There is no shortage of interest in AI right now. Boards are asking about it. Budgets are being allocated. Vendors are promising transformation. And yet, there is a gap between what AI is supposed to do and what it is actually delivering.

That tension — between the promise of AI and the reality of getting it to work — is the subject of this playbook.

At HSO, we have spent more than three decades helping organizations redesign how they operate. We have seen what happens when technology is deployed without the right foundations in place, and what becomes possible when it is.

Over the coming pages, we set out the foundations that give organizations the best chance of getting AI pilots into production: people, process, data and cloud architecture. In that order.

Every chapter draws on the experience of HSO practitioners who have seen programs scale, and on the clients who have made AI work in production.

— Peter J. ter Maaten, Founder and CEO, HSO 

In most of these cases, the technology worked. The problem was the order of operations.

Organizations are deploying AI without looking at the fundamentals first. They aren't redesigning processes, bringing end users into the conversation, or establishing whether their data can support what they're asking AI to do.

For AI agents to make it into production and change how businesses operate, organizations need to sequence transformation differently…

Then they redesign the process. Not a single task; the entire end-to-end flow, from input to outcome, reconsidering how day-to-day operations should function when intelligent agents are part of the team.

Then, and only then, they layer in the technology. A trusted data platform, cloud architecture, purpose-built agents, and the governance to hold them accountable.

Each chapter in this paper draws on the expertise of HSO’s global practitioners, who work at the intersection of AI technology and business operations, delivering projects and services to companies in more than 60 countries worldwide.

HSO’s starting point is always the same: successful transformation does not begin with software. It begins with strategy, people, and a clear understanding of the processes you are trying to improve.

Every AI program has two timelines: the project plan—model selection, integration, testing, deployment, and continuing improvements—and the time it takes for people to understand what is changing, why it affects them, and how they are supposed to work differently.

Most organizations focus their attention on the first timeline while underestimating the second. The gap between an agent going live and the workforce absorbing AI into their daily operations is where the majority of programs lose their return on investment.

AI and agents, at their most effective, solve human problems. But when people are not central to their design, adoption suffers.

Gallup's research finds that when leadership has communicated a clear plan for integrating AI, employees are three times more likely to feel prepared to work with it, and 2.6 times more likely to feel comfortable doing so.

User-centered design (UCD)—the discipline of building solutions around the needs, behaviors, and realities of the people who will use them—has been a cornerstone of good product development for decades. AI makes it even more critical because, unlike traditional deterministic software that provides fixed outcomes, AI and agents work probabilistically. Since AI and agents are affected by a user’s specific context, how an agent behaves cannot be fully anticipated before people start using it.

HSO's UCD approach covers five stages, based on the proven Double Diamond methodology and tailored specifically for AI agentic implementations: 

By the time a line of code is written, the solution has already been tested, refined and validated by its end users.

The result: the technical build is faster, late-stage changes are fewer, and the project is de-risked because the people who will determine its success are already bought into it.

Leading organizations build AI centers of excellence: cross-functional groups responsible for governance, literacy, and continuous learning. They create structured environments where employees can develop skills and build fluency at their own pace. When people understand what AI agents can do and discover that they can use them without a technical background, their fear recedes.

Before designing any agent, leaders also spend time tracking how work currently gets done — the volume, the time, the effort, the exceptions — to strengthen the business case for AI programs and create the baseline against which an agent's performance is measured.

Done well, an AI Center of Excellence turns the people most likely to undermine an AI program into the people most invested in making it succeed.

At the other end of the scale from fear, many employees are already using AI unregulated. Data from UpGuard reveals that more than 80% of workers use unapproved AI tools in their jobs, while research by CybSafe and the National Cybersecurity Alliance has found that 38% share confidential data with AI platforms without their employer's approval.

HSO has seen the consequences of this “shadow AI” firsthand. One company discovered that its employees had been using unvetted AI tools independently to overcome local challenges. However, this non-malicious “shadow AI” involved tasks using confidential data, leading to a preventable (and high-risk) security gap for the larger organization.

The effect is twofold: it reduces the risk of ungoverned AI use, and it turns your organization’s most enthusiastic early adopters into advocates.

In traditional technology implementations, change happens once. The system goes live, people are trained, and the organization moves on.

AI does not work that way. Unlike traditional software, which stabilizes after deployment, agents require continuous evaluation, retraining and adjustment. Organizational culture must reflect this: weekly knowledge sessions, cross-functional feedback loops, and regular performance reviews of the agents running in production. Go-live is not the end of the change management process, but rather the beginning of a new phase of change management.

When deploying AI, the instinct is to look for tasks to outsource to an agent. Which parts of the job are repetitive? Which decisions follow a predictable pattern? Which steps could AI complete faster than a human?

