Generative AI use cases that survive the pilot

Generative AI demos are easy to build and easy to admire. Getting one into daily use is harder. Pilots stall when nobody owns the result, when the data is not ready, or when the answers are good enough to impress but not reliable enough to trust. Choosing the right use cases up front avoids most of that.

Score use cases before building anything

We score each candidate on five questions. A use case that fails two of them rarely makes it to production.

1. Is there a named owner? Someone in the business who wants the outcome and will change a process to get it.
2. Is the knowledge available? The documents, records or systems the model needs must exist, be current and be accessible with the right permissions.
3. Can the benefit be measured? Time saved per case, faster first response, fewer escalations, higher conversion. Pick one primary metric.
4. What happens when it is wrong? Drafting an internal summary is low risk. Giving customers binding answers is not. Start where mistakes are cheap and a human reviews the output.
5. Does it fit into an existing workflow? Assistants inside the tools people already use get adopted. Separate portals often do not.

Use cases that tend to work

• Service desk and contact center assistants that draft replies from approved knowledge articles for agents to review.
• Document extraction and summarization for invoices, contracts, claims and medical or legal records, with human checks.
• Internal knowledge search across policies, runbooks and product documentation, with sources cited in every answer.
• Sales and account preparation that summarizes CRM history, open tickets and recent communications before a meeting.
• Engineering productivity, such as test case generation, code review assistance and log analysis.

Build evaluation in from the first week

The difference between a pilot and a product is evidence. Before tuning prompts, write an evaluation set with your subject experts: real questions, expected answers and the sources that support them. Run it on every change. Track accuracy, groundedness, refusal behavior, latency and cost per request. Without this, every improvement is an opinion.

Ground answers in approved sources

Retrieval-augmented generation, where the model answers using documents retrieved from your own knowledge base, remains the most practical pattern for enterprise use. It works well when:

• Documents are current, owned and cleaned of duplicates.
• Access permissions from the source systems are respected at query time.
• Answers cite their sources so users can check them.
• The assistant says it does not know when the sources do not cover a question.

Guardrails and governance

Governance does not need to slow things down if it is designed in:

• Filter inputs for sensitive data and prompt injection attempts.
• Filter outputs for policy violations, personal data and unsupported claims.
• Log prompts and responses securely for quality review, with retention limits.
• Keep a register of AI systems, their owners, data sources and risk ratings.
• Tell people when they are interacting with an AI system. The EU AI Act's transparency rules make this an obligation for many systems used in the EU, and it is good practice everywhere.
• Check personal data handling against the DPDP Act, GDPR and your own policies.

Watch the unit economics

Costs scale with usage, context length and model choice. Measure cost per task from the start, cache repeated results, use smaller models where they perform well enough, and set budgets with alerts. A use case that saves ten minutes but costs more than the time it saves will not survive a finance review.

Plan the change, not just the software

Adoption is a people problem. Train users on what the assistant is good at and where it is not, collect feedback inside the tool, and publish improvements. Measure the primary metric before launch so the result can be compared honestly.

A simple path from idea to production

1. Discovery: collect and score use cases, pick two or three.
2. Proof of value: build a thin slice with real data and an evaluation set.
3. Pilot: release to a small group inside their normal workflow.
4. Production: add monitoring, guardrails, support and cost controls.
5. Scale: reuse the platform and patterns for the next use cases.

Our data and AI team runs this process end to end, from discovery workshops to production platforms. Tell us which processes you are considering.