Est.

Enabling Non-Technical Employees With AI Agents at Scale

Most enterprise AI agent pilots fail due to infrastructure gaps, not lack of employee interest.

· 8 min read · Updated
Cover illustration for “Enabling Non-Technical Employees With AI Agents at Scale”
AI Change Management · September 26, 2026 · 8 min read · 1,910 words

Zapier has thousands of AI agents running inside the company, used company-wide, not as a pilot but as how the place runs day to day. Deloitte is putting Claude in front of hundreds of thousands of employees. Those two facts sit far outside the norm, and the distance between them and most companies' experience is the actual subject of this piece. Most organizations trying to roll out AI agents never get close. The Agentics' Enterprise MCP Guide 2026 puts the number of enterprise agentic AI pilots that reach production at 11 to 14%. The rest stall, and they stall for a specific, nameable reason: identity gaps, audit gaps, access-control gaps, and vendor lock-in, not because employees don't want to use the tools. The employees stuck waiting are concentrated exactly where the value is highest: sales, marketing, legal, and operations teams who could put an agent to work tomorrow if there were a governed way to give them one. Zapier and Deloitte didn't get lucky. They built the infrastructure that makes broad access possible, which is a different project than training people to want it.

That's the frame for everything that follows. This is an infrastructure problem.

What non-technical employees using agents in their daily work need

Diagram: Why Most Agentic AI Pilots Never Leave the Runway. Visualizes: Show the dramatic drop-off from pilot to production for enterprise agentic AI: 86–89% of pilots stall before reaching production, leaving only 11–14% that make it, according to…

Picture what it actually takes for someone in legal or marketing to use an agent well. They open a chat window, type what they need in plain language, and the right tools respond, with no separate login, no config file, no API key to track down. Accenture's 2026 analysis argues that conversational interfaces, not rigid commands or API setup, will define how enterprise AI actually gets used, and that's the experience employees need. It is not what most of them get.

What most of them get instead is friction that has nothing to do with whether they want to use the tool. Setting up an agent usually means a separate OAuth consent screen for every tool it touches, credential files sitting around in plaintext, and manual provisioning that needs a platform engineer involved for every new connection. A marketing manager does not have a platform engineer on call. Neither does a legal analyst. The tool sits unused because nobody built a path that doesn't run through an engineering ticket.

Where that path exists, the results speak clearly. Inside Anthropic's own legal team, a lawyer with no coding background built self-service tools that now triage issues before they ever reach the legal queue, using Claude-powered workflows. Marketing review turnaround dropped from two or three days down to within a single day. Nobody made that lawyer into an engineer. The tooling just got out of the way enough for someone who understood the legal work to build what the legal work needed.

The pattern holds elsewhere. Novo Nordisk's NovoScribe platform cut clinical study documentation time from ten-plus weeks down to minutes, with far fewer resources needed to get there. L'Oréal picked Claude for conversational analytics specifically because employees across multiple systems could ask questions in plain language instead of writing queries, and it handled complex, cross-system questions with high accuracy. Neither result came from training people to code. Both came from tooling that absorbed the technical complexity so the person with the business knowledge could just ask.

That's what a governed path to agent access has to deliver: the same ease for a legal analyst or a marketing manager that a developer already has. Runlayer, an enterprise MCP gateway and AI control plane that connects agents to 18,000+ integrations while enforcing policy and audit across every session, is built to give non-technical employees that access from tools like Slack without routing every request through a platform engineering queue. In the infrastructure world, this gets called a golden path, and building one at scale requires something to sit underneath it that handles the connections between agents and the dozens of tools they need to touch.

That something is the Model Context Protocol. MCP, built originally by Anthropic and now governed by the Linux Foundation's Agentic AI Foundation, solves a problem that used to require a custom connector for every single combination of agent and tool. Instead, one protocol handles all of them. Every major AI provider has adopted it: OpenAI, Google DeepMind, Microsoft, and AWS all run on MCP now, which makes it shared infrastructure across the industry rather than something Anthropic built for itself. Downloads of the MCP SDK have reached a scale that points to real enterprise use, and CRM platforms, marketing automation tools, and developer tooling have all added MCP support.

That ubiquity is also the catch. MCP is the connective tissue linking agents to the tools employees need, and that means its security is the enterprise's security for AI, full stop. A single MCP server often is in front of credentials for several enterprise systems at once. Compromise that one server and every system behind it is exposed. Plenty of MCP servers get deployed with no authentication at all, overprivileged credentials sitting in plaintext config files, and default network settings that leave them reachable from the open internet. Trend Micro found 492 MCP servers exposed to the internet this way, with zero authentication protecting them, and separate reporting citing an Astrix Security audit found that only a small minority of public MCP servers use OAuth. Every new employee given agent access through an ungoverned MCP setup adds to that exposed surface. Without governance built in, giving more people access means giving more attackers a way in.

