AI in Enterprise Data Work — What Actually Ships

There's a gap between the AI conversation happening on Twitter and the AI work that actually ships in enterprise environments. Having worked both sides — building local AI infrastructure and deploying data solutions for large organizations — here's what I've seen actually deliver value.

The Pattern That Works

The highest-ROI AI applications in enterprise data work aren't the flashy ones. They're:

• Automated analysis of existing data assets. Most organizations are sitting on analytics platforms (GA4, Adobe, CRM systems) generating data nobody has time to synthesize. AI that reads the data humans already collected and surfaces patterns is immediately valuable.
• Document-to-insight pipelines. Meeting transcripts, stakeholder emails, requirements docs — turning unstructured text into structured action items, risk flags, and decision logs.
• Multi-model verification. Running the same analysis through different models or prompting strategies and comparing outputs. Catches hallucinations and builds confidence in recommendations before they reach stakeholders.

What Doesn't Ship

In my experience, these consistently stall:

• AI replacing human judgment on stakeholder-facing decisions. The trust isn't there yet, and shouldn't be. AI augments the analyst; it doesn't replace the meeting where humans align on what the data means.
• End-to-end automation without human checkpoints. Every pipeline that skips human review eventually produces something embarrassing. Build the checkpoints in from the start.
• Model choice as a strategy. The model matters less than the data quality, the prompt engineering, and the human review process. I've seen better results from a well-prompted smaller model with clean data than a frontier model fed garbage.

The GTM Angle

For data teams trying to prove AI value to business stakeholders, the playbook is:

1. Start with data they already trust. Don't introduce a new data source and AI analysis simultaneously. Use their existing GA4, their existing CRM, their existing reports.
2. Show the delta, not the output. "Here's what AI found that you weren't seeing" is more compelling than "here's an AI-generated report."
3. Async first. Send the insight with a screenshot. Don't book a meeting to show an AI demo. Meetings should be for decisions, not demonstrations.

The organizations that get value from AI are the ones that treat it as an acceleration layer on existing workflows — not a replacement for the workflows themselves.