The short version
OpenAI says enterprise use of ChatGPT has entered a new gear: weekly messages in ChatGPT Enterprise rose roughly 8× year over year, and workers report saving 40–60 minutes per day. On the same day, BNY Mellon said it’s bringing Google Cloud’s Gemini Enterprise—powered by the new Gemini 3 models—into its in‑house Eliza platform so employees can build and run compliant AI agents at scale.[^1]
Why this week marks an enterprise AI inflection
Two signals landed on December 8, 2025:
- OpenAI published its first State of Enterprise AI report with hard usage telemetry: enterprise message volume up ~8× year over year, a 19× rise in use of structured workflows like Projects and Custom GPTs, and a ~320× increase in reasoning tokens consumed per organization—evidence that companies are shifting from casual chat to embedded, repeatable processes.
- BNY (BNY Mellon) and Google Cloud announced that Gemini Enterprise is being integrated into BNY’s enterprise AI platform, Eliza, with the bank continuing to leverage Google’s newest models—including Gemini 3 and Veo 3—to power agentic workflows for research, onboarding, reconciliations, and more. The move follows a year in which BNY put over a hundred “digital employees” to work and ramped to 100%+ agent usage in multiple teams.
These developments compress the distance between experimentation and durable value. OpenAI’s numbers show depth of use rising; BNY’s deployment shows how “agentic” systems are crossing from pilots into production in a heavily regulated industry.
What OpenAI’s 8× surge really tells us
OpenAI’s dataset (usage from enterprise customers plus a survey of 9,000 workers across ~100 firms) points to three patterns:
- Depth beats dabbling. Message counts rose sharply, but the bigger story is the 19× rise in structured workflows and the steep growth in reasoning tokens—signals that teams are encoding processes (research, QA, reviews, code changes) rather than tossing ad‑hoc prompts into a chat window.
- Time savings compound. Across roles, 75% of workers report improved speed or quality, with typical time savings of 40–60 minutes per day and heavy users reporting 10+ hours weekly. That’s unused capacity you can redirect to analysis, outreach, or customer conversations.
- Platform competition is heating up. Transaction data from the Ramp AI Index shows OpenAI remains the leader in paid adoption by U.S. businesses, but Anthropic is closing some ground—useful context as CIOs plan multi‑model strategies.
<<callout type="note" title="A quick sanity check on ‘8×’"> OpenAI’s 8× refers to weekly message volume in ChatGPT Enterprise year over year, not total revenue or seat count. It’s paired with telemetry showing a shift to Projects/Custom GPTs and more intensive use of reasoning models—useful leading indicators of value creation.
Why BNY Mellon’s Gemini 3 move matters
BNY has spent the past two years building Eliza as a governed platform that lets any employee compose solutions from approved models and data. By integrating Gemini Enterprise, the bank is:
- Equipping its workforce with a no‑code agent workbench (Agent Designer) and out‑of‑the‑box agents like Deep Research, wired into enterprise data sources under role‑based access control.[^1]
- Adding the latest multimodal reasoning from Gemini 3 for tasks like market synthesis and document triage.
- Continuing a multi‑vendor model strategy: a multiyear OpenAI agreement earlier this year brought ChatGPT Enterprise and API access to Eliza; now Google augments Eliza’s agentic layer.
Operationally, the bank is already in production at scale. CEO Robin Vince said in October that BNY has 117 AI solutions live and 100+ agents assisting in areas like payment validation, code repair, and client onboarding.
Reporting this fall indicated 96–98% of employees are trained on Eliza.
The takeaway: this isn’t a lab exercise. It’s a blueprint for how highly regulated firms can scale agents—tight data boundaries, human managers for digital workers, and explicit oversight.
The agentic shift: from copilots to doers
“Agentic AI” refers to systems that can plan, call tools, and execute multi‑step work with guardrails. Banks like BNY are pairing this with identity, access control, and audit so agents can carry a workload alongside humans.
Agentic building blocks (Dec 2025)
| Capability | OpenAI (Dec 2025) | Google Cloud (Dec 2025) |
|---|---|---|
| Web‑scale autonomous research | Deep Research (agentic, asynchronous web research) | Deep Research agent in Gemini Enterprise |
| Real‑world task execution | Operator → ChatGPT agent mode | Agent Designer, visual flow builder in Gemini Enterprise |
| Latest models in enterprise | GPT‑5 family, o‑series reasoning models (per OpenAI roadmap) | Gemini 3 Pro (Preview) in Gemini Enterprise [^1] |
| Data grounding & connectors | Model Context Protocol connectors (e.g., LSEG data in ChatGPT) | Workspace/Microsoft 365, Salesforce, SAP, BigQuery connectors in Gemini Enterprise |
For risk, both ecosystems emphasize enterprise controls. Google has added memories, sources panels, and admin toggles in Gemini Enterprise; OpenAI centralizes governance through ChatGPT Enterprise admin controls and compliance APIs. Heavily regulated adopters (like BNY) further layer internal risk reviews and human supervision of digital workers.
The enterprise playbook: how to capture value without chaos
The headlines are exciting, but value hinges on execution. Four practical steps:
- Start with “controlled complexity.” Pick 2–3 workflows with measurable lag (e.g., research synthesis, QA, reconciliations). Encode them as agents with clear inputs/outputs and a human sign‑off step.
- Treat data access as product design. Ground agents in the right collections (e.g., SharePoint/Drive, CRM, ticketing) with least‑privilege policies, redaction, and kill switches. Instrument every tool call.
- Make depth your KPI. Track agent‑run rate, hand‑back rates, and cycle time—not just usage. Favor models that reduce rework and context‑switching over raw benchmark scores.
- Build a portfolio, not a monoculture. Multi‑model strategies are winning. Use OpenAI for certain reasoning‑heavy flows and Google for multimodal or workflow‑oriented agent stacks; keep options open as costs and capabilities shift.
<<callout type="tip" title="A one‑page agent design checklist">
- Define mission, scope, and guardrails; document failure modes and escalation.
- Map data sources, tool permissions, and audit trails.
- Establish human‑in‑the‑loop points by risk tier.
- Add evaluation harnesses (golden tasks, sandbox environments, red‑team scripts).
- Measure ROI with before/after cycle‑time and error‑rate deltas.
What to watch next
- Enterprise bake‑offs. With Gemini 3 entering enterprise and OpenAI accelerating agent features, expect faster iteration and more cross‑vendor stacks.
- On‑prem and sovereignty. Google’s Gemini on Distributed Cloud widens deployment options for restricted data.
- Adoption vs. value. Even amid eye‑popping spend forecasts—Gartner pegs 2025 AI spend near $1.5T—CIOs still report friction moving pilots into production. Treat the BNY pattern (governed platform, widespread training, measured rollout) as a template rather than an exception.
Sources
- OpenAI — The State of Enterprise AI (Dec 8, 2025): Report overview
- BNY + Google Cloud — Press release (Dec 8, 2025): BNY integrates Gemini Enterprise into Eliza and PR Newswire
- Google — Gemini 3 announcements: Product blog and Enterprise release notes
- BNY — What is Eliza: Artificial Intelligence at BNY; culture & adoption: Unlocking Value with BNY’s Enterprise AI Platform
- Adoption telemetry — Ramp AI Index (Oct–Nov 2025): Overview and October write‑up
- BNY scale metrics: PYMNTS coverage; digital workers trend: Wall Street Journal
- OpenAI agent features: Deep Research and Operator → ChatGPT agent mode
- Ecosystem signals: LSEG data in ChatGPT via MCP connector (Reuters)
- Market context: Gartner: AI spending $1.5T in 2025
