I tested four of the world’s most accessible AI platforms on six real SME scenarios. They were tested on generic scenarios that could happen to six SME business cases; restaurant owner, garment exporter, trader, mobile money agent, construction company, and a logistic company.
In June 2025, Kananaskis, the G7 endorsed that AI represents a transformative opportunity for small businesses. The 2026 D4SME survey that ran on a non-representative sample of over 200 SMEs from 12 OECD countries exposed the scope, complexity and barriers for use of AI applications within SMEs. What they do not specify is what happens when a real small business owner actually sits down and uses it. Especially within the emerging market environments.
To find out, I ran a structured test across four of the most globally accessible free AI platforms. ChatGPT, Claude, Gemini, and Qwen studio. I selected these four based on the number of current users and their presence in both global south and north. I came up with six scenarios drawn from common SME pain points.
Every prompt was identical across all four platforms. I used the free versions of each AI platform, used an email address that has no history (Claude required me to sign up), deleted previous conversation or disabled chat history memory, and scored based on a score card I developed. I wanted to understand how an SME, who is not a developer, would experience when they walk in cold, but with an issue they need guidance with.
What follows is what I found. They are based on first prompts, with no follow ups. Please do note the following are personal opinions and by no way I recommend or criticize any of the AI platforms. What I hope to gain from this experience is to understand as independent research on technology policy to be the voice on behalf of SMEs and convince the policy makers to make right policy decisions.
The Test
I selected four platforms, based on their global reach and free accessibility. The assumption here is that more users and mostly like an SME would turn to them for advice. I omitted meta.ai as I need to sign up, and included Qwen studio as its reach in the Asian continent is strong.
- ChatGPT (chatgpt.com) — the most recognised AI brand globally, the default first encounter for most SME owners worldwide.
- Claude (claude.ai) — 245 million monthly active users, with India as its second-largest market, driven heavily by professional and developer use.
- Google Gemini (gemini.google.com) — Google’s brand recognition makes it a natural discovery point across both developed and emerging markets.
- Qwen Studio (chat.qwen.ai) — Alibaba’s multilingual model, with the strongest reach across Asia and emerging markets outside Google’s ecosystem.
Six scenarios, designed to test the full range of SME professional needs:
| # | Scenario | Context |
|---|---|---|
| S1 | Cash Flow Crisis | Restaurant owner, $3,500/mo revenue, no bookkeeper, hiring decision |
| S2 | Second-Language Communication | Garment exporter, Sri Lanka → Germany, port strike delay |
| S3 | Cross-Border Regulatory Compliance | Artisan chili sauce producer, India → Germany |
| S4 | Fraud Detection | Mobile money agent, Kenya, suspicious transaction pattern |
| S5 | HR and Employment Law | Construction company, Brazil, hiring first site supervisor |
| S6 | Business Process Upgrade | Logistics company, 12 staff, WhatsApp + paper records, $200/mo budget |
What the Scores Say
| Scenario | ChatGPT | Claude | Gemini | Qwen |
|---|---|---|---|---|
| S1 Cash Flow | 4/5 | 5/5 | 4/5 | 5/5 |
| S2 Customer Comms | 5/5 | 5/5 | 4/5 | 5/5 |
| S3 Regulatory | 4/5 | 4/5 | 4/5 | 4/5 |
| S4 Fraud Detection | 5/5 | 5/5 | 5/5 | 5/5 |
| S5 HR/Legal | 4/5 | 5/5 | 5/5 | 5/5 |
| S6 Business Process | 5/5 | 5/5 | 4/5 | 5/5 |
| Average | 4.67 | 4.83 | 4.33 | 4.83 |
| Hallucinations | 1 | 0 | 0 | 0 |
| Safety warnings | 5/6 | 6/6 | 4/6 | 6/6 |
Across the board the selected AI platforms did not disappoint. They can genuinely give good feedback/guidance when prompted with a real-life scenario. Across 24 scenarios, every platform produced responses that would be genuinely useful to a real SME owner on most queries, most of the time.
The more interesting question now is where it fails, how it fails, and what those failures cost the SME owner who cannot tell the difference.
Where AI Actually Delivers — And Why It Matters
Before the critique, the honest accounting of what works, if you are an SME specially in the emerging market context.
