The Hard Part of AI Isn’t Building. It’s Running

For years, enterprise AI conversations revolved around building better models. Then generative AI arrived, dramatically reducing the effort required to create software. Today, engineers can describe an application in plain English and generate a working prototype within hours, while business users can assemble AI agents capable of querying databases, invoking enterprise tools and automating tasks with minimal coding.

But while building AI has become easier, running it reliably in production has emerged as the real challenge.

According to workflow orchestration company Prefect, enterprises are entering a new phase of AI adoption where operational discipline, not development speed, will determine success. The company believes organisations that pull ahead will not necessarily be those deploying the most AI models, but those capable of governing, monitoring and continuously improving AI systems after deployment.

Founded in 2018, Prefect provides workflow orchestration software that helps organisations automate data pipelines, machine learning workflows and AI applications. Its platform enables engineering teams to monitor workflow execution, manage failures and maintain visibility across increasingly complex AI systems. 

More recently, the company has expanded into AI operations through Horizon, a secure gateway for agentic workloads, while its open-source FastMCP framework has become one of the most widely adopted ways of connecting AI agents to enterprise tools and data sources.

The Production Challenge Begins After Deployment

An AI application that performs well in a demonstration often encounters an entirely different reality in production. Once connected to customer data, financial systems or operational workflows, AI agents begin making decisions that directly affect business outcomes. At that point, organisations must answer questions that rarely arise during the prototype stage: Who approved the deployment? What systems can the agent access? What happens if a workflow fails? And who is monitoring spiralling AI infrastructure costs?

Wayne Grigsby, Lead Sales Engineer at Prefect & FastMCP, argues that many organisations discover these issues only after deployment. The company also warns that inefficient AI architecture can significantly inflate operational costs. 

Excessive token usage, unnecessary context retrieval and overuse of expensive foundation models increase inference costs. These design decisions, Prefect says, need to be addressed during development rather than after AI applications reach production.

Governance Cannot Rely on Prompts

Among the challenges emerging in enterprise AI, governance remains one of the most underestimated.

Many organisations are giving AI agents permission to access customer records, execute transactions and update internal systems while relying on prompts to regulate behaviour. Prefect refers to this as ‘vibe governance’, an approach that depends more on prompt engineering than enforceable operational controls.

Grigsby says that prompts cannot replace structured permission models, audit trails or access controls. AI agents operate within the permissions they are granted and execute instructions at machine speed without the judgement humans naturally apply. As AI moves deeper into core business processes, governance must therefore become part of the underlying infrastructure rather than an afterthought.

Reliability is Becoming the Competitive Advantage

Governance alone is not enough. Enterprise AI also requires reliability comparable to any other business-critical software system. Running production AI involves scheduling, retries, dependency management, failure recovery and continuous monitoring. 

Without these capabilities, organisations may only discover that an AI workflow has failed after incorrect outputs have already influenced reports or customer interactions.

Grigsby points to WHOOP, the human performance technology company, as an example. The company uses Prefect to orchestrate the data and machine learning workflows behind users’ sleep and recovery insights. After replacing internally developed tooling, the company reduced production incidents by 75% and improved mean time to recovery by more than 40%, largely because engineering teams gained greater visibility into workflow execution and failures.

For Prefect, this level of observability separates an AI experiment from a production-grade AI platform.

Building an AI Stack Organisations Can Trust

The company also challenges the industry’s understanding of what it means to be ‘AI-native’. Rather than simply deploying multiple AI models, Prefect argues that AI-native organisations are those capable of operating AI with predictable outcomes, governed access, observability, auditability and clear human accountability.

Instead of treating AI as a collection of disconnected tools, enterprises should think in terms of an integrated operational stack spanning applications, agents, orchestration platforms and governed enterprise data.

WHOOP demonstrates this approach. Its enterprise data resides within Snowflake, where governance policies and access controls are established centrally and inherited by downstream applications. Prefect orchestrates the workflows responsible for processing data and machine learning operations, ensuring reliability and observability while Snowflake governs the underlying data.

"The companies pulling ahead with AI are the ones that treat governed data as the starting point, not a cleanup step. When your data is governed in one place, and the workflows running on top of it are reliable and observable, teams can put AI into production without gambling on trust," says Remy Thellier, Senior Manager, AI/ML DCP Partners, Snowflake. 

As enterprise AI matures, Prefect believes the defining capability of market leaders will not be how quickly they can build AI applications, but how effectively they can operate them. Governance, observability and reliability may attract less attention than the latest models, but they will ultimately determine whether AI becomes trusted enterprise infrastructure or remains a promising experiment.