The news
Three enterprise AI rollout accounts published September 28 point to the same basics: Databricks meters fast access to new models, ServiceNow (NOW) plans graduated shutdowns for misbehaving agents, and Qiagen (QGEN) grounds agents in curated data. Together they show what moves AI past pilots.
Databricks, the data and AI software company, said in a September 28 blog post that it gives more than 10,000 people access to new frontier models on the day they launch (the post's headline puts the figure at 12,000 employees). Access runs through Unity Gateway, its central system for AI governance, cost control and monitoring. New models are labeled experimental and drawn from a separate experimental budget, and each user also has a monthly spending cap and a daily limit meant to stop runaway costs.
The company said it weighs private benchmarks, feedback from power users and cost per session measured from usage traces when deciding whether to promote a new model to general use. For the new Claude and GPT-6 models described in the post, it had gathered enough data within three days to move them out of the experimental budget and into general use. Its figures show why cost tracking matters: Claude Opus 5.5 cost $4.23 per session versus $5.94 for Opus 4.8, and OpenAI's GPT-6 Sol cost $2.34 versus $4.52 for GPT-5.6 Sol. But when Databricks gave a control group GPT Astra with no cost controls, the average developer spent 60% more. "Models that are marketed as frontier often aren't," the team wrote.
At Okta's Oktane 2026 conference, Bhakti Pitre, ServiceNow's vice president of AI platform security product, told SiliconANGLE's theCUBE that stopping an AI agent should not be a simple on-off decision. She described graduated responses: an agent that produces no results calls for investigation and a pause, while one offering unauthorized discounts should be stopped. According to the interview, ServiceNow can map an agent to the business processes that depend on it, uses identity security technology from Veza, which it acquired, to trim excessive permissions, and works with Okta (OKTA) to secure agent identities and session tokens.
That approach sits in ServiceNow's AI Control Tower, which the company expanded on May 5 around five functions: discover, observe, govern, secure and measure. ServiceNow said then that the product can detect an agent operating beyond its permissions and shut it down in real time.
Qiagen, the Netherlands-based life sciences company, is taking a data-first route in drug discovery. At Neo4j's GraphSummit, Iman Bhattacharya, a senior global product marketing manager, described the Qiagen Discovery Platform, as reported by SiliconANGLE. It layers Model Context Protocol (MCP) access, a standard way for AI tools to connect to data, and an agent interface on top of a knowledge base that more than 150 MD- and PhD-level experts have curated by hand for more than 25 years. Qiagen announced a collaboration with Nvidia (NVDA) on graph-based AI for drug discovery in May 2026.
The numbers
- Databricks staff with day-one access to new models
- More than 10,000
- Time for Databricks to move new Claude and GPT-6 models into general use
- Within 3 days
- GPT-6 Sol vs GPT-5.6 Sol cost per session at Databricks
- $2.34 vs $4.52 (-48%)
- Claude Opus 5.5 vs Opus 4.8 cost per session at Databricks
- $4.23 vs $5.94 (-29%)
- Rise in average developer spend with uncontrolled GPT Astra access
- 60%
- Years of manual curation behind Qiagen's knowledge base
- More than 25
Why CEOs should care
For CIOs and CFOs, Databricks' figures show that model choice is a cost decision as much as a quality one. Newer models cut its cost per session by 29% and 48% in the comparisons it published, yet uncontrolled access to another model raised average developer spending by 60%. Before opening a new model to staff, set per-user budgets, label it experimental, and decide in advance which benchmarks and cost measures will earn it wider release.
For CISOs and operations leaders, ServiceNow's point is that an off switch is not a plan. Each agent needs a named owner, a map of the business processes and permissions it touches, and a tiered response that spells out which behaviors trigger a pause and review and which trigger immediate shutdown. Ask platform vendors whether their controls cover agents built on other suppliers' tools, not only their own.
For business-unit heads and boards, Qiagen's approach is a reminder that agent accuracy depends on the data underneath. Companies with proprietary, well-maintained data, such as product catalogs, contracts or case histories, hold an advantage only if someone owns its quality. Useful questions for management: which data sources our agents rely on, who curates them, and whether each agent answer can be traced back to its source.
The bigger picture
The three cases come from different industries but point the same way: the hard part of enterprise AI is operations, not model selection. As new models keep arriving, companies that can evaluate one in days, route work to cheaper options and contain an agent quickly will absorb change more easily than those that treat each release as a fresh project. Vendors such as ServiceNow now sell these controls, which gives buyers a choice between building them in-house and buying them.
What’s next
Watch whether Databricks publishes results for the next wave of models it tests, and how ServiceNow's work with Okta on agent identities shows up in products. For most companies, the practical next step is an internal inventory: every AI agent in use, its budget, its data sources and its shutdown plan, completed before more agents are added.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
