Methodology

How the Lab calculates

The definitions and formulas were written down before any screen was built, so every task gives answers on the same basis. This page summarises them; the full method sits with the Lab's source.

Agreed, lab-method-0.1. Drafted on 13 October 2026 and agreed on 14 October. A change to any term or formula gets a new version, and every export prints the version it used.

Three costing paths

A use case is costed on the path that matches how it is bought or built. All three produce the same result shape, so any two can be compared, and a serving choice can replace a model step's cost in a custom workflow but is never added on top of it.

Path	Use when	Cost drivers	Headline unit
Packaged tool	Buying seats or a SaaS tier, such as an assistant or embedded AI	Seats and price, administration, enablement, review	Active user or accepted outcome, chosen by you
Custom workflow	Building on model APIs and services	Model and service steps including retries, human effort, fixed operations, build	Accepted outcome
Serving choice	Deciding between pay-per-token and dedicated capacity	Volume, token price, GPU rate, capacity, utilisation, people, switching cost	Cost per million tokens, handed to a workflow

One unit of work

Every cost is divided by one unit of work, set once per scenario, such as a ticket resolved or a document processed. The default is the accepted outcome: a completed request whose output is accepted for use, after any human review the scenario already pays for.

Two scenarios compare only when they share a unit. When a denominator is zero the result reads “Undefined”, never £0 or 0%.

Cost and value

AI cost is deducted once. Net resource value already has the full monthly TCO taken off, so nothing downstream subtracts it again. Payback uses run cost, so build is not counted twice. Resource value counts staff time at loaded cost, so it is not cash unless that spending actually falls. The cash view counts only the share of the baseline you stop paying, such as contractor or licence spend, against the new payments the AI brings; people time in the scenario is treated as existing staff, and one-off build as cash paid.

Monthly TCO: variable consumption + human effort + fixed operations + amortised build
Monthly run cost: monthly TCO − amortised build
Cost per accepted outcome: monthly TCO ÷ accepted outcomes
Gross benefit: baseline unit cost × accepted outcomes
Net resource value: gross benefit − monthly TCO
Resource-value payback: build cost ÷ (gross benefit − monthly run cost)
Cash saving: gross benefit × share of the baseline you stop paying
New monthly payments: model, service, seat or serving charges + fixed running costs
Net cash a month: cash saving − new monthly payments
Cash payback: build cost ÷ net cash a month

First-month outlay, which includes the whole build cost, is shown beside monthly TCO and never added to it. It is cash only where its lines are new payments rather than existing staff time.

Forecast, measured and booked

Every value figure carries one status. A scenario is a forecast, and it changes status only when a separate record is created and linked to it, never by editing a label. Totals never mix statuses. Every figure a cost scenario produces is a forecast; measured and booked figures come only from workspace records.

Status	Comes from	Needs
Forecast	A scenario	Assumptions with a stated source
Measured	An observation record	Baseline, observed value in the same unit and period, evidence level, method
Booked	A measured figure accepted in a named record	Measured, evidence at E2 or above, an eligible benefit type, the record named

Evidence levels run from E0 to E4, as defined in the Hub's value workbooks. They are a gate, not a multiplier: nothing below E2 goes into a monetised total.

Sources and unknowns

Every price or benchmark carries one of five labels, a source and a checked date. An unknown price stays blank and blocks the result with a message naming it; it is never treated as zero.

Public list price: From a public price page, with its URL and checked date
Quote: From a supplier quote, with the supplier and date
Measured: From your own telemetry or invoices
Estimate: Your judgement. Allowed, and always shown as such
Illustrative: Invented for an example, and flagged wherever it appears