# 5 lessons from getting AI agent workflows into production

> Forward-deployed engineers share 5 lessons for getting AI agent workflows from demo to daily use and running reliably in production.

Source: https://asana.com/inside-asana/ai-agent-workflows-demo-to-production

## From demo to daily use: 5 lessons for getting AI agent workflows into production

#### In brief: taking an AI agent from pilot to production in any business

_Getting an AI agent from demo to daily use has less to do with the model than with edge cases, missing data and trust. Alberto Arrighi leads customer success at StackAI, now part of Asana, and Max Poff is one of its forward-deployed engineers, who split their time between writing code and deploying agents inside banks, private equity firms, agencies and defense contractors. In a_[recent interview on MTS](https://www.youtube.com/watch?v=HJdg7bKlYAY)_, they shared five practices that take an agent from prototype to production: start with a tedious, high-stakes workflow, treat user testing as the real build, connect the data the whole team relies on, build in a learning loop, and design for security from day one._

## What's the hardest part of getting AI agents into production?

The hardest part is the stretch between a working demo and the moment users trust the agent in their daily work. "The hardest part is handling the edge cases that emerge between the demo and the point where users truly trust it in daily production," Max said.

Demos run on clean inputs and one person's assumptions. Production brings messy documents, missing permissions, and colleagues who work differently from the person who sponsored the project.

## 1. Start with a tedious, high-stakes workflow

The best first agents take on work that is essential but slow. Max described a risk and compliance team at a large European bank. Every time a business change proposal arrived, risk managers cross-referenced a large risk register to spot new risks, then a controls register to assign mitigations.

The team now posts the proposal to a custom Slack bot. A StackAI workflow returns a formatted Excel workbook with the full risk assessment, flagged risks and mapped controls. "A task that used to take an entire afternoon now takes roughly 10 to 15 minutes," Max said.

Two things made it stick: the agent met people inside a tool they already used, and the output arrived in the format they already worked in.

## 2. Treat user acceptance testing as the real build

User acceptance testing (UAT) is where a demo becomes a product. Max worked with the executive assistant team at a private equity firm to automate expense coding in SAP Concur. "You can build a demo for that very quickly," he said. "But once we put it in front of beta testers, the real work began."

StackAI runs a structured UAT phase on every deployment to surface the cases where the agent fails. Each fix makes the workflow more resilient to real-world use. Plan for this phase in your timeline; it is not a sign the project is behind.

## 3. Connect the data the whole team relies on

An agent built with one stakeholder sees only that stakeholder's world. Alberto's team built a media planner agent for one of the largest media and marketing agencies in Northern Europe. Given a client brief, it recommends channels, budget allocations and a rollout plan.

It looked solid with the primary stakeholder. When the wider team tested it, they pointed to critical datasets the agent couldn't reach. "Building a demo is easy; opening it up to real testers is when you discover missing system permissions or broken logic," Alberto said. The fix was to connect more tools and break down the silos between them.

This reflects a wider shift. A year ago, most use cases connected an agent to a single knowledge base. Today they are orchestration systems that pull from many sources, trigger on events, and take action across tools.

## 4. Build a learning loop into every workflow

Agents get more useful when they remember what they learn. StackAI agents keep shared memory files across users in a client organization. As people use the agent and give feedback, that memory updates, and the agent loads it into future runs.

Alberto pointed to a second loop: writing outputs back into the original knowledge base. "That continuous write-back loop is a huge value driver for our enterprise deployments," he said. Over time, every run adds to what the organization knows.

## 5. Design for security and governance from day one

In regulated industries, where an agent runs matters as much as what it does. Top financial institutions increasingly require agents to run in their private cloud or on-premise, not on shared SaaS infrastructure. StackAI also supports defense contractors running in air-gapped environments.

"For defense, banking, and hedge funds, data perimeter security is non-negotiable," Max said. Governance, observability and security become essential once agents touch sensitive data and take higher-stakes actions.

## What can AI agents do now that they couldn't a few years ago?

Models now handle structured data well. "When I first started, the common advice was: if your workflow depends on a messy, convoluted Excel spreadsheet, AI probably isn't the right solution," Max said. Today, many of the agents he builds produce complex, formatted Excel workbooks.

Agents can also run in the background, maintaining a data layer while people simply ask questions. Max believes this puts static BI dashboards under pressure: "Dashboards represent static snapshots of what someone guessed you might want to see."

## Where the Asana Work Graph® fits in

Every practice above comes back to one thing: context, institutional knowledge…understanding the who is doing what by when, how and why. An agent makes it from demo to daily use when it can see the data, decisions and owners behind the work. That’s exactly what lives in the Asana Work Graph®. StackAI adds cross-system execution, so agents can act across the tools a business runs on. Together, a workflow doesn't have to start in a chat window: assigning a task in Asana can trigger a StackAI agent to run end to end, with people approving and handing off work along the way. "Connecting agentic capabilities directly into enterprise work graphs unlocks an entirely new operational tier," Alberto said.

#### Get Started

Ready to move your own AI agent workflows from demo to daily use? Learn how StackAI by Asana helps enterprises build, test and deploy agents across the systems their business runs on.
- [Learn more](https://www.stackai.com/)

## About the experts

[Alberto Arrighi](https://www.linkedin.com/in/alberto-arrighi/) leads customer success at StackAI, now part of Asana. His team of AI strategists and forward-deployed engineers works with enterprise customers to find the highest-value workflows for agents and turn them into specs and requirements for production deployments. He began at StackAI as an AI strategist. 

[Max Poff](https://www.linkedin.com/in/maxpoff/)is a forward-deployed engineer at StackAI, now part of Asana. He splits his time between writing code and deploying agents with enterprise clients, including risk and compliance teams in banking and operations teams in private equity, and feeds what he learns back into the product.

- [Why Anthropic's "Worse Together" study is the most important read on agentic workflows right now](/resources/anthropic-worse-together-agentic-workflows)

Artificial Intelligence (AI)

Team Asana

#### Head of Global Communications

- [Introducing the Asana AI Teammates integration with Databricks Genie One](/inside-asana/asana-ai-teammates-databricks-genie-one)

Artificial Intelligence (AI)

Engineering

#### Content writer

Companies have invested heavily in bringing their most important data together. But for many employees, getting an answer still requires specialized tools, technical expertise, an ...

- [Introducing AI Meetings and Tables: Every meeting moves work forward, every Table keeps it current](/inside-asana/introducing-ai-meetings-and-tables)

Artificial Intelligence (AI)

#### Content writer

Campaign kickoffs and program reviews are where teams make their most important decisions. But too often, commitments aren’t tied to an accountable person, and the decisions thems ...

- [Asana as a governed data source: bringing Work Graph context into a secure multi-agent system](/inside-asana/asana-governed-data-source-work-graph-multi-agent)

Engineering

Artificial Intelligence (AI)

#### Developer Advocate

The data that makes AI evaluations great lives in data warehouses, production databases, and private systems. Only by accessing these governed systems can you get valuable busines ...

- [From demo to daily use: 5 lessons for getting AI agent workflows into production](/inside-asana/ai-agent-workflows-demo-to-production)

Artificial Intelligence (AI)

Team Asana

#### Head of Global Communications

What's the hardest part of getting AI agents into production?The hardest part is the stretch between a working demo and the moment users trust the agent in their daily work. "The ...

- [Artificial Intelligence (AI)](/inside-asana/artificial-intelligence)
