Solving AI ROI in software publishing

A closer look: solving AI ROI in software publishing

1. Introduction to Pierre Chaffardon

Pierre Chaffardon is 45, married and the father of three children. He joined Generix Group in 2023 and is now its CRO. After around two decades at IBM, he oversees the installed base, growth, market share and the international strategic plan. Nawfal Laghzali, who hosts the conversation, reports on debates heard among leaders, between those who say they are doubling their productivity, those who see no impact backed by figures and those questioning the cost of tokens (the units consumed by models) or dependence on vendors.

1.1 Generix and its critical software

Generix is a French software publisher that has been active for around thirty years in two areas: supply chain and logistics on one side, finance and invoicing on the other. The company has more than 3,000 customers worldwide. Major French groups have supported its international expansion.

Carrefour is among its largest customers, and this year Generix is celebrating twenty years of partnership with Louis Vuitton. Its software runs warehouses, organizes or optimizes transportation and manages invoicing. Pierre Chaffardon describes a partnership rather than a simple customer-supplier relationship, because these systems support critical operations.

2. Framing AI around business challenges

Generix is not an AI company, Pierre Chaffardon explains, and views AI as a tool serving a business challenge. The company has not followed the approach of some AI-native companies, with large token budgets or internal rankings of the biggest consumers. For about twelve months, he believes, most companies have been looking more closely at their token bills, which has somewhat encouraged Generix to keep tight control over them.

2.1 Segmenting internal and product use cases

Pierre Chaffardon distinguishes personal uses (mainly with Copilot), internal business applications, processes to transform end to end and then AI embedded in products. Generix introduced AI into cyber training “at least three years ago, I think.” Today, an agent in Teams regularly provides employees with micro-training, rather than training delivered through five hours of PowerPoint followed by a test taken as a mere formality.

2.2 Rethinking end-to-end processes

Generix finds that ROI expands when AI covers an end-to-end business workflow. Repetitive, high-volume and easy-to-describe tasks remain a common starting point, even if gains for a single employee can be challenged unless volumes are very significant.

The company is reviewing entire roles, including that of product manager, to rethink product management, design and specification. Some roles could potentially lose their place in a processing chain redesigned by AI.

2.3 Optimizing warehouse picking

Picking is a use case embedded in the software sold to warehouses. Previously, processes were optimized separately based on a model valid for a quarter or a year. AI makes it possible to simulate the entire chain almost in real time and select the most appropriate action.

AI-optimized algorithms can save between 20% and 30% of pickers’ time, the operators who prepare orders. Since payroll is a warehouse’s largest expense item, this time can increase productivity with the same headcount or reduce processing costs at the same volume.

3. Measuring the ROI of AI agents

Pierre Chaffardon says that, “I think,” his CEO Raphaël presents Generix as a company with 800 employees and around one hundred AI agents. He confirms that around one hundred agents have been identified and, at the time of the conversation, are involved in part of the company’s internal execution activities. An AI committee oversees initiatives eligible for structuring and scaling.

3.1 From individual gains to collective impact

A tool designed solely for its author’s use remains an individual initiative. The committee becomes involved when a tool can transform the work of a category of employees. A sales manager used AI to create tools that standardize presentations, proposals and contracts for part of Generix’s services. The committee then reviews the architecture, cybersecurity, consumption and costs.

3.2 Validating business cases

The committee estimates ROI and the business case to determine whether it is worth extending the tool to all employees in the relevant category. It sets monitoring KPIs: the time required before the tool, the time required afterward and then the intended use of the time freed up.

3.3 Reallocating time saved

Pierre Chaffardon believes that companies lose part of the gains linked to AI. “If time is not automatically reallocated, human nature being what it is, it is instantly lost to the company.”

Of the roughly one hundred agents identified, about twenty have so far demonstrated their value and received managerial approval that Pierre Chaffardon describes as “somewhat collective.” For these uses, the token costs are considered to have been recouped.

4. Governing AI experiments

Nawfal Laghzali describes an evaluation framework through which an initiative passes when it is submitted to the AI committee. Pierre Chaffardon confirms that a formal approval stage exists. The committee must also maintain an inventory of initiatives to prevent an organization from losing track of what is happening when AI is in the hands of all employees.

