Safety Audits That Took 14 Days Now Take One Hour (AES)

Imagine you run safety compliance for one of the largest energy companies in the Americas. Your team is responsible for more than 1,500 safety audits a year, across operations in multiple countries, each one touching physical infrastructure where a missed finding isn’t an inconvenience. It’s a safety failure. Now imagine that each of those audits takes roughly 100 hours of human time; real people reading through hundreds of pages of documentation, cross-referencing regulatory standards by hand, rafting protocols from scratch, with results that vary depending on who happens to be doing the work.

That was the situation at AES.

A Bet Worth Taking

Safety audits at an energy company are not the obvious place to experiment with AI. The stakes are too high, the regulatory environment too unforgiving, and the instinct in most large enterprises is to protect processes that touch physical safety with as many human reviews as possible. So when AES decided to move fast on this, it was a deliberate choice, not an accident.

Working with Google Cloud and Anthropic, the team built an agentic system on Vertex AI. The model at the center was Claude, chosen for a specific reason: these audits involve reading and synthesizing documents that can run to 400 pages, often in multiple languages across different regulatory regimes. Claude’s long-context handling and multilingual capabilities were the right fit for that kind of work. Gemini could have handled parts of the task, but this particular problem needed what Claude was best at, and through Vertex AI Model Garden, AES could access Claude with the same enterprise governance and security controls as everything else on their Google Cloud stack. No separate vendor agreement. No separate compliance review. Just the right model for the job.

The timeline was two months from initial concept to operational agents. For context, enterprise AI projects in regulated industries can take 18 months or more to reach production, and many quietly die before they get there. Two months is not normal. It was possible partly because AES already had a Google Cloud footprint, so the data infrastructure was already there, and partly because the companies signed a 10-year strategic alliance that meant the relationship wasn’t being built at the same time as the product.

What Happened

After running more than 50 audits through the system, the results came in. Audit cost was down 99%. Turnaround time from 14 days to one hour. Accuracy was up between 10 and 20%. Enough capacity was freed to double the annual audit volume with the same team.

The cost reduction gets the headline, but the accuracy improvement is the more interesting number. Safety audits exist to catch real problems. A faster audit that catches fewer issues isn’t progress, it’s risk dressed up as efficiency. The fact that accuracy went up while cost and time collapsed says something specific about where the old process was breaking down. The AI was doing the tedious cross-referencing work that humans find hardest to do consistently, and that turns out to be where most of the variance was coming from in the first place.

Why the Claude-on-Vertex Model Matters for ISVs

There is a detail in this story that’s easy to miss. AES didn’t go to Anthropic directly. They accessed Claude through Vertex AI Model Garden, which means the same governance layer, the same audit trail, and the same compliance posture that covered everything else on Google Cloud covered Claude too. One enterprise agreement. One security review. Same platform.

For ISVs building in compliance-heavy industries, this is the answer to a question that comes up in every enterprise deal: how do we use the best model for this task without creating a procurement and compliance nightmare alongside it? Model Garden’s the answer. You pick the model that fits the workload and the governance doesn’t change.

What AES Showed Everyone Else

The compliance problem AES solved isn’t unique to energy. Insurance carriers live with document-heavy regulatory reviews. Pharmaceutical companies manage submissions where a single filing can take years to prepare and the cost of an error is measured in delayed approvals. Financial institutions run audit functions with the same core pressures AES did: too much volume, too much manual work, too much variance in the output.

What’s different now’s that there’s a production reference in a genuinely high-stakes environment. AES didn’t run a cautious pilot on a low-consequence process and declare success. They went straight at safety-critical compliance work, at full scale, with a two-month build. For ISVs trying to move enterprise buyers off the fence on agentic AI for audit and compliance workflows, that story is more useful than any benchmark. It’s harder to argue that this is too risky when another company just proved otherwise on infrastructure where the stakes were as high as they get.

Want to go deeper?