Eye on AI: What moved at Brainstorm AI SF - and what it means for your roadmap
This year's Brainstorm AI in San Francisco had a different feel. Fewer hypotheticals, more proof-of-concepts going live. Teams are piloting AI agents in real workflows, and the hot questions were less "what can AI do?" and more "how do we govern and orchestrate agents alongside people?"
The theme underneath it all: build useful systems, reduce risk, and move fast without breaking critical operations.
AI agents are moving from demos to real work
Across industries, companies are deploying AI agents in specific, bounded tasks. Think customer operations triage, finance review prep, or internal developer support. The gains are real, but the orchestration challenge is bigger than most expected.
- Define clear handoffs between agents and humans.
- Instrument everything: prompts, decisions, escalation paths, audit logs.
- Start narrow, measure rigorously, then scale.
Risk posture: "fast follower" beats "first mover" in critical infrastructure
Exelon's leadership captured the mindset for any high-stakes operator: security and reliability come first. When "wrong" can mean the lights go out, you let others take the early hits, then adopt with evidence.
- Run controlled sandboxes for months, not weeks.
- Gate model changes behind formal change management.
- Separate experimentation from production with strict data and network boundaries.
No, software engineering and accounting won't be fully automated
Cursor's Michael Truell put it plainly: code compilation is speeding up, but humans still need to decide how the software should work. That's the job. In finance, DataSnipper's Vidya Peters echoed the same: qualified accountants remain essential, especially in regulated workflows.
Translation: expect fewer keystrokes, more judgment. And in many enterprises, domain-specific apps will beat general-purpose models because compliance and context win deals.
Data centers: going where the power is - and bringing their own
Speakers on the "new geography of data centers" agreed: build near abundant power today, and plan to generate on site tomorrow. As inference loads outgrow training, more compute will need to sit closer to users - but most cities are already power constrained.
One infrastructure leader summed up the bottleneck: we can move tokens over fast fiber, but we still need electrons. Expect creative power procurement, on-site generation, and new partnerships between utilities and AI operators.
Build vs. partner: why a joint venture can beat a services contract
One telecom leader explained why they created a JV with a major consultancy instead of a standard services deal. The JV unlocked speed, shared incentives, and access to talent and hubs they couldn't stand up internally. If your AI program is stuck, check your incentives and ownership model.
The future of enterprise AI is hybrid
Leaders from PayPal and Nvidia were aligned: open-source models give control and fine-tuning flexibility; proprietary APIs offer performance and managed risk. Most enterprises will run both.
- Use open models for private data and customization.
- Use proprietary APIs for scale and reliability.
- Standardize evaluation so you can swap models without drama.
People and culture decide whether pilots actually stick
Responsible AI leaders stressed what many teams learn the hard way: tech is the easy part. If you don't prepare people and org charts, your pilots stall. As one exec put it: "Don't just think about technology-think about people and the culture. It is so paramount."
Another useful framing: stop calling AI "a tool." Treat it like a capability that reshapes process design, incentives, and roles.
Healthcare's line: clinicians first, patients later
Clinician-facing AI is gaining ground; patient-facing agents still raise safety, liability, and UX questions. A veteran clinician pointed out that patients have always brought "second opinions" into the room - whether from Google, ChatGPT, or a neighbor. AI will be powerful for rare and edge cases, but the patient-physician dynamic hasn't fundamentally flipped yet.
AI news to watch
Disney invests $1B in OpenAI and licenses its IP
The headline is big: capital plus content. Expect tighter pipelines between premium IP and synthetic media tools, stricter guardrails on likeness and licensing, and a wave of AI-native entertainment experiments. For media and marketing teams, start drafting policies for rights, provenance, and watermarking before these tools hit your stack.
OpenAI debuts GPT-5.2 to counter "falling behind" concerns
OpenAI launched GPT-5.2 to push past chatter that rivals were catching up. The pitch: better quality, steadier reasoning, and improved efficiency across typical enterprise tasks. Developers should plan quick A/B tests against their current baseline and update eval harnesses to measure reliability, latency, and cost per task.
Oracle stock stumbles
Markets reminded everyone that AI infra is a long game. Watch the signals that matter: cloud bookings quality, GPU/AI accelerator supply, and margins on AI workloads. For buyers, multi-cloud and capacity hedging look smarter than ever.
What to do next
- Stand up an AI agent pilot in one process with clear guardrails, SLAs, and human escalation. Measure time-to-resolution, error types, and rework.
- Create an orchestration plan: which tasks agents own, which humans approve, and how exceptions get handled. Log every decision.
- Adopt a hybrid model strategy. Build a single evaluation suite so you can swap models (open and proprietary) with minimal regression risk.
- Review your infrastructure plan. If inference is customer-facing, explore edge or metro-adjacent colos and assess power constraints early.
- Strengthen Responsible AI: data retention, PII redaction, model isolation, red-teaming, and change management. Consider frameworks like the NIST AI RMF.
- Reskill your teams. Prioritize prompt patterns, agent design, retrieval, and evaluation. If you need a starting point, browse role-based options at Complete AI Training - Courses by Job.
- In healthcare and other regulated fields: favor clinician/operator-facing AI first. Build patient/customer agents only after you have auditability, escalation, and liability coverage.
- Explore strategic partnerships or JVs if your roadmap needs speed, shared incentives, and vetted talent you can't hire quickly.
Bottom line
AI is moving from novelty to operations. Keep your scope tight, your evals honest, and your people empowered. Do that, and you'll compound advantages while others debate the hype.
Your membership also unlocks:
