The Mid-Market AI Waste Index, Q3 2026

This index combines two sources: the spend ledgers we build during Reality Check engagements, and self-reported runs of our public AI Waste Calculator. Respondents answer in bands; we compute on midpoints; unit costs are our own field estimates. The sample skews to companies between ₹200 crore and ₹5,000 crore in revenue, because those are the companies that call us.

This is field data, not census data. It is directionally reliable and precisely wrong, and we publish the limitations because a benchmark you cannot interrogate is marketing.

The median pilot-to-production rate across the sample is 22%. Sector variation is narrower than expected: GCCs lead at roughly 30%, professional services trails at roughly 15%. The causes respondents could not name, the ledgers usually could: data readiness and integration, not model quality.

The useful way to read this: four out of five pilots in the median portfolio are not failing to work. They are failing to arrive.

Internal time is the single largest line in most ledgers we build, and the one no finance system captures. A mid-market company with four people nominally on AI work is spending roughly ₹2 crore over two years in loaded cost, almost none of it attributed to any initiative, because the initiatives themselves were never costed.

Where no documented business case exists, we estimate 25-40% of licence spend is shelfware within eighteen months of purchase. The pattern is consistent: tools are bought ahead of use cases, the champion leaves or the pilot stalls, and the renewal auto-pays because nobody owns the cancellation.

Four behaviours separate the top decile, and none of them are exotic. They write a baseline before they build. They name an owner for every system in production. They evaluate against a golden dataset on a cadence, not on complaints. And they kill initiatives on schedule: the kill rate in the top decile is higher than the median, not lower.

Self-reported inputs, midpoint arithmetic, our own unit costs, and a sample drawn from companies already worried enough to measure. The index will overstate waste for disciplined organisations and understate it for the rest. Treat your own ledger as the only authoritative version.