Most enterprise AI proofs of concept don't fail because the model is weak. They fail because nobody agreed, before the PoC started, what problem it was solving, how success would be measured, and who would sign off on moving to production. The PoC ends with a good demo and no decision.
We start every engagement with a one-page problem-to-metric matrix, agreed with the customer before any build begins.
The matrix
Each row is one business problem. The columns force the questions that decide whether it ever reaches production.
| Problem | Today's baseline | Target metric | How it's measured | Data needed | Owner who signs off | Go-live condition |
|---|---|---|---|---|---|---|
| Collections calls don't reach enough borrowers | Share of due accounts contacted each cycle | Contact rate and promise-to-pay rate | Dialler and CRM records, compared with the same period last quarter | Due list, call outcomes, payment records | Head of collections | Promise-to-pay rate at or above the human team on a matched sample |
| Customer queries wait too long | Average wait and first-response time | Share of queries resolved without an agent | Ticket and call logs | Past transcripts, FAQs, policies | Head of customer service | Resolution rate holds for 4 weeks with no rise in complaints |
| Reconciliation takes days every month | Hours spent and items left unmatched | Auto-match rate and close time | Finance team time logs and ledger exports | Bank statements, ledgers, settlement files | Financial controller | Auto-match rate agreed in advance, with every exception traceable |
The examples are illustrative; every customer's rows are their own.
Rules that make it work
- One owner per row. If no one will sign off on the metric, the row comes out of the PoC.
- Baseline before build. Measure today's number first, on the same data source you will use to judge the PoC.
- Compare like with like. Judge the AI against the current process on a matched sample, not against a perfect score.
- Name the go-live condition up front. "If we hit this, we move to production" is agreed on day one, not argued at the end.
- Plan production during the PoC. Security review, data access, deployment model (cloud, VPC or on-premise) and compliance sign-off run in parallel, not after.
- Time-box it. A PoC that runs past its agreed window without hitting its metric gets a decision, not an extension.
Why this matters in regulated industries
Banks, NBFCs and insurers add two gates most PoCs ignore until too late: information security review and regulatory compliance (RBI outsourcing rules, the DPDP Act). Putting both in the matrix as named owners, with their own go-live conditions, removes the most common reason a successful PoC stalls for months.
Use it yourself
Copy the table, fill one row per problem, and get each owner to agree their row before you start. If a row can't be filled in, that problem isn't ready for a PoC yet.