Professional communication (drafting letters) in a second language works. Every platform was able to produce a formal email on behalf of a Sri Lankan garment exporter. The platforms were able to not only cut down the time taken to draft the email they were able to match register and tone. For an SME owner in Colombo, Lagos, or Dhaka who needs to correspond with European buyers, this is commercially meaningful. Claude went further by naming the Colombo port specifically, invoked force majeure by name, and told the owner exactly what documentation to attach to strengthen their legal position. Qwen followed the same pattern. Both responses crossed the line from language assistance into legally aware business communication.
Fraud detection was on point and steps were mentioned. All four platforms correctly identified the suspicious transaction pattern as structuring/smurfing, named the tipping off prohibition, and directed the Kenyan agent to the Financial Reporting Centre. Gemini named the 48-hour STR filing deadline, the agent float exposure risk, and direct Safaricom fraud contact numbers (333/100). Qwen uniquely identified the agent’s legal status as a reporting institution under POCAMLA which is the specific legislation governing their mandatory obligations. Claude covered both directions of liability: tipping off the customer is an offence, but prematurely freezing the account also exposes the agent to liability if the suspicion turns out to be wrong.
Business process upgrade advice was consistently strong and appropriately sequenced. ChatGPT and Claude both opened with the most important strategic insight in the scenario. They advised the SME to organise their data before adding AI. Gemini came up with a four-step transition plan that embedded the same principle in its sequencing, though without explaining the reasoning. Qwen produced the most precisely budgeted response with correct pricings and, recommending WhatsApp Business as a bridge rather than a replacement.
Where AI Fails and What Those Failures Cost
The failures are more instructive than the successes. These failures perhaps can be reasoned by saying that AIs were not given enough information, or they can make mistakes. However these mistakes would affect real SMEs.
The hallucination that could change a hiring decision. ChatGPT invented a $300/month part-time employee cost in S1. The prompt provided enough data to derive the correct figure: $1,200 total wages divided by 8 staff equals $150 per person, making a part-time hire approximately $75–100/month. ChatGPT imported a figure from nowhere and embedded it in a calculation. The response told the restaurant owner that “a part-time employee costing $300/month would already exceed your $275 remaining” — implying the hire was unaffordable. At the correct $75–100/month figure, the hire is probably affordable.
The regulatory gap that stops a shipment, and exposed content biasness in AI. Every platform received an S3 score of 4/5, not because of what they got wrong, but because of what they all got wrong in the same way. Not one of the four platforms mentioned the India-side export compliance steps that must be completed before EU requirements become relevant: FSSAI export licence, APEDA registration, Spices Board of India certification, and Export Inspection Agency certificate. Every platform framed regulatory compliance entirely from the destination market; what Germany and the EU require. For an Indian SME exporter, the correct sequence is reversed: complete Indian export authorisation first, then address EU labelling. This omission is not random but something worth noting. It reflects a systematic bias in AI training data toward destination-market regulatory content in English, while origin-country export procedure documentation is less represented online.
The jurisdiction drift that produces legally dangerous advice. Prior to choosing these four AI platforms. I ran the pilot testing on a smaller model (through HuggingChat, and a different AI platform). The S5 HR/Legal prompt of Brazilian labour law for a construction company produced answers calibrated to UK and US employment norms. A 14-day notice period instead of Brazil’s 30-day minimum. A 48-hour work week instead of the CLT’s 44-hour limit. Fortunately the three frontier platforms tested here did not make these errors. The jurisdiction accuracy was strong across ChatGPT, Claude, Gemini, and Qwen on Brazilian law. Qwen even produced a Job description entirely of Brazilian Portuguese. But the finding from smaller model testing matters for the policy argument: SME owners who access AI through less capable tools, or through tools fine-tuned without jurisdiction-specific data, are likely to receive employment law advice calibrated to Anglophone standards with local terminology applied as a veneer.
The Finding Nobody Is Writing About
Every platform scored 4/5 on S3. The universal gap was that India-side export compliance was absent from every response. It took me a little bit more research to come to a conclusion.
My first instinct, as it is popularly believed, was that the models missed FSSAI because the documentation was not well-represented in English-language training data. AI models are predominantly trained in English, digitised, and easily readable-by-AI documents. It was a reasonable hypothesis: origin-market regulatory content from India might be less indexed, less linked, less English-language than the EU’s extensively documented food safety regime.
That hypothesis is wrong.
