---
title: "Build, Connect, Control: The Three Layers of an Enterprise AI App"
slug: "enterprise-ai-three-layers-build-connect-control"
author: "ibl.ai Engineering"
date: "2026-11-02 09:00:00"
category: "Premium"
topics: "enterprise ai architecture, agentic vibe, mcp, llm spend caps, ai governance, coding agents, self-hosted ai, internal tools, source code ownership"
summary: "A customer service app from one prompt in about ten minutes covers only the first of three layers. Enterprise AI apps also need connections to live systems and enforceable limits on spend, access and audit, and most evaluations test only one."
banner: ""
thumbnail: ""
linkedin: |
  One prompt to Claude Code. About ten minutes later, a customer service app is running with single sign-on, agent chat and conversation history, and no key pasted into the chat.

  That demo is real, and we recorded it. It is also one layer of three.

  Build: can one person get to a working, signed-in app quickly?

  Connect: can the agent reach live systems, with the right person's permissions, without a custom integration for each one?

  Control: when usage takes off, can you cap spend per person, narrow access and answer what the agent did last Tuesday?

  Deloitte's 2026 survey of 3,235 leaders found close to three-quarters of companies plan to deploy agentic AI within two years, and only 21% of those report a mature governance model. Fast building makes that gap wider, because a useful app gets adopted before anyone sets limits.

  So when you evaluate a platform, test the three together, in one afternoon, on one real app. Then ask a fourth question: who holds the code when the contract ends?

  #iblai #EnterpriseAI #AIGovernance #MCP #AIAgents
---

## The Short Answer

**An enterprise AI app needs three layers: a fast way to build it, connections to the systems where the data lives, and enforceable limits on spend, access and audit. A one-prompt demo proves only the first. Test all three on one real app. With ibl.ai you own all the code and the data across every layer, with no per-seat pricing.**

The three layers are not phases. If you add them one at a time, the second and third arrive after people already depend on the app.

## What does the build layer of an enterprise AI app look like?

The build layer takes a description and returns a working app with sign-in and an agent attached. In [our October 30 recording](/updates/agentic-vibe-customer-service-app-one-prompt), Jaione Amigot builds SupportDesk from one prompt to Claude Code in about ten minutes, with no code written by hand.

The [iblai/vibe](https://github.com/iblai/vibe) skills scaffold a Next.js app, connect it to an ibl.ai organization and attach a support agent that already existed in the tenant. A browser sign-in writes the credentials into environment files, so no token appears in the chat.

Be precise about what that shows. It is a recorded demo against the hosted backend, not a production rollout.

It proves a person can reach a signed-in, agent-connected app in minutes. What [Agentic Vibe](/blog/what-is-agentic-vibe-enterprise-ai-apps-coding-agent) adds to a coding agent is the part it can't guess: your sign-in, roles, memory and billing.

Speed matters because skilled people are scarce. [BCG](/blog/bcg-applied-ai-talent-technology-services-buyers) reports applied AI is 10 to 15% of the AI-tagged workforce at technology services firms, against 90 to 95% at AI-native firms.

## What does the connect layer of an enterprise AI app need?

The connect layer gives an agent governed access to live systems, so it doesn't work from a stale copy. The [Model Context Protocol](https://modelcontextprotocol.io/) is the common standard for this. An app that reaches none of your systems of record can answer questions, and little else.

Four details separate a demo connection from an enterprise one. Our [October 30 platform update](/updates/platform-update-2026-10-30) shipped each of them, in a two-week release that moved twelve services and packages:

<table>
<thead><tr><th>Requirement</th><th>Why it matters</th><th>What shipped</th></tr></thead>
<tbody>
<tr><td>Tenant-level credentials</td><td>One customer's OAuth client must not collide with another's</td><td>Per-tenant OAuth clients and admin-held secrets</td></tr>
<tr><td>Identity on every call</td><td>The agent should act with the user's access, not a shared super-user</td><td>User-scoped connections, enforced where that scope is set</td></tr>
<tr><td>User choice</td><td>Not every chat should reach every system</td><td>Per-user and per-session server toggles</td></tr>
<tr><td>A curated catalog</td><td>Anyone can publish a server</td><td>Vendor servers synced from the MCP Registry, curated by staff</td></tr>
</tbody>
</table>

Keep one caveat in view. Reaching data in place means it isn't copied into a separate store, but whatever a tool returns still goes to the model. That is why where the model runs belongs in the same conversation.

## What does the control layer of an enterprise AI app enforce?

The control layer turns usage into something you can bound and explain.

[Deloitte's 2026 survey](https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html) of 3,235 business and IT leaders in 24 countries found close to three-quarters of companies plan to deploy agentic AI within two years, and only 21% of those report a mature governance model.

The most dangerous deployment is the useful one, because people adopt it before anyone sets limits. So the controls need to exist on day one, and they need to be specific:

- **Spend limits.** [Three scopes](/updates/spend-limits) shipped in August, workspace, agent, and one person on one agent, each with a block or alert mode. A [per-user cap across every agent](/updates/platform-update-2026-10-30) followed in October, with an optional tenant default for new members.
- **Server-side enforcement.** Block mode refuses new requests for every client, so a custom app can't route around it. See [why budgets break](/blog/enterprise-ai-budget-overruns-spend-caps-2026) without this.
- **A record.** [Transcripts](/docs/os/analytics/transcripts) show what an agent said and did, and [Audit](/docs/os/analytics/audit) logs who changed its configuration.
- **Role boundaries.** [Roles](/docs/os/organization-settings/roles) define who can see and change what.

## How do you test all three layers in one evaluation?

Run one real app through all three layers in a single sitting. Vendors tend to demo the layer they are best at, and a test across all three exposes the seams. Ask for these five things, live:

1. Build a signed-in app against your identity provider, from a prompt, with no credential pasted anywhere.
2. Connect one real system of record, and show that two users with different access get different results.
3. Set a per-user cap, then hit it, and show the refusal.
4. Export the record of what the agent did during that test.
5. Turn off one connection for one chat, and show that the agent loses it.

Any step that needs a follow-up meeting has told you which layer is missing.

## Who should own the three layers of an enterprise AI app?

Ownership sits underneath all three layers, because a vendor can change the price or terms of any of them. If speed, connections and controls all live in a hosted product, the app you built in ten minutes depends on that contract.

With ibl.ai you own all the code and the data under a perpetual licence. The platform is model-agnostic across any LLM, has no per-seat pricing, and you can deploy anywhere: your own cloud, on-premise, GovCloud or a fully air-gapped network.

The Vibe toolkit is MIT-licensed, and the same app that runs against the hosted backend in the demo runs against ibl.ai on your own servers. Start from the [Agentic Vibe page](/product/agentic-vibe), or [book a pilot](/contact).

ibl.ai is family-owned and operated from New York, NY.

*Sources: ServiceNow's [third-quarter 2026 filing](https://www.sec.gov/Archives/edgar/data/0001373715/000137371526000092/erq3fy26.htm) shows the infrastructure side of this market: total revenues of $4,106 million, up 20.5%, and subscription revenues of $3,988 million, up 21%. Product facts are from our own update notes, linked above.*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.