Michael Goh: Why Singapore SMEs struggle to move AI beyond the pilot stage

AI adoption figures frequently obscure how far a company has truly advanced. Individual employees may already be using generative AI for specific tasks, while management teams experiment with targeted pilots. Turning that activity into repeatable business processes requires clearer objectives, reliable data and workflows that reflect how work is actually done.

Evidence from IPI Singapore’s collaborative work with small and medium-sized enterprises (SMEs) shows that while many businesses reach the trial or testing stage, genuine integration into routine operations remains uncommon. Several fundamental obstacles routinely stall this momentum. Projects frequently run into difficulties when the core business problem is poorly articulated, or when documented processes fail to align with how staff actually execute their duties on the ground. Furthermore, fragmented and incomplete data repositories regularly derail technical execution, whilst bringing employees into the rollout too late undermines internal buy-in and stalls practical adoption.

In this editorial interview, Michael Goh, Chief Executive Officer of IPI Singapore, outlines how SMEs can gauge whether an AI use case is genuinely ready for deployment and detect operational bottlenecks before scaling up pilot schemes. He also details practical methods for involving staff early in the process. Additionally, Goh highlights the role of advisory support and technology matching, explaining how these resources enable organisations to refine the specific problems they face before evaluating potential solutions.

SME AI adoption remains concentrated around experimentation

Headline adoption figures can combine individual usage, experimentation, pilots and production deployments. Based on IPI’s work with Singapore SMEs, where are most companies actually positioned in that journey?

Michael Goh: Most SMEs we engage with have moved beyond exploration. They are typically at the experimentation and piloting phases. That said, this is largely due to the nature of IPI’s services, Innovation Advisory and Technology Matching, which attract SMEs seeking assistance in adopting innovation for business transformation.

What we see is that people are using AI at the individual level. In some SMEs, business leaders are using these tools to drive productivity. These are encouraging signs. What is less common is the deliberate and mindful integration of AI into business processes, at least not with the SMEs we are working with at this juncture. This is not unexpected, because doing so requires considerable deliberation and effort from business leaders to gain clarity on the one business problem that needs solving and will yield the fastest ROI.

AI adoption is like any other innovation adoption. It is a journey. We work with SMEs on this journey across four stages: Discover, Strengthen, Scale, and Expand. Most arrive somewhere in the middle. They have taken a first step but need a roadmap to move from activity to outcome. Finding the right tool is rarely the hard part. Understanding where you actually are, where you need to get to, and what has to change internally to get there is where most businesses need help. This is where many of the SMEs we engage with find themselves.

A pilot should begin with evidence that the use case is ready

Starting with a business problem is widely accepted advice, yet unsuitable use cases still continue to reach the pilot stage. Where does use-case selection go wrong, and what evidence should an SME require before approving an AI project?

Michael Goh: Before any AI project gets approved, there are three questions an SME should be able to answer clearly. What specific outcome are we trying to improve, and how will we know if it has improved? Have we mapped the process this AI will touch, and does that map reflect what actually happens on the ground, not what we assume happens? Is the data available, and is it in good enough shape to produce outputs we can act on?

If any of those three cannot be answered with confidence, the project is not ready. That is not a reason to walk away. It is a signal of exactly where the preparation work still needs to happen.

There is one question that is not asked often enough. If AI fails after adoption, as any tool or machine can, what is the fallback plan for the next 30 days if the system cannot come back online? This question becomes even more important when AI is fully integrated into core work processes as agentic AI, or deployed in physical hardware.

We have seen autonomous vehicle fleets suddenly stop or cause accidents due to software issues. That does not mean we should stop adopting AI. Instead, we should apply well-tested systems engineering principles and treat AI as an integral part of a larger system that can fail. Ask the hard questions: what if AI becomes a single point of failure? As a smart adopter of AI, we need to address these questions upfront and plan for them. Not doing so would be irresponsible and unwise.

AI exposes the gap between documented and actual workflows

Once the use case has been established, which weaknesses inside an organisation most commonly prevent the technology from producing the intended result?

Michael Goh: There is often a gap between how work is documented and how it actually gets done. Most organisations have a version of their processes on paper. For a successful AI integration, you need an accurate picture of what people on the ground are actually doing, and an understanding of why they have deviated from the SOP. When AI is built on the assumed version, it optimises something that does not quite exist. Staff find workarounds, and adoption quietly falls apart.

On data, the problems cluster around three things: silos, inconsistency, and analogue capture. Data sitting in separate systems that do not talk to each other. Data recorded differently by different teams doing the same job. Data that still exists on paper or in someone’s head. Each one limits what AI can reliably do, and together they can make a sound use case unworkable. There is also the question of data quality and the scope of what is captured. If information is not captured at all, there is no way for AI to help.

