Spoke to a company last week with >50 MCP's running across their org → token cos

By Laurens Nys · October 27, 2026 · Curated by George's Blog

Spoke to a company last week with >50 MCP's running across their org

→ token cost through the roof

→ no clue on ROI numbers

Bet tons of companies are in this trap atm

But connecting tools does not equal deploying AI.

Tons of companies seem to just;

- ship an MCP for every tool

- let everyone build their own agents

- have nobody running evals

- get on some hackathons meant to "find use cases"

- pay third parties to give "prompting workshops"

- set up some Claude jobs

...

These are not how you deploy AI.

Deploying AI is:

- mapping how the work actually runs, and baselining the time, errors and cost before you change anything

- defining a clear shared vocabulary of the business (aka ontology)

- redesigning the process from first principles with AI at the center

- getting teams involved in deployment (FDE's work)

- testing on real examples and keeping the evals running, so you can put cost per run next to time saved

Granted, the first list is easier (and fine to start with).

But it’s also how you end up with 50 MCPs and no ROI number.

At Ortelian, we work on the second.