OpenAI - AI in the Enterprise

OpenAI recently released a set of case studies on the ROI of AI in the Enterprise. It's hard to separate hype from hard value in the current AI gold rush. Here's my distilled take.

1. Start with Evals – Morgan Stanley's Playbook

Morgan Stanley launched with three rigorous evaluation tracks before production. The payoff? 98% of advisors now use an AI assistant daily, and searchable research archive coverage leapt from 20% to 80%.

2. Embed AI into Products – Indeed's "Why This Job?"

Indeed fused AI into the core job-seeker flow. Personalized "Why this job?" blurbs raised application starts by 20% and boosted hiring success another 13%. They fine-tuned a leaner model to hit the same lift with 60% fewer tokens.

3. Invest Early and Iterate – Klarna's Support Transformation

Klarna wired GPT-4 into customer support across 23 markets and 35 languages. The assistant now closes two-thirds of all chats, shrinking resolution time from 11 minutes to just 2. The shift frees capacity equal to roughly 700 agents and adds an estimated $40 million in profit.

4. Customize Ruthlessly – Lowe's Catalog Cleanup

Lowe's fine-tuned GPT-3.5 on supplier feeds. Tag accuracy jumped 20%, while error detection rose 60%. Generic models get you to baseline; custom data unlocks the edge.

5. Put AI in Experts' Hands – BBVA's Bottom-Up Surge

In five months, employees spun up 2,900+ custom GPTs. A single legal-domain GPT now fields about 40,000 compliance questions annually.

6. Unblock Developers – Mercado Libre's Verdi Framework

17,000 developers now use Verdi to build AI features safely. Vision models tag product images, enabling 100× catalogue expansion, while fraud-detection accuracy hovers near 99%.

7. Set Bold Automation Goals – OpenAI's In-House Engine

An internal platform now runs hundreds of thousands of support tasks each month. By stripping away rote work, the team redirects energy to higher-impact problem-solving.

Where to Go from Here

1. Quantify before you scale. If you don't measure baselines and incremental lift, you're flying blind.

2. Build AI into the moment of need. Users won't adopt an extra step.

3. Custom beats generic—always. Your data is your moat.

AI's strategic window is wide open, but advantage accrues to those who execute with discipline.