AI Feasibility Studies: Can You Rely on Them, and When Do You Need an Adviser?

Type into any general AI chat tool, "Write me a feasibility study for a café in Nasr City", and within a minute you get several tidy pages: an executive summary, a market analysis, costs, a profit table and a recommendation to go ahead. It looks convincing. The real question is whether the numbers are right enough to put your savings on.

This article answers that question frankly, from the position of a team building an AI-based feasibility study product in Egypt. We explain what AI does well, where it goes wrong and why, how to test any study written with it, and when it cannot replace a human adviser.

What does "an AI feasibility study" actually mean?

The phrase is used for three very different things, and it helps to separate them before judging. The first is a founder describing the idea in a general chat tool and asking for a study; the model writes the whole text from memory. The second is templates or software that fill ready-made tables with numbers the user enters, with generated text around them. The third is specialised platforms that combine current market research with sources, a language model that proposes assumptions and writes the explanation, and a fixed calculation engine that computes the numbers.

Judging "AI feasibility studies" as one category is not very useful, because the gap between these three types is wider than the gap between any of them and a traditional study. What decides the quality of a study is where the numbers came from, how they were calculated, and whether you can check them.

What AI does well in a feasibility study

In four areas, AI is clearly better than the traditional approach:

• Structure: it knows the usual sections of a feasibility study and orders them, so the risk or licensing section is not forgotten.
• Speed: days of writing become minutes, which lets you try more than one idea or scenario before deciding.
• Breadth of research: models connected to search engines can gather competitor prices, rents and wages from many sources quickly.
• Clear writing: turning numbers into an explanation a founder understands, which many dry accounting studies lack.

These advantages are real, and they are why government bodies have started to use AI in this area too. In July 2026 the AI service «Ehsab Forsetak» ("calculate your opportunity") was announced; it prepares a preliminary feasibility study for industrial projects in minutes, as part of the digital support system for manufacturers¹, a step in line with Egypt's National AI Strategy².

Where AI goes wrong

1. Confidence in false information

Large language models are designed to produce text that sounds right and reads consistently, not to guarantee that every statement in it is true. Research has a name for this: "hallucination", when a model states something false or non-existent with the same confidence as something true³. In a feasibility study it shows up as a market-size figure with no source, a competitor price that does not exist, or a tax rule with the wrong rate.

2. Out-of-date information

A model learns from data up to a certain date. In a country that has recently seen waves of inflation, exchange-rate moves and a run of tax amendments, old prices mislead quickly. For example, in July 2026 several laws amending tax legislation were issued, including an amendment to the VAT law; the standard 14% rate stayed the same, but other details changed⁴. A model that does not search current sources will not know that.

3. Calculations that contradict each other

The profit table in a feasibility study is a chain of linked operations: sales from demand, costs from sales, cash flow from both, and break-even and capital from the cash flow. When a model writes each number as free text, contradictions appear: break-even in month 6 while the cash-flow table does not reach it before month 10, or a capital requirement lower than the investment cost itself.

4. Unjustified optimism

Ask a general tool about your idea and it tends to help you carry it out, not to talk you out of it. That is why many such studies recommend going ahead and assume high sales from the first month. A study that cannot say "no" is not a feasibility study.

5. A different answer every time

Ask the same tool about the same idea twice and compare the numbers. In our own tests while building Kleverz, running the same model on the same data could produce a peak cash need that differed widely from one run to the next. That does not make the model bad. It means one number from one run is not a basis for a decision.

How to test an AI-written feasibility study: 8 questions

Before you trust any study, whether written by AI or by a person, ask it these questions:

A study that fails two or more of these questions needs reviewing before use, whoever wrote it. Each item is explained in detail in our guide to feasibility studies in Egypt in 2026.