The decision-making gap is a common issue in agentic AI integration and can result in damaging decisions if left unchecked. If the system surfaces a recommendation and there is no clear, timely process for acting on it or questioning it, the insight just sits there. Accountability for AI-assisted decisions has to be designed in deliberately. It will not sort itself out.

Staff involvement has to begin before configuration

Process redesign can affect how employees work, which tasks remain manual and where responsibility sits. When should the people affected by an AI deployment become involved, and who should own the project in an SME without a dedicated transformation team?

Michael Goh: The earlier the better. By the time a tool is being configured, the window for genuine involvement has mostly closed. Staff who are brought in at go-live receive a decision that has already been made. That distinction, between being involved and being informed, matters more than most people realise, and it shows up directly in whether people actually use the thing.

The people doing the work being targeted need to be in the room when the process is being mapped. They will surface things no vendor assessment will catch, and they are far more likely to back a solution they helped shape than one that arrived and was explained to them.

On ownership, the most successful AI projects we see in SMEs are owned by someone with a direct stake in the business outcome. Someone who knows what success looks like in business terms and has the authority to keep things moving when the project hits friction. The technical questions can be navigated with the right support. The business judgment cannot be outsourced.

Instead of framing AI purely as a productivity tool, which in certain contexts can be interpreted as trimming headcount or removing steps, what if we reframe the narrative by positioning AI as an enhancer that enables staff to achieve more and be more creative in how they deliver impact? That is a far more compelling way to bring people along.

Workarounds are an early warning that a pilot will struggle to scale

Performance metrics may take time to establish during a pilot. Before those results are conclusive, which operational or behavioural signals indicate that an AI implementation is already moving in the wrong direction?

Michael Goh: The earliest signal is almost always behavioural. When the people who were supposed to use the tool start working around it, running a parallel process, logging things manually, or simply not logging in, something is wrong. If the staff have been involved right from the start, it is rarely about resistance to change. It is usually that the tool does not match how the work actually happens, or that nobody has established why the output should be trusted.

The second signal is outputs that look plausible but do not drive action. If recommendations are consistently being overridden or ignored, either the model lacks the right data, or it is solving the wrong problem. Either way, it adds no value, and continuing to invest will not change that.

Advisory value begins before vendor selection

Technology marketplaces can make it easier to discover potential providers, but SMEs may still have to determine whether they are solving the correct problem. Where can an advisory and matching process change the decision before vendor evaluation begins?

Michael Goh: An SME evaluating vendors on its own is testing answers to a question it has already written. The problem is that the question is often the first thing that needs to be challenged. Vendors respond to what they are asked. If the use case is poorly framed, a capable vendor will still deliver something that does not move the needle, and both sides will have technically done their job.

The advisory layer changes what gets examined before any matching happens. IPI pressure-tests the business case and works with business owners to define or refine what success looks like in measurable terms. That removes one of the most common reasons implementations fail, which is that the brief was wrong to begin with.

Technology matching has to be measured after the introduction

TechInnovation and AIMX Singapore, taking place on 26–27 August, brings technology discovery, matching and enterprise AI adoption together on a single platform. What would demonstrate that combining those functions has produced a meaningful outcome for participating companies?

Michael Goh: It signals something the market needed. There should be no separation between innovation discovery and AI adoption. For too long these conversations happened in different rooms, and the companies that needed both were left to connect the dots themselves. Recognising this, IPI and MP Singapore joined forces to bring TechInnovation and AIMX, platforms each organisation owned independently, together as a pilot to address this friction point and deliver value to the regional ecosystem and local enterprises.

A business that finds a relevant solution at a showcase still needs help assessing whether it fits, structuring a pilot, and managing what comes after. Putting those capabilities in the same place reduces the gap between a promising introduction and an actual outcome.

On measurement, attendance and introductions are proxies for potential, not evidence of impact. The outcomes that matter are downstream. How many introductions led to a structured engagement? How many progressed to a PoC? How many translated into a commercial or operational outcome? At IPI, those are the questions we hold ourselves accountable to.

Execution is becoming the dividing line in SME AI adoption

Evidence from IPI’s engagements with SMEs shows that obstacles to AI adoption often become apparent long before a pilot generates definitive performance figures. Everyday warning signs, such as staff operating parallel workflows, avoiding the tool entirely, or regularly disregarding its outputs, typically indicate that the software is misaligned with practical working routines or has failed to earn user trust. In these situations, simply upgrading the model or tacking on extra capabilities will do little to fix the underlying problem.

This raises the standard for advisory and technology-matching programmes. Introductory meetings and pilot trials may indicate active interest, but their value depends on whether projects progress into practical deployments that improve operational or commercial outcomes. For SMEs with limited capacity to absorb failed IT investments, decisions made before and during rollout can determine whether an AI project becomes part of the business or ends with the pilot.