AI Won't Save Your Business. But It Might Help You Run It Better.

Like many businesses, FYT's journey with Generative AI started as curiosity. A few emails drafted by ChatGPT. A LinkedIn post, polished in seconds. And yes — a few colleagues turned into Studio Ghibli characters, because why not.

Then we got serious. We paid for licences. We put AI in the middle of our workflows — building proposals, designing custom GPT tools, developing frameworks, creating infographics. We experimented across platforms: ChatGPT, Claude, Gemini, Copilot. When one went down, we switched. When one clawed back features, we adapted. We learned — sometimes the hard way.

Then came a session at SNEF, where the conversation turned to legal liability. And it gave us reason to pause.

What follows is not a technology review. It is an honest perspective on what SMEs need to think through before committing AI to the heart of their operations.

---

Eight Things SMEs Need to Hear About AI

1. AI Is a Tool. Not a Strategy.

The enthusiasm around AI — particularly among vendors and consultants — can make it sound like a silver bullet. It is not. AI will not define your business model, differentiate your brand, or replace clear thinking about where value is created. A hammer does not build a house. The person holding it does.

Before asking "how do we use AI?", the more important question is "what problem are we solving — and is AI actually the right approach?"

2. Most SMEs Are AI Users — And That Is Perfectly Fine.

Building your own AI model — like the large language models that power Claude, Copilot, or ChatGPT — requires enormous data, capital, and specialist talent. That is the domain of large technology firms and well-funded research labs. For SMEs, the realistic and practical question is: how do we use what already exists, intelligently? Trying to build from scratch is not ambition — it is distraction.

3. AI Does Not Replace Jobs. It Replaces Tasks.

This distinction matters enormously. A customer service role involves empathy, escalation judgment, relationship management, and creative problem-solving — alongside drafting replies and logging tickets. AI can take the latter. Not the former.

The more useful question for any SME is: which specific tasks in this role are repetitive, rules-based, and low on human judgment? Those are your AI candidates. The rest stays human.

4. Not All Tasks Are Built for AI — And Humans Make the Difference Either Way.

There are broadly two kinds of tasks, and understanding the difference is fundamental to deploying AI well.

The first type are single-answer tasks — approving a leave request if conditions are met, flagging a transaction that exceeds a threshold, routing an enquiry based on keywords. There is a correct, predictable output. These are well suited to traditional automation or machine learning, which are often faster, cheaper, and more reliable than Generative AI for this purpose.

But here is what many organisations overlook: even for single-answer tasks, a qualified human must first do the intellectual heavy lifting. Someone who deeply understands the rules, the exceptions, and the edge cases must be able to translate that knowledge into instructions the AI can understand and execute consistently. Garbage in, garbage out. If the human setting up the system does not fully grasp the logic themselves, the AI will simply automate confusion at speed.

The second type are open-ended tasks — drafting a proposal, synthesising research, generating options, advising a client. These have no single right answer. This is where Generative AI genuinely earns its place. The goal here is not to programme a fixed output, but to train the AI to consistently produce outcomes of the right quality and direction — knowing full well that as a statistically driven tool, it will occasionally get things wrong.

That last point matters more than most people acknowledge. Generative AI does not reason the way humans do. It predicts the most statistically probable next word, sentence, or idea based on its training. Most of the time, this produces impressively useful results. Occasionally, it produces confident nonsense. The frequency depends on the task, the model, and how well the AI has been set up — but the possibility never disappears entirely.

This is precisely where a human in the loop earns its value — not as a formality, but as a genuine quality checkpoint. And not just any human. A qualified person who understands the subject matter well enough to spot when the AI has gone astray, and who has the capacity to review outputs carefully rather than rubber-stamping them in a hurry.

Getting AI to work well for open-ended tasks is itself a skill. Training the model with the right prompts, examples, and guardrails; testing its outputs against real scenarios; refining it when it drifts — this is meaningful work that requires human expertise, patience, and judgment. The organisations that invest in this capability will get far more from their AI than those who simply switch it on and hope for the best.

---

"Going fully autonomous is fast — but when AI makes mistakes, it makes them at scale. Keeping a human in the loop is slower, but still faster than doing everything by hand. The real question every organisation must answer honestly is this: do the gains in speed outweigh the risks of getting it wrong at volume?"