4.1 The role of the AI committee

After this inventory, projects are categorized according to the process they target, making it possible to involve the appropriate business coaches and experts during the evaluation phase. An employee’s willingness to test AI is not enough to build quantifiable KPIs that can then be reviewed.

4.2 Sandbox, KPIs and formal validation

Generix leaves room for sandboxes (testing environments) when they appear relevant, then reviews demonstrable ROI once they are complete. Pierre Chaffardon believes that a measurement method can be learned or put in place with guidance, and that it would be unfair to expect employees to bring this skill intuitively.

The results are then presented to the committee. An agent recognized as generating gains receives approval, and the tokens associated with this type of agent are validated or prioritized by the engineer responsible for optimizing financial flows.

5. Orchestrating a multi-model AI stack

For Pierre Chaffardon, the challenge is not to identify the best-ranked LLM, or large language model, at a given time. AI serves either internal transformation or a value-added service for the customer. In both cases, Generix seeks to depend on it as little as possible.

5.1 Copilot, Claude, Mistral and open source

The partnership with Microsoft has a long history around Office 365 and collaboration tools. It has been extended to AI modules, including Copilot. Claude is used for sales strategy, go-to-market iterations and product teams. Generix has commitments with Mistral and is testing Asian open-source models within its evaluation scope.

5.2 Reducing vendor dependence

The CRO advocates an orchestrator position: select the model best suited to a given use, then, as far as possible, retain the freedom to switch to a different model. This approach relies on multi-sourcing and an engineering layer that manages the models. Generix is studying open source to retain control of services for which it makes contractual commitments to customers.

5.3 Balancing cost, performance and continuity

A price increase, a geopolitical constraint or a vendor outage can affect a service. For internal uses, Generix can absorb a drop in productivity. A service sold to customers, such as one that optimizes truck packing and loading for a given customer at a volume of around two million parcels per year, cannot stop.

This service is designed around a given model, but it must absolutely be able to change vendors. It can also, for reasons more closely related to profitability, adapt the model to a target margin without falling below the expected level of operational relevance.

Generix is betting that models will become commoditized. The difference would then lie in orchestration, selection and the ability of the higher layers to ensure operational rerouting.

6. Turning data into a competitive advantage

Pierre Chaffardon repeats a phrase heard at IBM: data was the “black gold” of the 2000s, and it remains so. Value comes from the connection between models and data, enriched by business understanding, rules, authentication and software-specific features.

6.1 Connecting models, data and business expertise

The fear of a “SaaS apocalypse,” in which AI would replace legacy software, has not materialized, in his view. AI will transform software, especially at the interface and application layer, without eliminating execution systems. Generix now combines them with simulation and optimization.

The intelligence of these systems relies, according to Pierre Chaffardon, on data that is as clean and structured as possible and on around thirty years of customer experience. These elements were integrated through custom development, then standardized in product options. Generix feels reasonably secure in this foundation and sees AI more as an opportunity than a risk.

6.2 Making the data lake reliable

Data quality is becoming a central issue. Pierre Chaffardon points out that “garbage in, garbage out” still applies: the best LLM applied to a degraded source produces a poor result.

In supply chain, Generix is conducting work that is not very visible to customers to reorganize a clean, consolidated data lake, a central data repository shared across its software components. The work is ongoing.

6.3 Getting value from e-invoicing

In finance, Pierre Chaffardon notes that regulation helps clean up data. The e-invoicing reform has a European scope. In France, its first wave took effect on September 1, 2026, while a second is expected on September 1, 2027. The structured format imposed for invoicing makes it possible to clean data as it enters systems.

Generix is still deploying its registered platform projects with customers in France. As this deployment progresses, the company will be able to accumulate structured invoicing data. Product teams are working on AI simulation and optimization use cases without having delivered them yet. For now, ideas exceed the available time, although they are not unlimited.

7. Accelerating go-to-market with AI

After three to four years of strong growth, Generix is preparing a five-year strategic plan whose ambition rests more on international markets than on France. South America still offers a wave of investment in industrialization, automation and standardization, whereas Europe is already fairly well served by software in Generix’s two business areas. North America, served from Montreal, is also a significant growth opportunity.

7.1 Analyzing markets, TAM and ICP

Market analysis and customer segmentation with Claude have become a regular part of Generix’s go-to-market. The tool helps define TAM (total addressable market) and ICP (ideal customer profile). Pierre Chaffardon estimates that the effort required to produce these analyses has fallen by around 80%. This time saved is measurable and applies to leadership teams.

