Forty AI Ideas. Three Worth Funding.

An industrial services company — 1,400 staff, ₹380 Cr revenue — had forty AI ideas from a hackathon, three vendor pitches and a board request. Nobody had costed any of them. The list had been under discussion for seven months.

Three questions, in this order

Is the data here today, in a system you control, with enough history to test against? Are we allowed, once security, legal and client consent are considered? Does it beat the real current cost of doing the task by hand? The order matters. Most companies start at the third question, build a business case, and only then find the data does not exist.

What happened to the list

• 23 failed on data. Not messy data — no data, or data on a client's server we could not read continuously.
• 9 failed on permission. The compliance-certificate idea took legal eight minutes to reject: a statutory certificate issued on model output is not defensible, and the signer is personally liable.
• 5 lost to the status quo. A SQL query written in 2019 caught 94% of duplicate invoices at no marginal cost; the model reached 97% for ₹40,000 a month.

The model is the cheap part

On the largest funded use case, the cost of processing one document with a model is ₹13.90. Inference, embeddings and storage are ₹4.80 of that. Human review and exception handling are ₹7.90. So the question that decides whether the project makes money is not which model to buy — it is how many documents a person has to look at.

There is a best accuracy, and it is not the highest

Net saving peaks at 78% straight-through processing. Pushing to 93% raises engineering cost by ₹12.5 lakh a month and turns a use case earning ₹8.6 lakh a month into one losing ₹4.2 lakh. We stopped improving the model on purpose.