bbv KI Webinar - Review/Outlook 2024/2025

Review and Outlook: Making Sense of AI in 2024, Anticipating 2025

In public discourse, AI is often treated as if it consists of ChatGPT and a few image generators. That understates things, yet the observation says a great deal about the dynamics of 2024: language models are becoming synonymous with "AI" and thereby dominate both perception and investment decisions. That is exactly where this review starts. Anyone developing an AI strategy for an organisation needs less to know every individual headline and more to understand the big movements: which model category prevails? Who shapes the market? Which patterns are establishing themselves in usage, and which assumptions about progress and costs are realistic?

2024 as the Year of Language Models - and Why That Matters for Enterprises

2024 is once again defined by Large Language Models (LLMs). This is not merely a continuation of 2023 but a shift in breadth: language models are no longer understood only as "chat" but as a generic capability layer. Organisations see that from the same basic capabilities (understanding, structuring, generating) a remarkable variety of business applications can be derived, from text work and research through to assistance in specialist processes.

At the same time the boundary between "pure language model" and multimodal AI is blurring. Image and audio capabilities are increasingly being integrated into the large models, making it appear as if everything is "one model" even though different modalities and product capabilities underlie it. For strategy this means: not every AI topic needs to be pursued in parallel. Anyone who consistently observes LLMs and their productisation already understands a large part of what moves the market. Other AI breakthroughs, for example in specialised domains such as protein folding, remain relevant but are often less directly transferable to general business processes.

OpenAI as the Discourse Driver - Even If "Best Models" Is Not the Whole Truth

A second dominant pattern in 2024 is the strong centralisation of attention: OpenAI shapes the narrative around generative AI. This is evident less in the fact that OpenAI is unchallenged "number one" in every category, and more in the fact that product announcements, releases and terminology have such strong pull that other developments are overshadowed in perception. Even where competitors catch up technically or pull ahead in specific areas, public attention often remains with OpenAI.

"You can really just watch what OpenAI does and get a reasonably good sense of what is happening generally in the generative AI market."

Multimodality is a central leitmotif here: models that understand image input, handle speech as input and output and are adding video as a capability layer on the horizon. The strategic question for organisations is less "which demo is impressive?" and more: which capabilities are becoming stable and affordable enough to migrate into products and processes as building blocks?

In the video space 2024 also shows an interesting paradox: OpenAI dominates the headlines with "Sora", but the actual innovation front is broader. Video models existed before, and there are models that excel in particular qualities (for example photorealism). For practice this means: market observation must not run on brand names alone but must compare capabilities, quality criteria and integration options.

Assistants, Copilots and the Pattern "Humans Stay in the Driver's Seat"

At the application level, 2024 establishes "assistant thinking". Organisations often introduce generative AI in a way that gives employees a tool to consult on a point-by-point basis: draft an email, summarise text, structure ideas, accelerate small sub-tasks. This applies both to general tools (ChatGPT, etc.) and to product-bound copilots (for example in office or coding environments).

"The common denominator: the human sits in the driver's seat. The assistant waits until asked."

This pattern is pragmatic, quick to introduce and often the first step. At the same time it has side effects: it delegates the actual usage strategy to every individual. Employees must figure out for themselves when the use is worthwhile, how to prompt cleanly, which data may be shared, how to verify results, how to avoid plagiarism and where the limits lie. Without an accompanying enablement structure, a patchwork emerges: some use it effectively, others not at all or incorrectly, and many stay with the same "low-hanging fruit" as always.

A further practical problem concerns differentiation: from the user's perspective a chat window is a chat window. Whether company knowledge, RAG, additional tools or policies sit behind it is not automatically visible. This leads to false expectations ("why can't the assistant do this as well as ChatGPT?") or, conversely, risky behaviour ("is this now internally safe, or public?"). Anyone who seriously wants to anchor assistants in organisations therefore needs not only technology but product thinking: clear positioning, clear boundaries and clear communication.

RAG Becomes "State of the Art" - and at the Same Time Less Dramatic Than Before

Retrieval Augmented Generation (RAG) stabilises in 2024 as the standard pattern. For many business-facing assistants it is the key: rather than hoping the model has "learned" the specialist knowledge "somewhere", deliberately provide relevant documents, guidelines, knowledge articles or process descriptions and generate answers from them. This gives considerably better control over timeliness, traceability and context fidelity.

At the same time, RAG design is shifting as context windows grow: when models can process very large contexts, "the perfect selection of the three best snippets" becomes less central. Instead the emphasis shifts toward robust retrieval strategies, data quality, permissions and a clean handover into context, along with the question of how much information is useful without diluting the answer.

AI Becomes Dramatically Cheaper - a Strategy Factor, Not a Side Detail

A core finding of 2024 is the price collapse: for comparable benchmark performance, costs are falling by orders of magnitude over just a few years. This fundamentally changes the economics of many use cases. Things that seem "too expensive today" can quickly slide into the range of "running operating costs are negligible". This creates new categories of applications: more volume, more automation, more continuous analysis, more "always-on" checks.

"Intelligence is getting cheaper. That opens up use cases that are currently still at the margin."

Importantly, not every task needs the "strongest" model. 2024 makes it clearer that model selection is an architecture decision. There are tasks where a smaller or cheaper model is perfectly adequate, while complex tasks are handled by targeted use of more powerful models. This differentiation is often a quick lever in organisations to reduce costs while enabling more applications.

At the Same Time, the Feeling of "Quantum Leaps" Is Stagnating

Parallel to the price collapse, 2024 presents a different picture on the pure "intelligence curve" of closed-source top models: the dramatic jumps perceived when moving from earlier model generations are flattening out. Rankings change, individual providers overtake each other in specific areas, but the steps feel incremental. This leads to a strategic correction: progress is not automatically extrapolatable in a linear fashion.

