Every vendor now has an agent. The demonstrations are genuinely impressive, and in most industries the procurement conversation that follows is about capability and price. Financial services has a different conversation, one the demonstrations rarely address: where does it run? In a regulated institution, agentic AI tends to fail on review, not on capability.

That question got louder after 2026 Dreamforce, where Salesforce framed its direction around the Agentic Enterprise and much of the news centered on keeping permissions and governance defined inside Salesforce. It also lines up with what institutions report about themselves. Deloitte's 2026 study of AI governance in banking found that only about 13% of banks reach leading governance maturity, that monitoring and oversight are weaker for agentic AI than for earlier forms of AI, and that stronger governance is associated with wider deployment. Gartner, meanwhile, predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls.

Capability Isn't the Constraint

In most sectors, routing data through a third-party model is a vendor management question. You assess the provider, check the terms, and sign.

In a regulated financial institution it's a control question, and the control framework isn't new. Firms have spent years establishing where customer data lives, who can access it, how that access is evidenced, and what happens in an examination. That framework was expensive to build, and it's unlikely to be relaxed because an AI feature is compelling.

So an AI layer that moves project, client, and communication data outside the governed environment doesn't fail on capability, it fails on review. That's rarely because anyone objects to the technology, it's because the person signing off has to be able to explain the data flow to a regulator.

The result is a sector that either holds off on deploying, or deploys in low-value areas where the risk is acceptable precisely because the stakes are low.

Why Delivery Data Is the Hard Case

Project delivery data is a particularly awkward category. It contains client communications, which may include material non-public information. It contains commercial terms, staffing decisions, and internal assessments of risk. And in a client implementation, it often touches the customer's own regulated data.

That's a harder category than marketing content or general productivity, and it's also where much of the AI value sits, because delivery risk is exactly what institutions most want early warning about.

Native Changes the Question

Agentic AI becomes deployable when it runs inside the environment that's already governed. That's the design choice behind Cloud Coach's AI, Delivery Intelligence, which runs natively on Agentforce, on live Salesforce records in the institution's own org, under the permissions and sharing model its administrators already maintain. Project data isn't exported to a separate AI vendor, and it stays inside the Salesforce boundary the institution has already assessed.

That can be the difference between a proof of concept and a deployment, because the security answer is the one already on file for Salesforce itself.

There's an operational consequence too. When the AI reads live records, there's no gap between what the model sees and what's actually in the system. Tools that route data through outside models can introduce that gap, and in a regulated context an answer that has to be verified before anyone can act on it isn't much of an answer.

What It Looks Like in Practice With Cloud Coach

A program director asks what's at risk across the transformation portfolio and gets a structured answer from live project records through Project Copilot, rather than a compiled status pack.

A client delivery lead sees stakeholder health across a multi-month implementation through Customer Signal Intelligence, drawn from communication patterns, before the relationship shows up as a complaint.

Meeting decisions land on the project record through AI Meeting Manager, which in a regulated delivery context is more than a convenience, it's the audit trail forming as a by-product of the work rather than being reconstructed afterward.

Why Speed Matters Here

In financial services the assessment cycle is long. A system that takes six months to implement after a six-month assessment turns into a year or more from interest to value, which is longer than many transformation programs can wait. Cloud Coach is typically live in days rather than months, and customers report about a 30% reduction in project overruns and roughly half as many manual tasks. For institutions, the manual-task reduction is often felt first, since so much delivery effort in regulated environments goes into producing evidence rather than outcomes.

The Question to Take Into Any Evaluation

Ask every AI vendor where the inference actually happens, what leaves your tenancy, and what the retention terms are for anything that does. The answers vary more than the marketing does, and in financial services that variance is the decision.

How We See It

Much of the AI on the market leaves Salesforce, and if yours does, the conversation you're having is with your risk function rather than your delivery team. In a regulated institution, agentic AI tends to fail on review, not capability, and where it runs isn't a technical detail, it's the decision.

Frequently Asked Questions Related to Fintech

• How is Agentforce used in financial services delivery?

Banks, insurers, and fintechs can use tools built on Agentforce to answer questions about project risk, track stakeholder health, and capture meeting decisions from live Salesforce records, under the same permissions model that governs the rest of the org.

• What does the agentic enterprise mean for banking?

In Salesforce's framing, it's an operating model where people, data, and AI agents work together. For banks, the practical test is governance, since agents become deployable when they run inside an environment that has already been assessed and can be explained to a regulator.

• Does Financial Services Cloud AI keep data inside Salesforce?

Tools built natively on Salesforce and Agentforce operate on records within the org and inherit its sharing and permissions model. Institutions should still ask each vendor where inference happens and what, if anything, leaves the tenancy.

Sources Cited

Deloitte, "Banking on Trust: AI Governance for Growth, Resilience and Scale," 2026 (Deloitte Trustworthy AI survey of 135 respondents across G-SIBs, D-SIBs, and other large banks). https://www.deloitte.com/dk/en/Industries/banking-capital-markets/perspectives/ai-governance-in-banking.html

Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," press release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027