Solutions / Use case

AI agent deployment, from pilot to production.

An agent that answers in a demo still has to read your real documents, act in your real systems and stop when it should. Tray.ai connects agents to your apps, limits what they can do and records every step, so they can run where the work happens.

Great — a couple more details

This helps us connect you with the right person. Optional, and you can skip it.
Jump to the stages
One request handled by an AI agent in production: asked for in Slack, checked against the credit policy in Confluence and the account in Salesforce, approved by finance in Slack, issued as a credit memo in NetSuite, and every step recorded in Snowflake.

The short answer

What is AI agent deployment?

AI agent deployment is the work of taking an AI agent from a pilot to a job it does every day in your real systems: grounding it in current company knowledge, giving it only the tools its job needs, testing it before each release, handing off to a person when it should stop, and watching every run. Most agents stall here rather than at the model, because each of those steps depends on connecting the agent to the apps the work lives in.

  1. 1 Ground it
  2. 2 Scope tools
  3. 3 Test it
  4. 4 Hand off
  5. 5 Observe
of IT tickets deflected in the agent's first two weeks at Apollo
40%
Slack channels managed by automation at Life360
2,500+
hours a week reclaimed by Zuora's AI customer brief workflow
170+
AI-powered processes in under 12 months at Wavo
30+

AI agent deployment, stage by stage

Five stages, each with what breaks and a how-to guide for building it.

  1. 1

    Ground it in company knowledge

    The agent answers from current documents, and only the ones the person asking may read.

    Where it breaks: The agent quotes a policy that was retired last year.

  2. 2

    Scope its tool access

    Each agent gets the few tools its job needs, running with the access of the person asking.

    Where it breaks: The agent was connected with an admin account in the demo, and can still change any record.

  3. 3

    Test it before every release

    Real requests run against every change, scored on what the agent did as well as what it said.

    Where it breaks: One line changed in the instructions, and the agent stops escalating what it used to.

  4. 4

    Hand off to a person

    When the agent should stop, the whole case goes to the right team and the user hears who has it.

    Where it breaks: The agent tells the user to raise a ticket, and they explain the problem again from the start.

  5. 5

    Watch every run

    Each request is traced from question to action, so a wrong answer can be found and explained.

    Where it breaks: A customer reports a wrong refund, and nobody can say which step made it.

Three ways to put an agent into production

  • The agent inside each app. Quick to switch on, but each one sees a single system, and people end up with a different assistant in every tool. Zuora’s CIO compared it to a hotel with several concierge desks.
  • An agent framework in code. Full control, and your team builds and maintains every connection, access check, approval step and log for as long as the agent runs.
  • An agent platform on your integration layer. The agent uses the same connectors, access control and logging as the rest of your automation, so a new agent starts from what is already governed. This is Tray.ai: teams build agents in Merlin Agent Builder, and IT sets what they can touch. For governing every AI tool and agent, including approvals and cost, see AI and MCP governance.

The Tray.ai platform behind it

Tray.ai is the enterprise orchestration platform for data and AI. These are the parts of it that run AI agent deployment.

Merlin Agent Builder

Build AI agents that act, not just chat.

Agent Gateway for MCP

The MCP Gateway that pays for itself.

iPaaS Governance

AI governance built in, not bolted together.

The metrics AI agent deployment moves

What teams running agents in production measure, and which stage each number depends on.

Resolution rate
Share of requests the agent finishes without a person.
What moves it Current knowledge to answer from, and the right tools to act with. How to build a knowledge base to vector sync 
Handoff rate
Share of requests the agent passes to a person, by reason.
What moves it Clear stop rules, and turning common handoffs into new tools. How to build an AI agent handoff to a person 
Release pass rate
Share of test cases an agent change passes, per kind of request.
What moves it A test set built from real requests that runs on every change. How to test an AI agent before release 
Time to explain a bad answer
How long it takes to find which step produced a wrong answer or action.
What moves it A trace of every run, from the question to each tool call. How to build an agent observability pipeline 

Connectors for AI agents

The systems agents run on, grouped by the seat each one fills. Each links to its connector page.

Models

The language models an agent reasons with. The model is your choice and can change.

Knowledge sources

Where the documents an agent answers from are written and kept.

Systems agents act in

Where the agent looks things up and makes changes, through tools scoped to its job.

Identity and access

Where the agent checks who is asking and what they are allowed to do.

Where people work

Where people ask the agent for things, approve its actions and hear what happened.

Data, monitoring and cost

Where runs are traced, searched for, and costed.

Connections between them

Field mapping, templates and common problems for the pairings AI agent deployment depends on.

Anything not listed connects through the full connector library, or directly through its API.

Teams running AI agent deployment on Tray.ai

“The moment it clicked was when I saw an agent run end-to-end and open a pull request in the repo. It showed me we could actually close the loop on issues, and that's real value.”

— Matt Currie, Senior Automation Engineer, Life360
Read the study

Processes next to AI agent deployment

These share systems and controls with agent deployment without being stages of it.

AI operations

How to build human approval for agent actions

Decide what needs approving by blast radius, show the approver what will happen, expire cleanly, and keep the record. The Headless prompts that build it.

AI operations

How to build LLM cost and usage tracking

Attribute spend to a team and a feature, alert on the rate rather than the total, and catch the loop before the invoice does. The prompts that build it.