7.2 Simulating acquisition strategies

With sales leadership, teams simulate several deployment plans, compare investments in cost of sales and cost of marketing, then test the gaps between direct and indirect market access. This simulation can help approach or accelerate repeatable sales.

7.3 Scaling campaigns and proposals

When a successful configuration is identified by customer profile or market, campaigns can be targeted and automated. Pierre Chaffardon also describes sales teams using AI applications to automatically generate proposals and contracts. He puts forward a hypothesis: the same ten people can produce two to three times more.

8. Keeping humans in the loop

Nawfal Laghzali believes AI does not eliminate difficulty, it shifts it. Pierre Chaffardon responds that every activity is a processing chain: accelerating one link shifts the constraint to another.

8.1 Shifting the constraint to expertise

In sales, the constraint becomes contextual commercial intelligence and sales instinct, which AI does not yet have. Pierre Chaffardon sees the difference in impact in preparing the application context and sorting the data submitted to the model. With the same tools, two people can obtain very different results, but above all work quite differently.

8.2 Making results and decisions more reliable

Predictability and trust in execution are becoming priority selection criteria. If two clicks produce two different results, it becomes difficult to assign a level of confidence to the answer. AI can provide a wrong answer with the same confidence as a good one.

Generix has hired AI specialists for its committee to make execution processes more reliable. Pierre Chaffardon believes that decisions, including investment decisions, will increasingly rely on these results: the impact of a poor sales hire may be limited, while an M&A policy carries different stakes. In go-to-market, a hypothesis delivered too quickly by AI requires human review by someone able to challenge it.

8.3 Defining levels of autonomy

The AI committee is still working on autonomy classes. It distinguishes processes left entirely to AI, those requiring an initial human review and those requiring the double review of a manager.

An investment decision, the launch of a warehouse operation or the invoicing of a critical customer engage the responsibility of the company, its leaders and the people in charge.

9. Sovereignty, resilience and continuity

Nawfal Laghzali imagines the case of a new model version that would degrade a previously satisfactory result and could also cost more in tokens. He refers to internal benchmarks, with the same set of tasks tested on two models or on the same model at two different dates. Pierre Chaffardon admits that Generix has not found a definitive answer on model stability.

9.1 Diversifying models and infrastructure

His first approach is to be able to run, with at least two providers, a use case sold to customers. A second is to combine SaaS and on-premises models in order to freeze a model considered relevant, while potentially accepting responsibility for managing its obsolescence.

The third approach aims for a dedicated, model-agnostic orchestration layer that retains the intelligence related to business objectives. Beneath this layer, model-agnostic execution must produce the same result twice when the same button is clicked twice, without performance degradation and while remaining reasonably cost-efficient.

Pierre Chaffardon considers true 100% sovereignty to have been impossible for a long time. In his view, a 100% European stack, without Chinese, Asian or American components, is impossible in today’s cloud service centers, and he believes he knows of no counterexample. He argues that it is necessary to seek some assurance from as many European players as possible with critical mass, investment and R&D resources, without requiring 100% French or European production. Dependence may lie in the model, the fiber equipping the data center or compute capacity. Generix relies on a French hosting provider, owned by an American group, with data centers in France that can be visited.

9.2 Prioritizing disaster recovery and cyber risk

For a SaaS provider, the absence of a sound disaster recovery plan, or DRP, is in his view a greater risk than a stack that is not entirely European. As a customer, he would challenge his provider on its recovery policy and backup site rather than on the nationality of its stack.

At the time of the conversation, cyber risk and industrial risk require more investment and leadership time at Generix than the nationality of every component involved in delivering the service.

10. Summary and recommendations

Nawfal Laghzali asks how a CRO decides on reallocating gains. Pierre Chaffardon responds that everything depends on the use case and priorities at a given moment. A warehouse manager seeking to reduce costs can use gains from picking or slotting, the placement of products in the warehouse, to reduce payroll.

He distinguishes between time saved and the use of recovered capacity. At Generix, this capacity is instead used to manage growth, serve more countries and deliver more projects in parallel, while keeping headcount growth contained.

For Pierre Chaffardon, the trade-off therefore varies from one use to another and can shift between financial optimization and operational optimization.