LLMs, agent frameworks, vector stores, observability: what we see reaching production with our clients, rated Adopt, Trial, Assess or Hold.
A synthetic mapping of generative artificial intelligence technologies, inspired by ThoughtWorks' Technology Radar and adapted to the challenges of European IT departments. Foundation models, agent frameworks, vector databases, MLOps platforms: we continuously assess the building blocks that structure the ecosystem, cross-referencing technical maturity, ROI observed in production and compatibility with compliance requirements (GDPR, AI Act, NIS2).
Mature building blocks, proven with our clients on critical workloads. Validated ROI, stable ecosystem, lasting support. Go ahead: these deliver on the promise.
High-potential technologies, already tested on concrete cases. Launch scoped pilots with governance and guardrails, and build your proof of value. Trust, but measure.
Promising technologies still stabilising. Watch them actively, experiment in a lab. Too early for production, just in time to prepare for what comes next.
Deprecated or risky technologies. Retired generations, lock-in, missing guardrails, or a stronger alternative available. If you have one in production, plan the exit now.
Filter by quadrant or by nature. Every entry carries the dated fact that justifies its position, a version, an adoption figure, a regulatory text.
Route each task to the model that fits it rather than picking a single one. Pricing tiers now span a factor of twenty-five: routing is a cost lever as much as a quality one.
GPT-5.6 has shipped in three tiers since July 2026, from $0.20 to $5 per million input tokens.
Claude Opus 5, GPT-5.6 and Gemini 3.7 lead on long-horizon reasoning and agentic work. Keep them for the tasks that warrant it: their token cost stays an order of magnitude above open models.
Claude Opus 5 shipped on 24 July 2026; Gemini 3.7 Flash reached general availability on 13 August 2026.
Open weights under Apache 2.0: downloadable, modifiable, self-hostable with no per-token fee. The default choice whenever data must not leave your perimeter.
Mistral Large 3 is the largest open-weight mixture-of-experts model released by a major lab.
The protocol that plugs tools into models. It has moved beyond one vendor's specification to become the market's interoperability layer, which settles the lock-in question on tooling.
Now under Linux Foundation stewardship, backed by Anthropic, OpenAI, Google, Microsoft and AWS; 41 % of surveyed organisations in production in 2026.
Vector and keyword retrieval combined, with reranking. It is the backbone of enterprise document services, provided it is treated as one building block rather than the whole architecture.
77 % of IT and data leaders consider RAG alone insufficient for a reliable production deployment.
Stateful orchestration for agents that reach production: failure recovery, checkpoints, human-in-the-loop. The longest real production track record among market orchestrators.
Generally available since October 2025, with public deployments at Klarna, Uber, LinkedIn, BlackRock and JPMorgan.
Vector search inside the PostgreSQL you already run: transactions, backups and access control are your database's own. One less system to maintain, as long as volume allows.
Past roughly ten million vectors a dedicated engine becomes relevant again, that is the threshold we observe.
Standardised gen_ai.* attributes to trace model calls, tokens consumed and latencies. Make it a buying requirement: it is what lets you change observability tools without re-instrumenting.
The four most widespread LLM observability platforms now export in OpenTelemetry format.
Governing everything that enters the context window, instructions, memory, tools, token budget, compression, instead of polishing a wording. The prompt becomes a subset of the discipline.
82 % of IT and data leaders say prompt engineering alone is no longer enough; 95 % consider context engineering important at scale.
Dedicated vector engines that hold up when metadata filtering gets demanding, because filtering sits inside indexing rather than being applied afterwards. Test them on your real filter set.
Managed costs 1.5 to 3 times self-hosted around ten million vectors, but removes the scaling workload.
Tracing, prompt versioning and continuous evaluation. Without this layer a quality regression goes unnoticed until the customer complains. Langfuse self-hosts under an MIT licence.
Langfuse passes 28,000 GitHub stars and self-hosts with no licence restriction.
The counterpart to MCP between agents: capability discovery, delegation, response. Still young, but already one of the two interoperability signals vendors point to.
MCP and A2A are the two protocols cited as interoperability signals in the 2026 agent framework market.
Llama, Qwen and DeepSeek now stand comparison on a good share of benchmarks. On a scoped use case at steady volume, in-house inference changes the business case.
