While almost every IT leader will tell you, “Yes, we’re using generative AI.” When asked, “How much of it is in production?” most will say, “Not yet.”

In a recent Enterprise Talk podcast, Martin highlighted a key distinction that separates the few who are shipping from the many who are stuck in proof-of-concept limbo: Are companies experimenting to learn, or experimenting to achieve a specific outcome?

Both are valid. But they lead to very different results. If your goal is “learn about AI”, you’ll get lots of demos and internal excitement – then stall. If your goal is “reduce processing time by 60%”, you can measure whether your agents are worth promoting to production.

In other words, while curiosity is essential, it’s commercial clarity that will get AI out of experiments and into production.

4 tips for moving agentic AI into production

Martin advocates four guidelines for those who want to move beyond experiments and turn agentic AI into a reliable, scalable capability.

1. Start with outcomes, not experiments

One of the easiest ways to get stuck with AI is to treat it as an open-ended exploration. Teams “try things out,” build demos, plug into an API or two, and then nothing ever quite makes it into production.

According to Martin, a better starting point is deceptively simple: be explicit about what you want to change.

Instead of launching an “AI initiative”, identify a handful of business measures that genuinely matter: turnaround time on a core process, cost-to-serve, customer satisfaction, incident rates, or the total cost of ownership of your application portfolio. Frame your AI work around concrete questions such as: Can we cut claims processing time in half? Can we reduce manual effort in onboarding by 60%? Can we automate the majority of level-1 support triage?

When you begin this way, you give your teams something firm to design around and something measurable to judge success by. That’s what moves agentic AI from “interesting” to “investable”.

2. Leverage your investment

There’s a persistent myth that to benefit from AI, you need to start with a clean slate. In reality, your existing systems, APIs, and data can be an advantage.

Most enterprises already have investments in SAP, Salesforce, core banking systems, HR platforms, data warehouses, and integration layers. They’ve spent years codifying processes and business logic in these systems. 

Martin believes this is where modern platforms and orchestration layers, such as OutSystems Agent Workbench, come into play. They allow you to plug agents into existing systems, call existing APIs, and use established data models and security controls. You aren’t discarding what you’ve built; you’re activating it in new ways.

A practical pattern looks like this: pick a cross-system process—lending, onboarding, claims, or procurement— and map where humans are doing repetitive, rules-based work, such as checking documents, reconciling data across systems, or applying standard rules. Instead of rebuilding everything, you let AI agents interact with those same systems, but at machine speed and with perfect consistency. The architecture you’ve already built becomes the infrastructure your agents rely on.

3. Think platform, not point solution

In the early stages of AI adoption, it’s tempting to treat each use case as a self-contained project: a bot for one function, an automation for another, an agent embedded in a single team’s workflow. That’s useful for learning, but it doesn’t scale.

If every team chooses its own tools and patterns, you quickly accumulate a tangle of AI scripts, prototypes, and shadow integrations. You may not know which agents exist, what data they touch, or how they’re making decisions. At some point, that becomes a governance and risk problem.

To avoid this chaos, Martin asserts that organisations need to approach agentic AI as a platform capability. This means one way of working. You’ll want consistent approaches to how agents are defined, how they authenticate, what they’re allowed to do, how they’re monitored, and how their actions are audited. You’ll want a shared place to register agents, see their status, and coordinate their interactions with each other and with your systems.

It’s reasonable to assume that most major SaaS and enterprise systems you use today will eventually become “agentic” in their own right. CRM platforms will come with sales agents, ERPs with finance agents, HR systems with talent agents. On top of that, you’ll need a neutral orchestration layer to coordinate cross-system work. In this environment, agents from different products and platforms can participate in end-to-end workflows without stepping on each other.

Designing that layer early, before you have dozens of unsupervised agents in the wild, is one of the most important architectural decisions an IT leader can make.

4. Aim AI at the work humans shouldn’t be doing

If you’re unsure where to begin with agentic AI, look for the parts of your organisation where the work is mundane and repetitive. In most enterprises, there are processes where well-qualified staff spend their days reading similar documents, comparing values, copying data between systems, checking checklists, or chasing missing inputs. These processes are usually document- and data-centric, rules-based, and high-volume. Martin says they’re also prime candidates for agentic AI.

Consider workflows like mortgage origination, insurance underwriting, legal intake, compliance checks, KYC, or supplier onboarding. In all of these, large volumes of semi-structured information need to be checked, classified, and evaluated against a known set of rules. An agent can determine which documents have been submitted, compare them to what’s required, extract the relevant fields, and apply scoring or decision logic. Only ambiguous or exceptional cases need to be escalated to human experts.

When this model is implemented well, the impact can be dramatic. Organisations move from handling hundreds of cases per week to thousands or tens of thousands with the same headcount. Human effort shifts away from data verification and towards customer conversations, complex judgment calls, and relationship building: the work that drives differentiation rather than simply keeping the lights on.

Moving beyond the pilot

Most agentic AI efforts stall in pilots because they rely on experimentation instead of intentional design. To break the cycle, Martin believes that IT leaders must anchor AI initiatives to clear business outcomes, activate existing systems, and adopt a consistent enterprise-wide approach to deploying and governing agents.

Agentic AI isn’t just another tool – it’s an architectural shift. The organisations that recognise this will be the first to move beyond prototypes and realise real impact: faster workflows, more thoughtful decisions, and teams focused on higher-value work.