The specific threats that show up when agents operate without governed tool access

These risks are not hypothetical. Several have already caused real, documented damage, and they share one mechanism: an agent treats an instruction planted by an attacker inside the content it's processing the same as one its user meant to give it.

Tool poisoning is one version of this. An attacker hides malicious instructions inside an MCP tool description, text the user never sees but the agent reads directly into its working context. Invariant Labs showed this could go all the way to complete SSH key theft, carried out through a poisoned tool description inside a real Cursor IDE setup.

Prompt injection through outside content is another. In the Supabase MCP incident, an agent processing support tickets followed instructions buried inside one of those tickets and handed back live integration credentials, because the agent had the access needed to do it and no way to judge whether the instruction was legitimate. The GitHub MCP incident worked the same way: a public GitHub Issue carried instructions that caused an agent to leak private repository code. Different systems, same architecture, same failure.

Some of these attacks need no click from anyone. EchoLeak, tracked as CVE-2025-32711, hit Microsoft 365 Copilot through a specially crafted email that silently pulled out email contents, documents, and Teams messages, without the user doing anything at all.

The clearest picture of what this looks like at scale came out of Mexico between December 2025 and January 2026. A single attacker used Claude and ChatGPT/GPT-4.1 to breach the federal tax authority, the electoral institute, four state governments, and a water utility, pulling out a massive volume of data that included hundreds of millions of taxpayer records. The attacker's own conversation logs sat publicly accessible online before anyone defending those systems acted on it.

Supply chain attacks round this out. Check Point Research found a way to achieve remote code execution in Claude Code through poisoned repository config files, and a large number of malicious skills turned up across ClawHub.

What connects every one of these cases is the same structural fact: once an agent can both understand content and act on it, anything it reads becomes a potential instruction. That's a different threat model than the one traditional security tools were built to handle, and it's why firewalls and data-loss prevention systems, all built to watch for known bad patterns, don't know what to do with a perfectly normal-looking instruction that happens to come from the wrong place.

The identity gap underneath the tool-access problem

The deeper reason these attacks keep working comes down to identity. Identity and access management systems were built for people logging into systems with a fixed set of permissions, not for agents, and the distance between what IAM tools can see and what agents are actually doing is where most AI security failures start.

The Cloud Security Alliance's 2026 research on AI agent governance found that 92% of security leaders at large enterprises lack full visibility into their own AI identities. Most don't enforce access policies for those identities at all, and most of their AI systems can already reach core business platforms, ERP, CRM, financial systems, with very little actual governance over what they do once they're in.

Part of the reason this is so hard to fix is that an agent's scope changes task by task. The same agent might act on behalf of one employee with one set of permissions in one moment, then act for a different employee with different permissions the next. Authentication needs to carry that delegation context with every single invocation, and conventional IAM tools were never built to track permissions that shift from one request to the next.

Left without a fast, governed path to agent access, employees build their own. Shadow MCP servers go up outside any formal process, and that activity happens entirely outside what security teams can see or audit. Only a fraction of organizations have a formal identity strategy for AI agents, and only a fraction of agents make it to production with full security approval. Gartner projects that through 2029, more than half of successful attacks on AI agents will come through access-control weaknesses specifically. That projection makes access control the single highest-priority gap to close.

What the 2026 MCP specification and enterprise authorization provide

The good news is that the protocol itself caught up in 2026. The spec now supports enterprise-grade authorization in ways it simply didn't before. The gap that remains is between what the spec supports and what most companies have actually built on top of it.

The July 28, 2026 update to the MCP spec tied authorization directly to OAuth 2.1 and OpenID Connect. MCP servers now operate as OAuth 2.1 resource servers, and clients have to implement Resource Indicators under RFC 8707, which bind a given token to one specific server. That one change closes off a real attack: a malicious server tricking a client into handing over a token meant for a different, legitimate server.

The enterprise-managed authorization extension, finalized in June 2026, goes further. Under EMA, the company's existing identity provider becomes the authority for who gets to use which MCP server. An employee logs in once, through the identity system the company already runs, and the identity provider automatically provisions every MCP server that employee is cleared to use. There's no separate OAuth consent screen to click through for each new tool. Every access decision lives in one place, the identity provider's own admin console, with one audit trail covering all of it.

That's the technical foundation Zapier and Deloitte's results were built on: identity-based provisioning, tokens bound to the servers they're meant for, and one governed record of who has access to what. The spec supports it now. What separates the companies still stuck in the 11 to 14% from the ones running agents across their whole workforce is whether they've actually built on top of it.

Sources

  1. MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure
  2. The 2026 State of AI Agents Report How enterprises are building and

More in AI Change Management