During the summer of 2023, we met with Mohammed Sijelmassi, Sopra Steria’s CTO and Digital Transformation Office head and discussed how Sopra Steria was deploying internally and helping clients externally with GenAI and LLMs. Earlier this month, we met with Mr. Sijelmassi again to assess Sopra Steria’s progress on its transformation, its work with clients, and use cases.

Use Cases: Knowledge Management & Reverse Engineering Remain the Priority

Sopra Steria currently has two use case priorities: conversational agent/knowledge management to complement self-service portals and bring a higher quality of chatbots, and reverse engineering for creating documentation out of existing applications: documentation generation is time-consuming and is a popular candidate for GenAI.

Code generation is next. Sopra Steria finds that code migration between object-based languages and Java version migration (e.g., to JDK 17 LTS) works fine: the challenge is to increase code quality, create concise code, and reuse software components (LoCs) where possible. Sopra Steria is identifying internally reusable components of high quality that can be used across the company’s software developers.

Areas of mid-term investment include COBOL migration to an object-based language. There are challenges in migrating between languages relying on sequential instructions (e.g., COBOL applications) to object-based programming languages. Sopra Steria is working with IBM and is taking a reverse engineering approach to identify specifications and generate code. It also points to proprietary programming languages (e.g., SAP ABAP) where the volume of LoCs in the public domain is insufficient to have best practices.

Prompt Engineering vs. Data Finetuning

Mr. Sijelmassi also provided some light on the LLM pre-training vs. client-specific (finetuning) discussion. The company believes that the quality of pre-trained LLMs is high enough for its own and its clients’ software engineering needs. Therefore, Sopra Steria is currently focusing on prompt engineering and creating libraries of standard prompts.

The company is exploring how to finetune LLMs on its own data and specially for its software products. Sopra Steria is selectively working with clients on this and has internal pilots around its HR Software and Sopra Banking products. The company finds that training LLMs on client-specific data is a relatively limited engagement (with costs ranging from €100k to €1m, depending of the size of available data). However, it points out that the finetuning exercise relies on the quality of existing data and its curation to avoid issues such as data bias. In other words, the finetuning effort will truly depend on the quality of the client’s data estate.

Internal Deployment is the Priority

Sopra Steria has embedded GenAI into its software development tools (Digital Enablement Platform) to improve productivity: so far, the company has purchased 10k Github Copilot licenses, which it is rolling out gradually, 1k per month in a phased approach aiming to build experience and gradually derive best practices.

Achieving developer buy-in is of course critical: in early internal surveys, Sopra Steria has found most developers believe that GenAI brings productivity gains and increases software quality. It expects different levels of GenAI adoption across its developers, with freshers/recent graduates and experts driving adoption. The company is accordingly targeting middle-aged developers with training and change management programs to raise GenAI awareness. Sopra Steria also intends to create career paths where developers can become experts rather than move to management positions.

Sopra Steria is targeting 10k licenses deployed and its current 58k employees by the end of 2024. 2024 will be the year of deployment and best practices.

Vertical Use Cases Are Next

Consulting with Sopra Steria Next still accounts for Sopra Steria’s bulk of GenAI activities. Sopra Steria Next continues to identify new use cases. Recently, it deployed Copilot 365 services to help clients use it and equipped 300 consultants to collect feedback.

Beyond best practices and use case identification, Sopra Steria Next highlighted its work around:

• GenAI and its impact on the environment (energy consumption)
• Selecting LLMs, differentiating between products and open-source models
• TrustAI.

Also, Sopra Steria Next is exploring vertical use cases, initially targeting conversational agents. The company has a project with a retail bank to use GenAI and help agents cross-sell products. It is also cross-pollinating the same approach to similar industries such as retail & distribution.

GenAI To Have a Contract Impact bv 2025

2024 is primarily a year for internal deployment. Sopra Steria expects to derive productivity gains from 2025 onward and embed them in its contract discussions. Depending on client maturity the company is looking to see productivity gains from 10% to 25%. Of course, this will not translate to an equivalent price drop: additional costs in terms of license, GPU, and training need to be factored in.

Beyond productivity, Sopra Steria’s investments in GenAI demonstrate two points:

• GenAI will go beyond the hype and bring productivity gains
• GenAI brings two easy-to-deploy use cases that are quick wins. Other use cases require investment and will take time to be rolled out.