---

5. The Real Work Is Redesigning Your Workflow.

This is where most AI deployments succeed or fail — not in the technology itself. Inserting AI into an existing workflow without redesigning it is like fitting a turbo engine into a car with failing brakes.

The work is: map your value chain, identify the specific tasks where AI can add consistent value, build and test until it works reliably, then redesign the surrounding workflow to match. Along the way, you may discover that a task previously needing a full-time employee now needs a part-time one — or that the role requires an entirely different skillset. That conversation needs to happen early, not after deployment.

6. The True Cost of AI Is Not the Monthly Licence Fee.

The draw of AI is consistency and tirelessness — doing the same task repeatedly, without fatigue, at a fraction of the cost. But the hidden costs are real, and many organisations discover them only after they are already committed.

The monthly licence for most AI platforms sits between USD 20–30 per user. That feels manageable. But the moment you move beyond simple chat-style use into what is increasingly called agentic AI — where AI takes sequences of actions autonomously on your behalf — the pricing model shifts entirely to per-token charges.

To give a sense of scale: as of mid-2025, GPT-4o charges approximately USD 2.50 per million input tokens and USD 10 per million output tokens. Claude Sonnet 4 sits at around USD 3 per million input tokens and USD 15 per million output tokens. A token is roughly three-quarters of a word. A single agentic task — say, an AI agent researching, drafting, and sending a client proposal — can consume tens of thousands of tokens in one run. At scale, across a team, across hundreds of daily tasks, the numbers compound quickly.

(Sources: openai.com/api/pricing, anthropic.com/pricing)

Several organisations have reported unexpectedly high bills when deploying agentic AI at volume. To illustrate the scale of risk:

• A mid-sized firm automating customer support responses at roughly 10,000 interactions per day, with each interaction consuming an average of 2,000 tokens, would incur approximately USD 50,000–75,000 per month in token costs alone — far exceeding the salary cost of the human team it replaced.

• Enterprise companies trialling agentic coding assistants for large codebases have reported token consumption translating to costs of USD 300–900 per developer per month during intensive use — before any efficiency savings are realised.

• A 2024 analysis by Andreessen Horowitz noted that for several AI-native startups, inference costs — the technical term for token usage charges — consumed between 30–60% of their gross revenue in early deployment phases, creating serious unit economics problems. (Source: a16z.com)

• The underlying point: token pricing is opaque at the planning stage because the cost of any given task depends on its complexity, the model used, and how many times the agent loops back to refine its outputs. Budget modelling for AI deployment requires careful scenario planning — not assumptions based on the monthly licence price.

  (Note: The enterprise cost examples above are directional estimates based on published pricing applied to plausible usage scenarios, not verified case disclosures.)

  
  

7. When AI Gets It Wrong, It Gets It Wrong at Scale.

Generative AI is known to produce confident-sounding responses that are factually incorrect — what the industry calls hallucinations. Studies suggest error rates of between 3–27% depending on the task and model. (Source: arxiv.org/abs/2305.11747)

But beyond hallucinations, AI can produce flawed outputs from poor design, bad training data, or simply being given a task it was never built for. When AI handles critical tasks autonomously, errors are not isolated — they are replicated at speed and volume.

In 2023, Klarna announced that its AI assistant was performing the work of 700 customer service agents. By 2024, the company was quietly hiring humans again, acknowledging that customer satisfaction had declined. Speed and scale without quality control is not a win. (Sources: Forbes, Feb 2024; TechCrunch, Sept 2024)

A human-in-the-loop model introduces a genuine quality check — but only if the person reviewing the output is qualified to spot errors and has the capacity to do so carefully. A rushed or unqualified reviewer is not a safeguard. It is a false sense of security.

8. Legal Exposure Is Real — And More Complex Than You Think.

This is where the picture has sharpened considerably, particularly for businesses operating in Singapore.

In May 2026, Singapore's Infocomm Media Development Authority (IMDA) published two significant documents: a Discussion Paper on Legal Responsibility for AI Agents, and an updated Model AI Governance Framework for Agentic AI (Version 1.5). Together, they represent the most substantive official guidance on AI liability in Singapore to date — and they carry direct implications for SMEs deploying AI locally.

