The ideaROI across a portfolio of AI apps
An organisation does not run one AI feature — it runs a portfolio. This dashboard models each application's full monthly cost and the business value it delivers, and rolls them into a blended ROI. Click any row to drill into what actually drove the cost and what value is assumed — the "double-click" on every element.
PortfolioApplication-by-application
| Application | Model | Cost /mo | Value /mo | ROI |
|---|
Illustrative numbers — swap in your own volumes, models and value assumptions. The point is the structure: token cost is usually small; build + maintenance dominate, and value must be named, not assumed.
At a glanceThe cost shape, and the gap it has to clear
Two readings of the table above, drawn from the same arithmetic. The first says where the money goes; the second says whether it comes back.
Monthly cost per application, split the four ways the method below describes. Model tokens are the thin band against the baseline — between 3% and 16% of each application's bill; build and maintenance are nearly all the rest. Hover a segment for its share, or click the row in the table above for the same numbers written out. These are the page's illustrative volumes and prices, so the proportion is a shape, not a measurement.
Cost and value on one shared dollar axis — never two — so the distance between a row's two bars simply is its net, printed at the end of the row. Six applications clear their cost several times over. On the pilot agent the two bars are all but level, and the cost bar is the longer one: $36,840 of cost against $35,000 of value, the single row in the portfolio that runs negative. Same illustrative inputs as the table above: a shape, not a measurement.
MethodWhat goes into each number
Cost = the whole iceberg, not just tokens
- Model tokensinput × price × volume (minus caching) + output × price × volume.
- Platformvector DB, observability, gateway, storage.
- Buildthe one-time engineering to ship it, amortised over ~12 months.
- Maintenancethe fraction of an engineer keeping it healthy (prompts, evals, drift).
Value = named, not hand-waved
- Labour deflected(tickets, hours saved × loaded rate), revenue influenced, or risk avoided.
- ROI= (value − total cost) ÷ total cost. Payback = build ÷ monthly net.
- Flag the losers: the pilot agent here runs net-negative — high loops, heavy maintenance, unproven value. That is the dashboard doing its job.