30 to 50 % of AI initiatives fail to return what they should: abandoned projects, under-used licences, agents that cost without serving (Gartner, McKinsey). Idun Group, a consultancy specialised in AI, practises consulting built on tooling: we analyse your real AI usage from the logs you already produce, then we deliver a roadmap prioritised by return on investment that says what to industrialise, what to consolidate and what to cut. Target set and measured at your site: 30 to 50 % lower cost on the targeted usage and 2 to 4 times fewer errors, on-site in Paris and remotely anywhere in France.
of AI initiatives fail to return what they should (Gartner, McKinsey)
AI diagnostics run in a year
users measured
of AI initiatives fail to return what they should (Gartner, McKinsey).
The finding is harsh: according to Gartner, at least 30 % of proofs of concept are abandoned, while McKinsey estimates that more than 80 % of initiatives leave the operating result untouched. A budget is approved, an isolated automation is built, then nothing. The agent stays in a test environment, never connected to the teams' real data nor measured on a concrete use case. Without scoping or measurement of real usage, the result stays at zero.
Two causes come back again and again. First, weak initial scoping: the perimeter stays vague and nobody defines precisely the task to be done. Then the absence of business buy-in: the agent is built by an isolated technical team, with no operational owner involved in the decisions. Waste follows the same slope.
of AI initiatives with no impact on the operating result (McKinsey)
of the AI licences paid for go unused, across the estates we analyse (field finding)
Idun Group builds its method against those two causes. Every engagement starts with the analysis of your real AI usage, which estimates the return on investment of a use case before committing to any development. The roadmap that comes out is prioritised by return on investment, effort and adoption, arbitrated with your business owners. Only the use cases proven profitable are then industrialised, by your teams, your integrators or a partner. Here is, cause by cause, how that method neutralises the risk of failure.
We start from the logs and exports you already produce, on the platforms you administer: Azure OpenAI and in-house models, Dust, Copilot, custom agents. Installing nothing and re-instrumenting nothing: handing over exports or a read-only connection, the exact arrangement is settled in the scoping meeting. The scoping interviews with the IT department and the business units are a mandatory step: the tooling does not replace them, it makes them factual. The debrief is grounded in your reality: a map of usage tool by tool, use cases detected with estimated return on investment, waste identified, the consolidated full cost of AI and a report ready for the steering committee. The full approach is set out on our AI agency page.
The advice turns the analysis into decisions: use cases prioritised by return on investment, effort and adoption, licence arbitration at renewal time, standardisation from one tool to another and support through to the steering committee. Execution then stays with your teams, your integrators or a partner: we sell neither the agents, nor the agent platform, nor the training we recommend, and the analysis serves your decision, not a catalogue. The team works on site for companies in Ile-de-France and remotely anywhere in France.
Others measure usage, fix an interface friction or push you to build agents. Visibility alone settles nothing: it still leaves you to decide what to industrialise, what to consolidate, what to cut. We arbitrate the AI portfolio (industrialise, consolidate, cut) with a quantified return on investment and a roadmap you can present to a steering committee.
Three principles frame the analysis: an isolated network is possible; results are aggregated by team and by tool, never by person; and we read the logs you already have without installing anything on workstations. By default the analysis starts from the exports provided and your data stays in France, at Scaleway in Paris. If your data protection officer or your security policy requires it, the chain runs inside your own perimeter and the data never leaves your premises. The arrangement is settled in the scoping meeting. The exact status of the processing is validated with your data protection officer: we analyse usage, never people, and never by scanning workstations. The deliverables stay yours.
The French AI agency market advertises wide ranges, which often fund a prototype stopped at the proof-of-concept stage. Our posture is different: no pricing is published, every quote follows the analysis and is given in a meeting. What we sell is a fixed-scope engagement, analysis included, delivered within 15 days, whose scope is framed by the discovery phase. The continuous insights option (recurring re-analysis, the same deliverables kept up to date) extends the engagement.
The real budget question is not the price of the advice, it is the cost of what already runs without being measured. Across the estates we analyse, 20 to 40 % of the AI licences paid for go unused. Every renewal signed without a usage measurement carries that waste forward, while the profitable use cases stay unfunded for lack of quantified proof.
Target set and measured at your site: 30 to 50 % lower cost on the targeted usage and 2 to 4 times fewer errors. Those savings do not come out of a calculator: they come from four levers the analysis identifies use case by use case, and that your teams or your integrators then pull.
Every request goes to the cheapest model able to handle it, instead of the largest model by default.
Repeated contexts and answers are not recomputed on every call: they are served from cache.
An oversized context burns tokens and causes hallucination: the data sent is cut down to what serves the answer.
Unused licences, agents that cost without serving, duplicates of the same use case: what does not pay gets cut.
You commit to a development only after the proof: the analysis of your real usage estimates the return on investment of each use case before any investment. Only those proven profitable get built.
Book a discovery callAn Idun engagement on the generative AI platform of all the perfume and cosmetics brands.
real activation at this consulting and training firm: 33 users for 39 Claude seats.
A 30-minute discovery call.
Scoping: interviews with the IT department and the business units. A mandatory step: the tooling does not replace them, it makes them factual.
The analysis, from the logs and exports you already produce. No access to your data before signature.
The debrief: recommendations to industrialise, consolidate or cut, and a roadmap, in Ile-de-France or anywhere in France.
We analyse your real AI usage and deliver a roadmap prioritised by return on investment, ready to present to a steering committee. It all starts with a 30-minute discovery call.