Open-weight models rival proprietary alternatives on many benchmarks in 2026.
Serving an open model on your own hardware, from developer workstation to GPU cluster. The building block that makes sovereign inference operable rather than theoretical.
Test first on low-variability workloads, where GPU sizing can actually be computed.
Several specialised agents splitting a task. Demonstrations convince; tuning cost and the difficulty of reproducing an incident still keep production at arm's length.
Handle in the lab with a capped token budget and a full trace of every exchange.
Useful when context cannot carry the knowledge: house style, business terminology, latency reduction on a repeated action. The prerequisite is still an evaluation set that proves the gain.
Without a prior evaluation set, fine-tuning is indistinguishable from an expensive hunch.
Agents that take a ticket and return a merge request. The gain shows on scoped, well-tested tickets; it collapses as soon as the specification is vague.
Measure against your own DORA metrics before any rollout.
Keeping an agent's state between sessions rather than rebuilding it on every call. The topic raises as many compliance questions as efficiency gains: what is remembered becomes data to govern.
Memory is one of the four pillars of context engineering, alongside instructions, retrieval and tools.
GPT-4, Claude 3, Gemini 1: these generations are no longer served, or will not be for long. If one is still called in production, migration is a planning matter, not a watch item.
Every retired generation takes its tuning with it: plan the evaluation campaign before the switch, not after.
A loop that decides and executes with no human checkpoint and no audit log. The risk is not model performance but the absence of any way to regain control when the chain drifts.
The AI Act has been fully enforceable for general-purpose models since 2 August 2026, with fines up to 3 % of global turnover or €15 million.
As soon as a model reads content it did not produce, a web page, an email, an uploaded file, that content can carry instructions. Without a detection layer and permission partitioning, the agent becomes an attack surface.
Partitioning the agent's permissions matters as much as the filter: an agent that cannot write cannot be diverted into writing.
Calling a single model from across the application code. The day pricing changes, the generation is retired or the service goes down, there is no fallback path.
Models already on the market have a compliance window until 2 August 2027: abstraction is also what will make that switch manageable.
Decision trees and scripts, renamed. Nothing wrong with them as tools, the problem is the AI budget they consume and the expectation they create without ever being able to meet it.
The test fits in one question: does behaviour change when the data changes, without a rule being rewritten?
Generative AI has stopped being an innovation topic and become an industrialisation one. The IT departments that succeed are not those adopting the most technologies, but those choosing the right building blocks, governing them properly and deploying them with guardrails. This radar filters the noise down to what we see reaching production.
RAG keeps knowledge outside the model: you update a document rather than retraining, you keep control of your data, and you can trace an answer back to its source. Fine-tuning still applies to house style, business terminology and latency on a repeated action, but it demands an evaluation set that proves the gain, which is exactly what most projects lack. That said, 77 % of IT and data leaders now consider RAG alone insufficient for a reliable production deployment: the discipline that surrounds it, context engineering, is what carries the result.
It is not a binary choice, and it is not decided once for the whole company. Mistral Large 3 and Small 4 ship under Apache 2.0, so they can be self-hosted with no per-token fee: that is the default whenever data must not leave your perimeter. Frontier proprietary models earn their cost on long-horizon reasoning and agentic work. What matters is the abstraction layer between the two, which lets you route by task rather than commit the whole codebase to one provider.
Both, and the difference is entirely in what surrounds them. An agent with a human checkpoint, a capped token budget, partitioned permissions and a full trace of every exchange is an operable system. The same agent without any of that is an attack surface: as soon as it reads content it did not produce, that content can carry instructions. Start supervised, then loosen the loop on what you have measured.
Since 2 August 2026 the Commission has held its full enforcement powers over general-purpose model providers, with fines up to 3 % of global turnover or €15 million. Models already on the market have a compliance window until 2 August 2027. Concretely, for a company deploying rather than publishing models, this makes two things non-negotiable: knowing which model is called where, and being able to produce the trace of what it did.
Each edition is dated and reflects what our teams observe in client production, not a market forecast. This one is the September 2026 edition. A building block only moves quadrant when something verifiable moves with it: a version, an adoption figure, a regulatory text, or a failure we saw for ourselves.
Where each building block sits in your own architecture is a conversation, not a table. Tell us what you run and we will go through it with you.