Claude rarely arrives through a procurement plan. It arrives through one engineering team that adopts it, then through projects and skills each person builds alone. Six months later the technical leadership holds premium seats, an invoice that follows the models being called and no view of what any of it produces downstream.
Claude Code inside engineering teams, projects that carry a shared context, reusable skills and instructions, connectors into the information system through the MCP protocol, and uses outside engineering across support and business functions. Measurement gives real adoption per unit, the depth of each usage, and the assets that stayed private while several teams rebuilt them in parallel.
Fit with the working environment, quality on the intended task, governance and administration, cost per seat against real use, sovereignty and hosting. Idun Group publishes no ranking between platforms: we publish the grid a leadership team uses to decide for itself, filled in with your own data rather than a market average.
Two seat tiers rather than one, with Claude Code reachable only from the premium seat on team plans. Usage that can be billed beyond the seat at programming interface rates. Ceilings that flip consumption to pay as you go. The model called, which changes the invoice on an identical task. Spend concentrated on a few people. And the hidden cost of rework, which appears on no invoice. Across the estates we analyse, 20 to 40 % of paid AI licences go unused.
We start from the logs and exports you already produce, across Claude and the rest of the estate, read only, with nothing installed on workstations and no analysis of your source code. A vendor console covers its own perimeter by construction and cannot compare one use case from one tool to another. Interviews with IT and the business units are mandatory. We analyse usage, never people.
The console is one of the most complete on the market. It exposes weekly and daily active users, seat utilisation, usage and cost by group and by person, a spend tab broken down by model, Claude Code metrics including accepted lines, edit acceptance rate and pull requests merged with and without assistance, mapping to identity groups, a programming interface refreshed every four hours and an OpenTelemetry export. It does not expose what happens after the output, since a merged pull request is not a feature in service. It does not expose competing assets, the same need written in several contradictory versions. The time saved it displays is a vendor estimate, not a measured business result. It gives no intent, nothing outside the Anthropic perimeter, and no history at all when the telemetry export, which is opt in, was never switched on.
At Inskip, a consulting and training firm, the analysis established 85 % real activation, 33 users for 39 seats, with the three heaviest users returning 12 to 25 times their seat and 74 ceiling hits on the most productive account while 12 seats produced no trace. The finding that moved things was invisible from the inside: an assessment framework circulating by copy and paste, carrying two incompatible scales in the same message, with five competing versions of the same asset. The limit, stated up front: the export had a rolling thirty day retention while the rollout had been running five months, so one month in five was measured.
Idun Group is an independent consultancy and resells no platform. We have to be able to conclude that the tool is not worth what it costs at your site. More than 40 AI assessments in under a year across more than 10 large enterprise clients.
