The IgnitionAI blog
Articles, field notes and insights on enterprise AI: RAG, conversational agents, LLMs and multi-agent systems.
AI agents connected to business tools: where to place human approvals
A human approval is useful when an agentic action changes data, triggers an external effect, widens a permission or leaves the expected scenario. This method helps define control points without turning every step into a bottleneck.
Multi-agent architecture: set operational limits before automating
A multi-agent architecture becomes operable when each delegation has an explicit task, data, rights, scope, output and stop condition.
Scoping an enterprise RAG project: 7 decisions to document before the prototype
Before a RAG prototype, document the need, the sources, access rights, evaluation, operations, risks and the go/no-go decision. A scoping method for CTOs and Heads of Data/AI.
Access control in a RAG: the architecture choices to document
Before indexing internal documents, document identity, the authorisation rule, rights propagation, refusals, revocation and audit evidence. This method helps design a RAG that does not widen existing access.
The AI Act and GenAI systems: organising documentation that is useful to a technical team
A useful AI Act file links the use case, the organisation's role, technical decisions, risks, tests, traces and changes. Here is a working method to prepare usable evidence without confusing internal documentation with legal advice.
Evaluating a RAG in production: quality, latency, cost and security
A RAG is not validated by a single good answer. Evaluate retrieval, answers, latency, cost and access separately, on a representative test set. A decision method for CTOs and Heads of Data/AI.
RAG evaluation: building a representative test set without skewing the decision
A useful RAG test set starts from real tasks, roles, sources and refusals. This method lets you compare versions without confusing an easy demo with production evidence.
Measuring the cost of an LLM system: a scoping method for CTOs
To scope the cost of an LLM system, measure a complete request, separate variable and fixed cost items, then compare assumptions with observed costs.
Moving from an LLM prototype to production: go/no-go criteria
An LLM prototype alone does not justify a deployment. Decide on a scope, evaluation evidence, controls, operations and a documented decision.
RAG on sensitive documents: how to define the data scope
For a RAG on sensitive documents, define the sources, uses, rights, transformations, exclusions and withdrawal rules before indexing.
GenAI systems in production: defining the conditions for suspension, rollback and escalation
For a GenAI system, plan triggers, a suspension capability, a reference version, an escalation chain and tested resumption criteria.
AI agent traceability: events, decisions and auditability
AI agent traceability links a request to a plan, controls, tool calls and a result. This method helps design traces that are useful for operations without copying business data without limit.
Handing over a GenAI system: preparing operations with internal teams
A GenAI handover is ready when the internal team can operate, restrict, evolve and explain the system without relying on an informal exchange with the build team.
Enterprise RAG: versioning, updating and withdrawing a source document without losing traceability
To maintain an enterprise RAG, link each chunk and each answer to a source version, test changes before promotion and verify withdrawal across the whole chain.
The OWASP Top 10 for LLMs 2025 applied to an enterprise RAG: the 10 risks and their countermeasures
The ten risks of the OWASP Top 10 for LLM Applications 2025, read in the context of an enterprise RAG: one concrete example per risk, the official countermeasures, and what they mean for your architecture. Includes a hands-on case of an injection and isolation test, with its results and its limits. For CISOs, DPOs and CTOs at French mid-market companies.
Kimi K3: what a frontier-class open model changes for your sovereignty and your bill
Moonshot AI released the Kimi K3 weights on 27 July 2026: 2.8 trillion parameters, 104 billion active, a 1-million-token context window. For most mid-market companies this changes nothing about self-hosting: the weights alone take around 1.4 TB of GPU memory. It changes three other things: reversibility, the market's price floor, and sovereign hosting. An analysis of the licence, the break-even maths and the decision grid. For CIOs, CTOs and innovation leaders.
RAG and access control: making a RAG inherit Active Directory (Entra ID) permissions
A RAG inherits Active Directory permissions by propagating the user's group membership to the search filter, on three levels: an ACL tag at ingestion, a pre-retrieval filter at query time, a check at response time. An Entra ID / Microsoft Graph tutorial, the six cases that break it (nested groups, token overflow, revocation), and the IAM/RBAC generalisation. For architects, lead engineers and mid-market CTOs.
Why AI projects don't reach production: what the failure numbers actually measure
Depending on the study, 30 to 95% of AI projects don't make it: RAND measures over 80% failure, Gartner at least 30% abandoned after POC, MIT 95% of pilots with no measurable return, S&P Global 42% of companies giving up. These numbers don't measure the same thing, but they point to four causes you can address at scoping. A sourced breakdown and a go/no-go method for CIOs, CTOs and innovation leaders.
How much does an enterprise RAG project cost in 2026? Real ranges, from POC to 100 million vectors
An enterprise RAG is committed in phases: €8,000 for a scoping that decides, €25,000 to €60,000 for a prototype, €80,000 to €250,000 for production, without ever signing the big budget blind. On top sits the line that derails budgets: recurring infrastructure, from $25 to over $5,000 per month depending on architecture. A line-by-line breakdown, a 3-year TCO of 9 vector databases, and the three scoping decisions that bound the cost. For CIOs, CTOs and innovation leaders.
Enterprise RAG in production: 5 critical decisions your first POCs hide from you
The laptop RAG POC holds up in a day. The same system deployed for 500 users over 100,000 documents cracks in three weeks. Five structuring decisions (ingestion, chunking, search, observability, access control) determine whether your RAG will clear that bar. A decision grid for CTOs and Heads of Data at mid-market companies.
Multi-agent architecture in production: three authorization-error patterns observed on engagements
Multi-agent systems fail at the composition layer between agents: an orchestrator with no access control, a sensitive-data sub-agent in a generic pool, loss of user context in the delegation chain. A breakdown of the three recurring error patterns and a mapping to the OWASP LLM06 Excessive Agency framework. For AI architects, lead engineers and CTOs at mid-market companies deploying multi-agent systems.
Microsoft 365 Copilot in the enterprise: the governance audit nobody does
Microsoft 365 Copilot reveals the oversharing accumulated in your tenant and makes it actionable at organisation scale within weeks. Why the prior Microsoft Purview audit changes everything, and how to run it. For CTOs, DPOs and CISOs at French mid-market companies deploying or having deployed Copilot.
The European AI Act: what mid-market CTOs must prepare after the Digital Omnibus
Regulation (EU) 2024/1689 on artificial intelligence, amended by the Digital Omnibus of 7 May 2026, imposes binding obligations on companies operating AI systems. Updated deadlines, risk levels, article-by-article obligations, action plan. For CTOs and Heads of Data at French mid-market companies.
Access control in an enterprise RAG: 5 architectures so you don't expose the executive's salary to the operator
A poorly scoped RAG chatbot exposes the leadership team's payslips to a production operator. Five access-control architectures compared, with their complexity, their security guarantee and their use case. For CTOs, DPOs and CISOs at French mid-market companies.
Tell us about your AI project
A first 30-minute call with a senior consultant. You leave with a documented opinion on feasibility, scope and order-of-magnitude costs. If we believe the project is not ready, we put that in writing.
Reply within 24 business hours from a named consultant.