AI for an Iranian business

Do not start from the tool, start from the task

Most failed AI projects in small businesses come from one ordering mistake: the subscription is bought first, and then a task is hunted for that could use it. The right order is the reverse, and it starts with a plain file, not a tool.

For one week write down every repetitive task with a rough time next to it. Then add three columns to the list and measure every row against them.

Repetition. A task done once a month is not worth building a process around, however annoying it is. Weekly repetition is the minimum that pays back the setup cost.

Cost of error. If the output is wrong and nobody notices, what happens? A product description with one extra adjective can be fixed. A number that goes onto a customer invoice cannot. This column decides more than the other two.

Review time. If working out whether the output is correct takes half an hour by itself, you have not handed the task over, you have only moved it. The practical bar: you can be sure in under two minutes.

A row that meets all three is a handover candidate. A row that misses one waits. A row whose cost of error is high never becomes a candidate, however much better the models get. That split is not about model capability; it is about what a wrong output costs you.

One experience repeats everywhere: the model makes the first half of the work lighter and the second half heavier. The draft is ready in minutes, but an almost right output takes longer than an obviously wrong one, because you have to read every sentence to find which one is off. Leave that time out of the calculation and the saving stays on paper.

Strengths, weaknesses, opportunities and threats, without the flattery

The figure at the top of this page has four cells and its axes are real: internal against external, helpful against harmful. We open the cells here, because a cell with three words in it is no use for a decision.

The strengths are internal and mostly linguistic. Writing, summarising, translating, answering a customer message, tidying meeting notes. A small Iranian team has one more strength that large companies lack: the distance between decision and execution is short, and no five people have to sign for a process to change.

The weaknesses are internal too and have nothing to do with the model. Payment friction is real. Persian quality still trails English, particularly in specialist text and formal register. And more important than both: in most small teams nobody owns the review of the output, and that single weakness kills more projects than model quality does.

The opportunities are external and sit exactly where the work was not worth doing yesterday. Translating a catalogue, writing descriptions for hundreds of items, first replies to repetitive messages, summarising internal documents for a new team member. The second opportunity is elsewhere: this work can be sold to clients, and handing repetitive office work over is one of those services.

The threats are external and two of them are serious. First, dependency: a process riding on an account that is not yours can stop working one morning, and that morning a client is waiting for an answer. Second, the absence of advantage: anything the model does easily, your competitor can also do; so more speed buys no durable advantage, it only raises the floor of what the market expects. The third threat gets its own paragraph in the next section, because it comes for websites more than for anything else.

The position that falls out of these four cells: for a small Iranian business AI is not a competitive advantage, it is a new floor. Someone who does not use it falls behind, but someone who does use it does not move ahead. Moving ahead still comes from where it always came from: the thing only you know and the model does not.

Three scenarios that actually come up

An online shop with a few hundred items. The handover task is obvious: draft the product description from the spec sheet. It meets all three conditions. High repetition, correctable errors, and fast review because you only have to check the description against the sheet. The trap sits exactly there: when a specification is missing from the input the model invents it, and a waterproof rating or an eighteen month warranty nobody mentioned ends up on the product page. The fix is structural, not moral: tell the prompt that anything not on the sheet must be left empty, and fill the empty fields yourself.

The third threat comes for the shop right here. Google own guidance says plainly that appropriate use of AI or automation is not against its guidelines, but that generating content with the primary purpose of manipulating search rankings violates its spam policies. The difference between the two is not intent, it is result: a thousand product descriptions that each have something to say are fine; a thousand pages built only to fill a gap are exactly what is being targeted. We unpack the detail in the AI and content SEO lesson.

A freelance translator or writer. The change here is deeper, because the job changes rather than the tool: from writing to editing. Whoever accepts that gets faster; whoever does not and delivers raw output gets the work sent back and spends their reputation on it. The economics show up immediately too: if you bill by the hour, the faster you get the less you earn. Pricing has to attach to the output rather than the time, and that decision cannot wait past the next project. The freelance market understood this shift before the employers did.

An agency, meaning us. This site runs three internal marketplaces, and a person used to review every listing. Now an automatic classifier looks first and a person looks after it. What mattered in that project more than the choice of model was one design decision: the confidence required to reject a listing was deliberately set higher than the confidence required to publish it. The reason is economic, not technical. Holding a healthy listing for an hour costs nothing, while wrongly rejecting one loses the seller for good. Every automated system that acts on users needs an asymmetry like that, and it is almost never the default.

Where the real cost is written down

We do not print subscription prices in this lesson, and that is deliberate: any number placed here will disagree with the real one within three months, and a reader will trust the wrong figure. The price of each plan, with the date it was recorded, sits in this site own AI reference.

But the subscription is not the only cost and for an Iranian business it is not even the largest. Count three items together.

The subscription. The number on the vendor page. The simplest item and the only one everybody looks at.

The payment route. An Iranian card is not accepted, so the real amount includes the fee of whatever route you choose and the risk that route carries. The routes and their differences are collected on the buying and paying page. This item sometimes exceeds the subscription itself and appears on no pricing page anywhere.

Review time. The hours your team spends reading and correcting output are a cost, even if they never reach an invoice. Leave it out and every project looks profitable on paper.

A practical way to decide without risk: before buying any subscription, do the same work for a month on the free plan and measure one thing only, the review time per output. If that number falls and stays down, buying the subscription is an easy decision. If it does not, a subscription will not fix it. One condition on this method cannot be waved away: the free plan is where your data is most likely to be used to improve the product, so that trial month has to run on public or invented data and not on a client file.

Where you should not hand over

There are three boundaries and none of them moves as models improve, because none of them is about model capability.

Where errors are expensive and surface late. A tax figure, a contract amount, a legal claim, medical advice, a promise made to a client. The problem with this group is not how fast the mistake happens, it is how slowly it is discovered: the output looks convincing and nobody notices until somebody chases it. Wherever feedback is slow, handing over is dangerous.

Where client data is involved and you have no contracted account. This one is a ruling, not a preference. The privacy in AI tools lesson carries the per plan policy table for each vendor and a two minute method for redacting a text before you paste it.

Where a signature is involved. The model writes, but responsibility for what was written belongs to whoever sends it. That difference shows up quickly in client work: the client has a contract with you, and if the text is wrong, explaining that a model wrote it solves nothing.

If the whole lesson had to fit in one sentence: hand over the tasks where, if the output is wrong, you find out before the client does. The rest is still your work, and it is probably the part you are paid for.