A general tool versus a specialised platform

The answer is not to drop AI but to use it for what it does well and not for what it does badly. That principle is how the Kleverz detailed study is built. We describe it openly here because it works as a yardstick for judging any similar platform:

1. Research with visible sources: the model searches the web for current prices, competitors, rents and wages in Egypt, and returns with each fact the link it came from, so you can open the source yourself. Grounding a model's answer in retrieved sources is a method research has shown to make model answers more factual⁵, and Google's Gemini API offers it as grounding with Google Search⁶.
2. Numbers from a calculation engine, not from the model: the model only proposes assumptions (price, cost, reachable customers, fixed costs); fixed code then calculates 36 months of cash flow, break-even, capital and scenarios. One part cannot contradict another.
3. Three independent drafts: instead of relying on one run, the assumptions are built three times independently and the middle one is used. If the drafts disagree widely, confidence is lowered and the study says plainly that the result depends on an uncertain assumption that should be tested first.
4. Your numbers first: if you give your price, cost and expected monthly sales, the study is never more optimistic than you are.
5. Legal rules from a reviewed list, not from the model: wages, social insurance, taxes and licences come from a reviewed list of facts with their sources; the model does not write them from memory.
6. A check before delivery: the study is delivered only when every promised section is complete; if the check fails, it is regenerated automatically.

Example: the same idea, two ways

Suppose a founder wants to open a small coding centre for children in a residential neighbourhood, with EGP 300,000 of capital. The figures below are an illustration of the difference in thinking, not a study of a real project.

A typical answer from a general tool starts with the size of Egypt's education market, assumes the centre attracts 80 students on a monthly subscription from month one, reaches a net profit in month three and ends with a strong recommendation to go ahead. It does not say where 80 students came from, does not calculate the employer's insurance share for the trainers, does not ask about licensing for an education activity, and does not test what happens if half that number come.

The structured approach starts from a different question: how many families within reach of the neighbourhood have children of the target age, how many can be reached each month with the marketing budget, and what share will sign up? If the answer is 25 students after six months rather than 80 from the first month, everything changes: break-even comes later, the capital needed rises, and the right decision may be to start with one trainer and a room rented by the hour instead of a full centre. The study also flags that training centres with premises have licensing requirements to confirm before signing a lease.

The difference between the two is not the quality of the writing. The second forces every number to come from somewhere that can be challenged. That is what makes a study a decision tool rather than a file.

How to read a study's confidence level

When assumptions are built more than once independently, you can measure how far they agree. In the Kleverz detailed study, the confidence level appears directly under the verdict, at three levels:

• High confidence: the independent drafts broadly agreed on the cash need and the break-even month. You can build on the result while tracking the weekly indicators the study lists.
• Medium confidence: the drafts agreed on the direction but differed on details. The result is fit for a decision; watch the most sensitive assumption in the first months.
• Low confidence: the drafts disagreed on a basic question, such as whether the business breaks even at all. The verdict is then "validate first", and the study names the assumption that flips the result, such as sales volume or margin, and how to test it cheaply before spending your capital.

Low confidence is not a flaw in the study. It is useful information: the available facts are not enough to settle the decision, and the right next step is to collect a real fact from the market before investing.

What to give the tool for a more accurate result

The biggest source of inaccuracy in any study, human or automated, is vague input. These details make the most difference:

If you are not sure of a number, give your best estimate. A good study compares your estimate with market prices and tells you if it is far off, which is better than leaving it blank for the model to guess.

When do you need a human adviser?

AI answers "should I start, and how much do I need?" well. But in some cases a human adviser is a necessity, not a luxury:

• When the lender asks for professional certification: Banque Misr, for example, requires "a feasibility study from a certified public accountant" for newly established projects in its small-business finance programme⁷. You then need someone to review and sign the study.
• In heavily licensed activities: medical laboratories, manufacturing, tourism and financial activities, where licences drive much of the cost and timeline.
• When there are partners: splitting shares, valuing non-cash contributions and drafting the company contract need legal and financial expertise.
• For foreign investment: the legal form, residency and restrictions on some activities for foreigners need specialist review before capital goes in.