It’s a reasonable starting point, and one that will yield efficiencies. But the organizations generating the most impactful results start with inputs, outputs and desired outcomes — then redesign the process around those, with AI as part of the answer.

For many companies, the barrier to process improvement is legacy: systems built over decades, customized beyond recognition, and running on infrastructure that was never designed to be changed quickly. The temptation is to bolt AI onto what exists. But that is why pilots stay pilots.

McKinsey data confirm this: high-performing organizations are three times more likely to have redesigned their workflows before deploying AI into them.

Not every process benefits from an AI agent. The ones worth prioritizing share a set of broad characteristics:

a. Map the Current Process in Detail

Log what the process costs today: the volume, time, effort and exceptions. This baseline is both your business case and the most meaningful measure of whether a future agent is delivering.

Mapping the current process can reveal surprising insights. One HSO client discovered that a single team was fielding hundreds of queries a month within what was meant to be a straightforward customer service operation. The answers existed in the company handbook, but at that volume, the team did not have the capacity to retrieve and relay them efficiently. An agent connected to the same knowledge base could resolve most queries instantly.

Your AI Center of Excellence will play a critical role in ensuring that these outcomes translate into meaningful operational change, and it is important to have both board-level leaders and operational experts in the room. Senior leaders commissioning AI programs tend to see the operation from a distance. They know which functions are expensive, which are slow, and which have been flagged in board meetings, but they do not always possess ground-level experience.

Team leaders understand each team’s volume of work, system workarounds, and the knowledge that exists only in people’s minds. Leveraging this combination of strategic perspective and ground-level expertise is key to identifying and achieving impactful outcomes.

c. Design the Agent Around That Outcome

With the outcome defined, the design work begins. From our experience across hundreds of agent deployments, there are four design principles that consistently determine whether an agent delivers in production:

We discuss the architecture of successful AI agents in greater detail in chapter four.

Before designing any AI solution, the most valuable thing your organization can do is understand exactly what your data estate looks like.

Think of data quality as an iceberg. Most organizations currently see around 10% of the factors that influence it: the duplicate records, the missing fields, the failed reports. The remaining 90% has been absorbed so completely into operations that it no longer registers as a data issue.

Getting beneath the surface of your estate — tracking the manual interventions, the workarounds, the time spent compensating for what your systems cannot do — will tell you which processes your data can support, and which it cannot.

A multi-year data remediation program would cost more than it delivers, take longer than anyone budgets for, and be partially out of date before it ends. A targeted approach is far more productive: identify what data a specific process requires, assess its reliability, and establish continuous controls to maintain it. In most organizations, this starts with a flagship initiative — a use case capable of delivering visible value early, while being sufficiently contained to manage data risk. That early success builds trust in the program, creates momentum, and gives the business confidence to improve the broader data estate in parallel.

With all data now visible and accessible through a single layer, stichd can extract value from every system in its estate, and the business can innovate without waiting for IT to catch up.

Enterprise integration landscapes are almost always more complex than they first appear. When HSO worked with a commercial bank to automate its consumer credit processing, the bank believed it had a manageable integration challenge: six dominant banking institutions covering 90% of statements. What nobody had mapped was that the remaining 10%, spread across 11,000 payment processors. An integration architecture scoped around the majority would have failed on contact with the rest.

Moving to the cloud makes enterprise AI architecture feasible, as on-premises infrastructure was not designed for the connectivity that agentic AI requires. It also frees up capital previously tied up in physical infrastructure — capital that can be redirected into the AI program itself — and removes the structural constraints that limit how agents connect and scale.

HSO has made MCP central to our agentic AI architecture. Clients are not building a new connection every time a system changes or a use case is added. Instead, they are operating from a governed, extensible layer that any agent, from any vendor, running on any model, can connect to.

Let us return to that meeting about the AI proof of concept that never made it into production. The question in the room was almost certainly, What went wrong with the technology? when it should have been, What should have happened before we built anything?

The organizations generating significant returns from early AI projects have not found better tools. They have done the groundwork — and in most cases, that groundwork has led them to build fewer agents than originally planned. However, the agents they are building frequently make it to production and deliver measurable, sustainable ROI.

Repeating this groundwork at enterprise scale takes structure and meaningful collaboration within the organization and with external partners. At HSO, we have seen firsthand the multifaceted benefits of developing an AI center of excellence. By bringing people, process, data, architecture, and delivery together in one team, organizations set themselves up for success from the very beginning.

Run well, an AI Center of Excellence turns each deployment into a foundation for the next — so agents are faster to build, cheaper to run, and more likely to deliver rapid ROI.