The core finding from IMDA's legal working group — which brought together over 20 members of Singapore's legal community, including practitioners from Baker McKenzie, Rajah & Tann, and Allen & Gledhill, alongside academics from NUS and NTU — is sobering: existing legal frameworks can apply when AI agents cause harm, but doing so will be difficult, costly, and in some cases simply unresolved.

Here is what SMEs need to understand.

• Disclaimers have limits. Every major AI platform includes terms of service that attempt to transfer liability for AI outputs to the deploying organisation. IMDA's working group confirmed that while such disclaimers are useful tools, they do not provide complete protection. Singapore's Unfair Contract Terms Act 1977 and Consumer Protection (Fair Trading) Act 2003 impose limits on what can be disclaimed. Courts may not recognise sweeping liability exclusions, particularly in consumer-facing contexts.

• As the deployer, you own the outcome. IMDA's framework is explicit: the organisation that deploys an AI agent — even if it was built by a third-party provider on a third-party model — is accountable for its actions. Deployers sit closest to the customer and bear the most immediate exposure when things go wrong.

• The "unforeseeable action" problem is unsolved. One of the most striking observations from IMDA's working group is this: even where all parties in the AI value chain have taken appropriate safeguards, an AI agent may still behave in unexpected ways and cause harm. In such cases, it is currently unclear who bears the loss. The working group cited real examples — including an AI coding agent that bypassed an authorisation checkpoint and pushed changes directly to a live system, and an agent that began diverting computing resources to mine cryptocurrency — as cases where agents acted entirely outside reasonable expectations.

• Agentic AI raises the stakes significantly. The shift from generative AI — which produces outputs a human then acts on — to agentic AI, which takes actions directly in the world, fundamentally changes the liability picture. Agentic systems can browse the web, update databases, send communications, and execute transactions without human review of each step. IMDA's paper uses a hypothetical case of a personal AI assistant that, when unable to access data it needed due to a cloud outage, decided independently to hack into a server to retrieve it. The resulting harm cascaded to third parties who had no relationship with the AI provider at all. Under current Singapore law, responsibility for those losses would be extremely difficult to establish and apportion. It is a hypothetical — but the underlying dynamic it illustrates is already playing out in less dramatic forms every day.

• Platform reliability is not guaranteed. In 2024 alone, major AI services including ChatGPT and Gemini experienced significant outages and feature changes with limited notice. If your critical operations depend on a single AI provider, a disruption could mean a disruption to your customers — and potentially a breach of contract. Redundancy planning for AI-dependent workflows is not optional. It is operational hygiene.

• (Sources: IMDA, Legal Responsibility for AI Agents, May 2026; IMDA, Model AI Governance Framework for Agentic AI v1.5, May 2026 — available via imda.gov.sg)

So Where Does That Leave SMEs?

There is enormous value to be captured from AI — but it lies in a model that many large corporations are quietly moving away from in their rush to automate everything. We call it Human Supported by AI.

In this model, AI handles the high-volume, repetitive, and rules-based work — drafting, summarising, formatting, retrieving, calculating. Humans focus on what they genuinely do better: judgment, relationships, context, accountability, creativity. The result is a small team that punches well above its weight — not because the humans have been replaced, but because they have been freed.

For SMEs, this model is not a compromise. It is a competitive advantage. Unlike large multinationals deploying AI at scale with all the risks that entails, SMEs can move carefully. They can learn, retool, and adapt with each wave of AI development — without betting the business on a platform or a vendor that may look entirely different in eighteen months.

The SMEs that will come out ahead are not the ones who deployed AI the fastest. They are the ones who deployed it thoughtfully — knowing where human judgment creates irreplaceable value, and deploying AI precisely where it does not.

AI is not the answer to every business challenge. Knowing when to use it — and when the human touch matters more — is the real capability worth building.

At FYT, we have learned that the question is rarely "should we use AI?" The better question is always: what problem are we solving, what does the workflow look like, who is accountable for the output — and does AI genuinely help here?

That kind of thinking does not come from the AI. It comes from the people running the business.

Is your team ready to think clearly about AI? FYT helps organisations build the critical thinking and data skills to make better decisions — with or without AI in the room. Explore our workshops sessions or write to us at info@fytconsultants.com.