In our view the practical formula combines both: an AI study settles the first decision quickly and cheaply, then a specialist adviser handles the steps that need a signature, a negotiation or a registration. That is why Kleverz works with Maher Consulting Firm, which handles company formation, licensing and tax after the study.

Do small business owners actually use AI?

In an OECD survey of more than 5,000 small and medium enterprises in seven countries in 2024, generative AI was in use in 31% of them. The sample is from OECD countries, not Egypt, but it shows how quickly these tools are spreading among small businesses⁸. As use spreads, the important question is no longer "should I use AI?" but "how do I make sure what it says is true?", which is what this article has tried to answer.

Common mistakes when using AI for a feasibility study

1. Copying numbers as they are: treating every number the model writes as fact, without opening its source or comparing it with a real market price.
2. Hiding uncomfortable numbers: adjusting the inputs until the tool produces the result you want. The study becomes a way to persuade yourself, not a decision tool.
3. Reading only the summary: reading the recommendation on page one without checking the cash-flow table, which is where the real problems show.
4. Ignoring the date: using a study prepared a year or more ago in a market where prices and rules change quickly.
5. Confusing a study with certification: assuming any study will do for a finance file, when the lender may require professional certification.
6. Not returning to the study after opening: a good study sets weekly indicators to track after launch, and comparing them with reality in the first months is its most valuable use.

How to use AI safely yourself

If you want to use a general tool to form a first view, these simple rules reduce its errors:

1. Give it your numbers: your price, cost, area and capital. The more specific the input, the less room for guessing.
2. Ask for a source for every number: and open the sources yourself. If the number is not in the source, do not rely on it.
3. Build the table yourself: move the assumptions into a spreadsheet and calculate cash flow and break-even yourself, rather than relying on numbers written in the text.
4. Ask for the pessimistic case explicitly: what if sales are a quarter lower? What if opening is two months late?
5. Ask twice: if the numbers differ widely between the two answers, the key assumption is unsettled and needs testing in the real world.
6. Check legal rules at their official source: wages, insurance and taxes change; do not take them from the model's memory.

Conclusion

AI has changed the cost and speed of a feasibility study, but not what makes a study good: numbers with sources, demand built from customers you can reach, linked calculations, a pessimistic case, and the courage to say "no" when the numbers say no. Use AI for what it does well, which is research, structure and explanation, and do not make it the only source of truth. When the decision is bigger than whether to start, such as certified finance, a partnership or complex licensing, add specialist human expertise.

Frequently asked questions

Sources

1. Akhbar El Yom. أيمن البياع: إطلاق خدمة «احسب فرصتك» بالذكاء الاصطناعي للمشروعات الصناعية (in Arabic) (2026-07-28) — secondary source; El Watan. خبر إطلاق خدمة «احسب فرصتك» (in Arabic) (2026-07-28) — secondary source. Checked 14 October 2026.
2. Egypt National Council for Artificial Intelligence (ai.gov.eg). الاستراتيجية الوطنية للذكاء الاصطناعي، الإصدار الثاني (2025–2030) (in Arabic) (2025-01). Checked 14 October 2026.
3. ACM Computing Surveys. Ji, Z. et al. Survey of Hallucination in Natural Language Generation, vol. 55, no. 12 (2023). Checked 14 October 2026.
4. ETA. القانون رقم 149 لسنة 2026 بتعديل قانون الضريبة على القيمة المضافة، الجريدة الرسمية (in Arabic) (2026-07-28). Checked 14 October 2026.
5. NeurIPS 33. Lewis, P. et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020). Checked 14 October 2026.
6. Google AI for Developers. Grounding with Google Search (Gemini API) (2026-10-13). Checked 14 October 2026.
7. Banque Misr. تمويل المشروعات الصغيرة (in Arabic). Checked 14 October 2026.
8. OECD. Generative AI and the SME Workforce: New Survey Evidence (2025-11-04). Checked 14 October 2026.