Pick a business area and the work it needs doing, and this prices it three ways: build it alone, adopt a build that already works, or join a shared service. The workloads are drawn from Inland Revenue's own published statements about its use of AI, priced against 3,895 systems four governments have published.
Each business area has a handful of workloads that come up again and again. Pick the one you recognise and the model opens on it, already set up — volume, time saved and the closest published systems.
Whatever a system is called, underneath it is doing one of a short list of jobs — reading documents, answering questions, sorting a queue, checking a claim. The same handful turn up inside almost every government agent, which is why they are worth building once. Pick one to see every published system doing that job, and what it costs.
See all patterns ›The full case base and the current selection, as they sit behind this page. CSV opens in Excel or Sheets.
The pattern is read from the published description. Complexity and replication follow from the pattern, and together they set the cost model below.
Decomposed on the FinOps cost schema. Confidence reflects the strength of the source, not the precision of the figure.
Both routes are priced the same way: people × months × working days × day rate. Nothing here is a constant — change any figure and the totals move with it.
| Line | Working | Amount |
|---|---|---|
| Delivery day rate | per day blended across roles | |
| One person-month | working days × the day rate | |
| Shared model · first build | people months | |
| Conventional · delivery | people months | |
| Conventional · RFP and contracting | people months | |
| Conventional · total | delivery plus procurement | |
| Each agency after the first | % reusable % reuse factor % local adaptation | |
| Shared model · annual maintenance | % of build a year % shared agencies | |
| Conventional · annual support | % of build a year per agency |
Under conventional procurement each agency funds a full build. Under the shared model the first build produces a reusable pattern which subsequent agencies inherit.
Three steps. Pick the business area, pick the work, and see what it costs to build alone, adopt a build that already works, or join a shared service.
Classified by the function the system performs, the case base reduces to a small number of recurring patterns. Cross-agency reach indicates replication potential: the more agencies already running a pattern, the lower the expected cost of each subsequent build. Select a pattern to carry it into the cost model.
Reach is the share of agencies across all four countries running at least one case of the pattern. Case counts come from what the governments published. Complexity class and replication character are Tilicho classifications applied to the observed pattern.
The calculation is set out below in four stages, using the parameters currently selected in the cost model. Adjusting any input there updates these figures. Each input is stated with its source and an assessment of confidence.
Total task volume is reduced at three stages: suitability for automation, staff take-up, and acceptance of the output. Benefit accrues only at the final stage.
Full annual running cost by layer. Token charges are one component of seven.
The lead agency funds the initial build. Subsequent agencies fund adaptation only, at a proportion determined by the replication factor.
Case counts and token prices are published and verifiable. Delivery and volume figures are planning estimates and should be substituted with agency data before commitment.
Shared government AI platforms elsewhere, assessed against the same architecture. Only elements confirmed by a government, audit-office, parliamentary, intergovernmental or peer-reviewed source are recorded. Where a figure is not published in such a source, that is stated rather than estimated.
Ranked by how much of the architecture is confirmed in a qualifying source, not by size or ambition.
One programme in the survey publishes both an adopter count and a before-and-after time measurement.
The complete set of platform cost and funding figures located in qualifying sources.
Metrics sought across ten jurisdictions and seven enterprise programmes, and whether any qualifying source reports them.