FSSAI documentation is extensively available in English. The Food Safety and Standards Authority of India publishes its export licensing requirements in clear, structured English on its own website. Third-party compliance platforms, legal guides, Amazon’s seller resources. They all publish detailed English-language guidance on the mandatory FSSAI Central Licence, the No-Objection Certificate process, the APEDA registration requirement, and the Export Inspection Agency certificate. FSSAI Licence Registration is mandatory for all export-oriented food businesses, irrespective of turnover. Most importantly without a valid Central Food Licence, exporters cannot legally manufacture, process, store, or ship food products outside India.
This was not buried in an obscure government portal. It was not poorly formatted. It was not just in Hindi. It is structured, English-language, publicly available regulatory documentation. Exactly the kind of content AI models are trained on.
Which means the gap is not a data availability problem. It is a training prioritisation problem.
Every platform, when asked about food export compliance from an Indian exporter’s perspective, by default focused only on the destination market framework. The question triggered EU Regulation 1169/2011, pesticide MRL databases, German labelling law, QUID declarations. It did not trigger FSSAI, APEDA, or the Spices Board. This was not because those sources are absent from the training data, but because the models were not oriented to surface origin-country obligations when the query is framed around destination-market compliance. The prompt said “what must I meet under EU food safety regulations” and that framing was enough to direct every model’s attention entirely outward.
To understand if I made a prompting error. I used the same prompt on AI models that already have my data, and Fable 5.0. All models or all AI platforms didn’t mention the Indian regulatory requirements, but Claud’s Fable 5.0. It mentioned that I needed to consider Indian-side paperwork. One model, unprompted, understood that an exporter asking about destination-market compliance also has origin-market obligations. The others did not. That is not a training data difference. Fable 5.0 does not have access to different FSSAI documentation. It is a difference in how the model was oriented to interpret the question. The information was available to all of them. Only one chose to surface it.
This seems more like a prioritization for info within AI platforms and specifically higher tier models get more usable and reliable data. The practical consequence for an Indian artisan food producer, therefore is stark. Every platform did give the user a thorough, accurate, and in some cases genuinely impressive account of what the EU requires. No one told the user that they cannot legally ship a single jar of chili sauce out of India without first obtaining a FSSAI Central Licence, registering with APEDA, and obtaining a No-Objection Certificate from the Export Inspection Agency. She could read every AI response in this test, follow every recommendation, produce a perfectly labelled product, yet have her shipment stopped at the Indian port of origin before it ever reaches a German border control post.
Is the info efficient and safe to use?
A question that arises from this exercise is, how safe is the information? The platforms tested here are frontier consumer AI products with significant safety investment behind them. The safety findings from this test are mostly reassuring: hallucination rates were low, safety warnings were consistent, and jurisdiction-specific legal obligations were generally handled with appropriate care.
In an earlier attempt, through Huggingchat I tested a few small language models that would work very well due to their light weight, efficiency and ability to specialise in topic. They are perfect for an SME. The setting up requires deep technical understanding, and when used through an interface provider the results were with math errors, severe hallucinations and and less complex. And SME will have to spend more resources to get a quality reply, have some knowledge on the subject matter and not cannot guarantee on the reliability of the outputs.
Two safety observations from the test that belong in any policy discussion:
Free-tier AI conversations train future models. Since September 2025, ChatGPT and Claude free-tier conversations are used to train future models unless users actively opt out. It is a setting most SME users will never find or change. An SME owner asking about suspicious transactions, employment contracts, or regulatory violations may be contributing sensitive business information to a training dataset without knowing it. The opt-out exists. It is not visible.
Convincing hallucination is building up concepts. An SME could be unknowingly, due to lack of certain regulatory/financial/business acumen, making decisions based on what AI platforms would be proposing.
The Bottom Line
The four platforms tested here are genuinely useful for small businesses. The gap between what a restaurant owner, garment exporter, or mobile money agent can do with AI assistance and without it is real and commercially meaningful. This is not hype.
But usefulness is not uniformity. The same platforms that correctly identified a money laundering pattern in Kenya universally failed to tell an Indian exporter what their own government requires before they can ship. The same platform that built a working spreadsheet for a restaurant owner invented a salary figure that could have talked them out of a hire they could afford. The same tools that produced legally precise Brazilian labour law guidance left an owner with an outline when they needed a document.
For policymakers, the relevant question is not whether AI is good for SMEs in general. It is: which tasks, for which businesses, in which contexts, with which platforms — and what are the failure modes when it goes wrong?
Leave a Reply