Interview with Anup Sable, COO & Board Member KPIT, MD Technica Engineering

“The real bottleneck is not software – it is validation”

Why speed of change becomes a risk without a platform and validation strategy

Software-defined vehicles promise faster innovation, continuous updates and new business models. Yet the higher the speed, the greater the risk of losing control. Anup Sable, COO & Board Member KPIT, MD Technica Engineering, argues for a shift in perspective: it is not software development that limits progress, but integration, validation and platform maturity. A conversation about AI as an accelerator, fragmented architectures, and why the traffic jam usually occurs just before the destination.

AUTOMOBIL-ELEKTRONIK: Dear Anup Sable, at last year’s Automobil-ELEKTRONIK-KONGRESS (AEK) you described speed of change as the key metric for the success of software-defined vehicles. How do you assess this metric today – especially in light of generative AI?

Anup Sable: Its relevance is even greater today than it was a year ago. Generative AI has massively increased the pace of innovation in engineering. What used to be incremental development has now become acceleration by orders of magnitude. Code, test scripts and analyses are created in minutes rather than weeks.

But this is precisely where the risk lies. The real bottleneck is no longer development itself, but integration and validation. I like to compare it to a journey from Munich to Frankfurt: in the past you travelled by bicycle, today you drive a supercar at 200 km/h – only to end up in a traffic jam shortly before reaching your destination. That traffic jam is validation.

Fragmented architectures as an innovation bottleneck

You have criticised fragmented E/E architectures and underestimated integration complexity for years. Where do you currently see the biggest technical stumbling blocks on the road to the SDV?

Anup Sable: There are two core issues. First, early and systematic platform integration is often missing. In classic, highly distributed architectures, the OEM’s role was clearly defined: specifying requirements, defining signals and integrating functions. Tier-1 suppliers delivered complete ECUs, and integration complexity was manageable.

Today, OEMs develop large parts of the software themselves. Functions are distributed across multiple ECUs, domains and compute units. From a functional perspective, this usually works. What fails later – and often painfully – is the platform: performance, timing, load behaviour and resource management. These effects typically only become visible very late in the development cycle.

That is why platform build-up and its validation before feature integration are critical. This is exactly where we at KPIT, together with our sister company Technica, focus our efforts.

SDV thrives on updates – and on regression

The real value of software-defined vehicles lies in continuous software updates. Yet many OEMs still struggle with this. Why?

Anup Sable: Because fast updates require fast regression. If you want to roll out software monthly or even quarterly, you must automate the entire validation cycle. Manual testing is neither time-efficient nor qualitatively suitable for this.

The key point is that automated integration and regression cannot be an afterthought. They must be considered upstream and anchored architecturally. The reality is sobering: most SDV programmes do not fail because of feature development, but because validation does not scale.

AI as a tool – not a miracle cure

What role does artificial intelligence play at KPIT in concrete terms?

Anup Sable: We have developed our own AI workbench with so-called deep agents. Of course, we use base models from OpenAI, Anthropic or Google. But the real value is created by everything around them: clean data preparation, consistent context management and memory across iterations.

Today, our AI agents support defect triage by analysing fault patterns across domains and identifying root causes far more quickly than traditional manual processes. They are also used for automated code generation to shorten development cycles and relieve engineers in a targeted way. Another key focus is the creation and maintenance of validation scripts, which makes regression testing scalable and reproducible.

Defect triage: from months to minutes

Where do you currently see the greatest leverage from generative AI?

Anup Sable: Clearly in defect triage. A vehicle is one of the most complex systems we build. Faults can originate in any domain – infotainment, ADAS, body, chassis or networking.

In the past, it took weeks or even months for specialists to identify the root cause. Today, AI can perform this analysis in minutes – provided the context is complete. The boundary is clearly defined: garbage in, garbage out. Incomplete logs, missing videos or unclear specifications lead to hallucinations. AI is extremely powerful, but it does not forgive imprecision.

Validation from the test bench to the fleet

What role does continuous monitoring of fleet vehicles play?

Anup Sable: It plays a central role. Our validation tools synchronise stimulation and observation with an accuracy of up to 80 nanoseconds across multiple CAN, Ethernet and IO channels.

This enables identical tests on the test bench, in laboratory vehicles and in production vehicles without any changes to the test logic. Field issues can be replayed one-to-one in the lab and analysed under controlled conditions. Faults become reproducible rather than random, which significantly accelerates root-cause analysis and makes the entire development cycle more robust.

Architecture trends: evolution instead of revolution

Zonal, central or hybrid – which E/E architectures are currently prevailing?

Anup Sable: Many OEMs have realised that a direct leap to a fully centralised high-compute platform was too risky. Some Gen-4 programmes have been stopped.

The current trend therefore follows a clear development logic: from distributed architectures to domain-based approaches and finally towards centralised compute platforms. A typical example is replacing classic gateways with powerful domain controllers. This approach reduces risk, creates learning curves and is economically viable. In this case, evolution clearly beats revolution.

Regulation: AI strengthens safety

How do the AI Act, functional safety and cybersecurity influence your SDV strategies?

Anup Sable: AI is not a safety risk – it is a safety tool. Used correctly, it can even improve safety validation.

Of course, AI in the vehicle must comply with safety requirements. But AI for engineering, validation and compliance is a huge gain. This is precisely where we support OEMs, both technically and at process level.

New role distribution in the SDV ecosystem

What do OEMs learn from collaborations such as that between KPIT and Mercedes-Benz R&D India?

Anup Sable: OEMs want to develop software themselves – that is irreversible. Tier-1 suppliers will increasingly focus on hardware.

Our role lies in between: early integration, platform validation and automation. Together with Technica, we develop validation platforms, and with Qorix– our independent company with ZF and Qualcomm – middleware that works independently of the SoC. Our goal is to prepare platforms in such a way that OEMs can develop faster, more safely and at scale.

Looking ahead: SDV 2030

Will there be a global SDV reference architecture by 2030?

Anup Sable: My wish is clear: yes. The reality will probably result in several scalable reference architectures.

OEMs differentiate through the user experience, not through architecture. Architecture is complex, but it is not a differentiator. That is why we should harmonise it. Every additional variant increases the probability of errors. Harmonisation accelerates development – and AI amplifies this effect massively.

Dear Anup Sable, thank you very much for this conversation.
Anup Sable: Thank you. I am looking forward to AEK 2026 – and to continuing the dialogue with the industry.