reported best-case figures (up to, Siemens-reported), not an average and not an audited value.
Source. Siemens product and case material · 2023 to 2025
Moderna
ChatGPT Enterprise ran company-wide, with over 750 Custom-GPTs built in two months. 40 % of weekly active users built their own GPTs, averaging around 120 conversations per user per week, with 100 % adoption in the legal team.
Context. Vendor-produced case study. Adoption metrics are concrete and show pharma-wide reach, not to be read as primary ROI evidence.
Source. OpenAI Case Study (CEO Stéphane Bancel, CIO Brad Miller) · 2024
ABOUT YOU
On-model product imagery is AI-generated instead of photographed. Reported are around 90 % lower image production costs, cumulative figures in the millions, conversion holding steady or higher, plus roughly 50 % of first-level service requests automated.
Context. Self-reported in a podcast context. To be read as a management statement, not an audited figure.
Source. Tarek Müller (Co-Founder & Co-CEO), OMR podcast · 02.03.2026
As of: May 2026
If you have an operational layer that's a candidate for autonomy, or aren't yet sure whether it is: get in touch. We answer directly, no funnel.