AI operations

How to expose an internal system as an MCP tool

Scope the tool to a job rather than an API, resolve identity per call, make writes idempotent, and log every invocation. The Headless prompts that build it.

AI operations

How to build shadow AI discovery

Find the AI tools nobody registered, from expenses, SSO and DNS, then route to a path rather than a ban. The Headless prompts that build it.

AI operations

How to build AI support deflection

Answer only what the knowledge base actually covers, escalate early with the context, and measure resolution rather than deflection. The prompts that build it.

AI operations

How to build a document intelligence pipeline

Score confidence per field rather than per document, validate against a system of record, and route only the uncertain fields to a person. The prompts.

Reading and resources

Guide

AI agent strategy playbook: How to run agents at enterprise scale

Learn how teams move AI agents from pilots to production with the right operating models, governance, and orchestration strategies.

Guide

From prototype to production: AI agents that actually ship

How platform teams ship AI agents to production with Tray and Amazon Bedrock—featuring architectures, frameworks, and real use cases.

Survey

How enterprise teams are scaling AI agents

New data from 1,000+ IT leaders on adoption, investment, and what’s blocking scale.

eBook

Not everything needs AI. A framework for knowing what does.

Not every workflow needs an AI agent. A framework for knowing when to use rule-based, AI-powered, or agentic automation — before you build.

On-demand webinar

Architecting AI agent deployment that works

What separates successful AI agent deployments? See how IT teams address integration, architecture, and scale — with a demo in Slack and Okta.

On-demand webinar

How to deploy AI agents faster with Merlin Agent Builder

Learn how IT leaders are scaling AI agent deployment faster — with real-world strategies, practical frameworks, and a live demo of Tray’s Merlin Agent Builder.

On-demand webinar

What to do with your second agent

Learn from Tray.ai CEO Rich Waldron and Amalgam Insights CEO Hyoun Park on how to master scaling AI agents, from orchestration and ROI to scalable deployment.

On-demand webinar

Build and launch agents faster with Tray Agent Hub

See how Tray Agent Hub helps teams cut build time, reuse tools and data, and scale adoption inside Tray Merlin Agent Builder. Watch the demo on demand.

From the blog

BlogAI agents in production: where pilots stallBlogReduce costs, boost efficiency with AI knowledge agentsBlogFirst of its kind native vector table for iPaaSBlogDrops of determinism: the fix for MCP agents that keep getting it wrongBlogAI agent evaluation: why agents pass the demoBlogHuman in the loop: when an AI agent should stopBlogHow to think about measuring agent performanceBlogAI agent sprawl: why the fifth agent breaksBlogEnterprise AI platforms that handle integration and agent orchestration, comparedBlogHow to build AI agents that actually work (not just chat)BlogAI architecture is why most AI agent projects failBlogThe four blockers to AI agent adoption (and how to fix them)

Speaking the same language

Definitions for the terms that come up in this conversation.

AI agent

An AI system that can reason about a task, use tools to act on it, and learn from feedback — not just generate text.

Human-in-the-loop

A design pattern where human approval or review is a step inside an otherwise automated process — keeping people accountable for the decisions that matter.

AI Agent Orchestration

How multiple AI agents are coordinated to work as a system: which agent acts, when, on what data, and with what oversight.

Agentic AI

AI systems that don't just generate responses but take action — planning, using tools, calling APIs, and completing multi-step tasks toward a goal.

Frequently asked questions

Why do so many agent pilots never reach production?

+

Because the pilot proves the model can answer, and production asks different questions. Can it read current documents and only the ones the user may see? What can it change, and as whom? Who picks up what it can't finish? Each answer is a connection to another system, and that is the work a pilot skips.

How is this different from building on an agent framework in code?

+

A code framework gives you the reasoning loop. You still build the connections to every system, the access checks, the approval step, the logging and the handoff yourself, and you maintain them. On Tray.ai those come with the platform, along with 700+ connectors an agent can use as tools.

Do agents need approval for everything they do?

+

No. Approval for everything trains people to approve without reading. Agents act on their own for actions that are small and easy to undo, and ask first for anything that reaches a customer, moves money, changes many records or can't be undone. Setting up those approvals is part of AI and MCP governance.

Which models can our agents use?

+

The model is your choice. Tray.ai connects to OpenAI, Anthropic (Claude), Gemini, AWS Bedrock, Cohere and Mistral AI, and you can change models without rebuilding the agent.

Where do people use the agents?

+

Where they already work. Agents built in Merlin Agent Builder run in Slack and Microsoft Teams, in a chat window, or behind an API, from one build.

How long does it take to get agents into production?

+

It depends on the job, and teams that reuse parts move fastest. Life360 builds its sub-agents in weeks rather than months by sharing tools and data sources between them. Wavo put more than 30 AI-powered processes in place in under 12 months.

How does this relate to AI and MCP governance?

+

AI and MCP governance covers every AI tool and agent the company uses, including the ones it didn't build, from finding them to approving high-risk actions, tracking cost and auditing every MCP tool call. Agent deployment covers building and running the agents you own. Many teams need both, and the same connectors and access controls serve each.

Last reviewed 2 October 2026.

See AI agent deployment running on your stack.

Walk through your systems and stages with a Tray.ai expert.

Great — a couple more details

This helps us connect you with the right person. Optional, and you can skip it.
Browse the how-to guides