Engineering AI is half the work. Justifying it inside a large enterprise is the other half. The frameworks that work:
Identify the value lever
- Cost reduction. Automate tasks engineers / analysts / agents currently do. Easiest to measure.
- Revenue uplift. Conversion, retention, cross-sell, faster sales cycle.
- Risk reduction. Fewer compliance violations, fewer escalations, faster anomaly detection.
- Capability creation. Things you couldn't do at all before — e.g., 24/7 multilingual support across 50 languages.
Anchor the case on one of these. Don't fluff.
ROI math template
Annual ROI = (Annual benefit - Annual cost) / Annual cost
Annual benefit:
= (tasks_per_year × saved_minutes × loaded_hourly_rate) +
(extra_revenue or avoided_loss)
Annual cost:
= LLM API + infra + engineering FTE + ongoing eval/governance
The cost line that gets forgotten: ongoing eval and maintenance. Budget 20-30% of build cost annually for upkeep.
Pilot → production journey
1. Discover (1-2 weeks)
- Pick a real use case with a single owner.
- Define success metric. One number.
- Define guardrail metrics — what must NOT regress (cost, latency, quality, compliance).
2. Prototype (2-4 weeks)
- Build the smallest thing that solves the use case.
- Eval on labeled set, real users in shadow mode if possible.
- Decide go/no-go based on success metric vs threshold.
3. Pilot (1-3 months)
- Deploy to subset of real users.
- Real metrics, real costs, real failures.
- Critical: measure user adoption — a feature people don't use is zero ROI even if it works.
4. Productionize (1-3 months)
- Hardening: observability, eval pipeline, guardrails, runbook, on-call.
- Cost optimization: caching, cascading, prompt prefix caching.
- Compliance review: governance, audit, model card, sign-offs.
5. Scale & operate (ongoing)
- Quarterly model audit.
- Monthly cost review.
- Continuous eval drift monitoring.
Metrics that resonate with execs
- $ saved / $ spent. Direct ROI.
- Time-to-value. Weeks from idea to production.
- % of relevant tasks now AI-handled. Penetration.
- User satisfaction / NPS delta. For user-facing.
- Risk metric reduction. For risk-driven cases.
Avoid: token counts, model leaderboard ranks, prompt sophistication. Execs don't care; they care about outcomes.
The traps to avoid
- Boil the ocean. "We're going to put AI in every workflow." Pick one. Win. Then expand.
- Tech-first framing. "We need to use GraphRAG" is not a business case. "We need to answer multi-hop compliance questions in <5s" is.
- Ignoring the human change cost. Workflow changes are change-management projects, not just engineering.
- No baseline. If you can't measure the current process cost, you can't measure savings.
- Skipping eval. "It looks good in demos" doesn't survive scale. Bake eval in from day one.
A one-page exec brief skeleton
- The problem — one paragraph, business framing.
- The proposed approach — one paragraph, no jargon.
- Success metric and target. One number.
- Cost & timeline. Engineering + LLM + ops, three milestones.
- Risks & mitigations. What could go wrong, how you'll handle it.
- What you need. Funding, headcount, access.
The execs who fund your work read 100 of these a year. Make yours readable in two minutes.