My Take on Capgemini's Generative AI Enterprise Report

Last month Capgemini released a report about applying GenAI in enterprises. The report is based on a survey of 1100 executives from organisations across countries and sectors. Below are some interesting info and my take.

1. Generative AI moves from pilot to mainstream

93 % started exploring generative AI
46 % are piloting use cases
21 % are already implementing gen-AI in certain functions.

I see that in 2025, LLMs are already mature enough to be applied in some simple use cases. Besides, the barrier to apply Gen-AI is way lower comparing to before LLM-era: Just pick a simple project without sensitive data, and use API from a reliable provider (OpenAI, Google) is enough to start. There is no reason for the other 7% of organizations to not start exploring Gen-AI.

2. Telecom and consumer sectors lead adoption: Sector adoption varies widely

Telecom operators report that 49 % have used gen-AI to some functions
Consumer products and aerospace/defence are not far behind, 40%+
Retailers, however, remain stuck in pilot mode, with only 9 % deploying in operations

We see that adoption is different sector-by-sector. That gives a offer benchmarks for executives to see where they are comparing to peer. Telecom and consumer leading are doption are quite expected. but I’m quite surprised that aerospace/defence adoption is that high, given accuracy and safety are paramount in these sectors. Capgemini doesn’t provide more details though.

3. Customer operations and marketing explode with AI

Customer operation: rise from 4 % to 41 % from 2023 to 2025
Marketing adoption surged from 2 % to 41 % over the same period
Risk management, finance and human resources also have double-digit growth

Customer support and marketing are where Telecom and Consumer sectors apply GenAI, that explains why their adoption are so high comparing to other sectors. For finance and human resources, GenAI can go through a long list of documents and provide a report to support personel processing cases fasters, as you see in recent examples of big banks I mentioned in my weekly newsletter.

4. Budgets vary substantially:

2/3 now allocate a dedicated gen-AI budget
Average budge for gen-AI are roughly 12 % of IT spend
Big firms allocate more percentage of budget to GenAI than smaller firms

This looks like the AI race may widen the gap between large incumbents and smaller competitors. But I don’t see it as discouraging to smaller firms, as the bigger the organization, the more overhead when applying new technology.

5. Cost is a bigger barrier than hype suggests:

57 % say high implementation costs hinder scaling
55 % believe the initial investment may outweigh the benefits
74 % report an unexpected surge in cloud consumption
51 % have suffered bill shocks.

As I mentioned before in how to choose LLM models for projects, people usually just pick the most popular frontier models (GPT, Claude) and then are shocked by the API bill. There are a lot of room to optimize the cost, eg. picking the right models and the right context. Doing correctly could cut the bill up to 100x.

6. Proprietary models dominate, but open-source is gaining traction

Half use proprietary models from big providers
35 % choose smaller vendors
13 % rely mainly on open-source models
65 % are exploring open-source alternatives to reduce costs and avoid vendor lock-in.

Unless you really need the best of the best, you can always find some open-source models with similar performance and that are way cheaper than frontier proprietary ones. Besides, you don’t have to worry about proprietary vendors updating their models suddenly.

7. Small language models (SLMs) will become ubiquitous

Adoption of SLMs is expected to rise from 24 % in 2024 to 44 % in 2025 and 92 % by 2028.
Some favour SLMs for their ease of development (85 %), cost-effectiveness (71 %) and customisability (69 %).

I see many reasons that SLMs are the way to go in the next few years:

As language models get better, small models will be enough for your needs
You can host smaller models with commodity hardware
You ensure your data stays on-premise, which is a requirement in some industries.

However, SLMs usually require fine-tuning to get the best results, so a bit more work is needed.

8. Vendor stack readiness

Out-of-the-box products from platforms (eg. Salesforce, SAP, ServiceNow) reduce build time.
Most organisations prefer to partner rather than build agents from scratch

Both buy and build options have merits. It is better to weigh both options, but beware the sales pitch that agents can save you millions right off the bat. If you don’t have expertise, starting with pre-made solutions from platforms is a solid choice, ideally by subscription to reduce upfront cost.

9. AI agents are new but expected to take over tasks

1 % currently use AI agents at scale
31 % are exploring them
12 % are piloting use cases.
85 % expect agents to manage one of innovation, marketing, sales or operations.

It is true that marketing, sales, HR, etc. are great for applying agents, as some parts of the work are repeatable, so can be handled by AI agents. However, it is better to start with a simple goal: how to automate some steps of our workflow, then the “agentic” parts will slowly come into the picture.

10. Employees worry about job impacts

78 % plan to invest in AI agents to perform specific tasks
31 % expect agents to take over full roles
62 % report that employees are concerned about AI’s impact on their jobs.
Capgemini: AI agents will take over individual tasks rather than entire roles.

Even if AI takes over individual tasks rather than entire roles, it is unavoidable that the work of 3 employees could be done by just 2. This means that some (not all but surely some) companies will reduce their headcount. So the concern of employees is totally valid. The bright side is, if an employee is good with AI, they will be sought after in the market. So upskilling with AI now is a good move to be future-proof.