Several possible limiting factors underlie this: high-quality training data is not available in unlimited supply, scaling runs into technical and economic limits, and major breakthroughs often require new research approaches (architecture, training, data, optimisation). In practice this means: organisations should not wait for everything to be "magically twice as good in six months". The better approach is to amortise the capabilities of the current generation while observing the next wave in parallel.

Open Source Grows Up - "Privacy vs. Performance" Is No Longer a Hard Choice

2024 is also a strong year for open-source models. The central change: open source is reaching a proximity to top models in many benchmarks that was previously not realistic. This shifts the classic dichotomy ("closed source = good, open source = weak but private"). Open-source models increasingly offer a combination of solid performance, low cost and high control.

"The switching costs are almost zero: a different endpoint, and suddenly massively lower costs."

For organisations this is strategically relevant for two reasons: first as a cost lever (depending on setup and hosting), second as a dependency question. When alternatives become realistic, the risk of being fully dependent on single providers or pricing models decreases. At the same time it remains important to think through the governance question: where do the models come from, what biases do they carry, which licences apply and how can this be used responsibly in a business context?

Looking into the Crystal Ball: What Could Shape 2025

"Computer Usage" as a Bridge: Models Operating Interfaces Like a Human

One obvious trend is that models not only generate text but "see" the computer and operate it via UI actions: moving the mouse, clicking buttons, filling in fields, typing text. This looks like an interim solution but a very practical one: instead of building clean integrations and APIs everywhere, you use the interface that already exists. Particularly for software that is not AI-ready, this can enable quick automation.

"It is like using a sledgehammer on a nail, but a cool transitional phase."

Long-term, UI automation remains less elegant than direct interfaces. Short-term, however, it can close a gap: bridging processes, connecting tools, making legacy systems usable. Strategically this is less an "end state" than an accelerator for getting agentic workflows into real enterprise environments in the first place.

Reasoning Models as a New Era - with a New Price/Time Profile

Probably the strongest shift in 2025 is the "reasoning era": models that do not answer immediately but think internally for longer, examine intermediate paths and generate step sequences before committing. The effect: on hard tasks, the hit rate rises significantly, but the price is not only monetary; it also shows up in latency and planability.

"The longer the model thinks, the better the performance, but under different conditions."

A decisive point: more performance costs disproportionately more. The last few percentage points of accuracy can become exponentially more expensive. This changes the decision matrix in organisations. The question is no longer only "which model is best?" but: how much accuracy is necessary for this use case? What does an additional quality level cost me, and is it economically justified? In many cases the pragmatic optimum lies not at maximum reasoning but at a well-balanced setup of cheaper base models, targeted context (RAG) and reasoning only where it truly counts.

From Assistants to Agents: Virtual Employees Rather Than Passive Helpers

If 2024 keeps the human in the driver's seat, 2025 shifts the focus toward agents: systems that pursue goals, plan chains of tasks, execute intermediate steps autonomously and involve humans only where approvals, decisions or context are needed. This is a role reversal: the human assists the agent selectively rather than the other way round.

At the same time, 2024 shows why agents do not "just work": an open task, tools, data access and then "go ahead" is often unreliable. Agents lose focus, run into dead ends or burn budget without results. 2025 therefore sees a methodological shift: workflow-based agentics. First guide tightly, then grant freedoms incrementally. This creates not only more reliability but also better engineering: agents become steerable process components rather than unpredictable demos.

Agentic Process Automation: Integrating Processes - and Then Rethinking Them

The next step after "agent in an existing process" is redesigning the processes themselves. Agents are different from humans: less deeply specialised, but broadly deployable, quick to switch context, strong in text/analysis/structure and with access to tools. This can cut across silos. Processes that were previously strictly distributed across departments can in future be structured differently: more end-to-end, more data-driven, more organised around capability bundles rather than role profiles. 2025 is more the year of experiments and discussions than of widespread implementation, but precisely these discussions are often the beginning of genuine productivity gains.

China as Accelerator - Open Source as a Geopolitical Variable

A further trend is the growing role of Chinese models, particularly in the open-source world. If capable models from China come to define the open-source standard, this will reinforce not only technical questions but also governance and values questions: cultural biases, content guardrails, training data, political frameworks. At the same time it holds true that biases and interests exist in US models as well. The shift makes the topic more visible and forces organisations to decide more consciously which dependencies they are willing to enter.

The AGI Debate Gets Louder - and Remains Hard to Grasp

2025 is likely to be "AGI-heavy" also because the term serves as a marker for progress and market leadership. At the same time AGI remains vaguely defined. Even if a model reaches "human level" on many benchmarks, this does not automatically mean it immediately delivers massive business value in practice. Reasoning can be expensive, slow and genuinely superior only on certain types of tasks.

It is also relevant that economic and strategic interests can play a role in communications, for example through contractual constructs where declaring "AGI" could influence certain partnerships or conditions. For organisations the conclusion is a sober stance: observe the AGI debate, but weight use cases, cost profiles, integration capability and risk assessment more highly than labels.

Concluding Line for Strategy

2024 shows an operational reality: language models are the dominant capability layer, OpenAI shapes the discourse, assistants are the most widespread entry point, RAG is becoming standard and costs are falling dramatically. At the same time, the impression of "intelligence quantum leaps" in classic models is flattening, while open source is maturing and creating genuine alternatives.

2025 feels like a transition into a new era: more agentics, more process integration, more reasoning, but under new conditions: higher costs for top performance, more latency, more engineering discipline, more governance questions. The pragmatic consequence is two-pronged: consistently make the capabilities of the current generation productive and amortise them, and observe the next wave closely enough to be prepared when price and maturity tip.