Algeria AgencyWeb & software
Artificial intelligence: what it does well, what it invents, and what you cannot hand it
Its failure mode is not breakage, it is confident error. It produces text just good enough that nobody checks it.
It has never been easier to sell artificial intelligence to a business, and never harder to get an honest description of what it does. This article attempts the second.
These tools are genuinely useful, on a specific list of tasks shorter than the sales material suggests. They are also dangerous in a particular way: they do not break down, they are wrong with confidence, in impeccable prose.
That is what separates them from every other tool described in this series. Software that fails tells you; this produces a plausible, well-built, incorrect answer, and it takes somebody who knows the subject to notice.
You will find no figures in this article, and that is deliberate. We explain why in the section on measurement — it is precisely the subject where quoting an undated one would be least honest.
What it genuinely does well
The list is real and shorter than the sales presentations. These tools are excellent at rewording text you have already written, at shortening it, at changing its register, and at translating into a language you know well enough to review.
They are very good at the first version of a repetitive text: a product description from specifications you supply, a standard answer to a frequent question, a chasing message. The important word is "first" — the version delivered is rarely the one they produce.
They are useful for sorting and extracting: classifying two hundred incoming messages by subject, spotting the orders containing a particular mention, pulling dates out of a text. Those are tasks where an error is visible and recoverable.
They are good at summarising a document you have in front of you, which is very different from answering a question about a document they do not have. That distinction is the most useful in the whole article: given the text, they are reliable; deprived of it, they fill the gap.
And they are excellent at getting you started. A blank page is a real cost, and a mediocre draft to correct beats an empty document — provided the correction actually happens.
The failure mode: it is wrong with confidence
An ordinary tool that fails shows an error. This one produces a complete, well-phrased answer in the expected register, and sometimes an entirely invented one.
That applies first to checkable facts: a figure, a date, a reference to a text, a name, a technical specification. Those are exactly the elements a hurried reader does not verify, because they are presented with the same assurance as everything else.
The effect worsens when the question is precise and local. Asked about an administrative obligation, a regulatory deadline or a practice specific to one market, the tool will produce something shaped like an answer, assembled from whatever most resembles it in what it has seen.
So the real risk is not the large error, which gets spotted. It is the plausible error inside an otherwise correct text, published because the rest was good, and discovered by a customer six months later.
The practical consequence is not to abandon the tool. It is never to hand it a task whose result you could not verify — which is a selection criterion, and the only one that matters.
The reviewer rule
One rule protects against everything above, and it has to be written rather than implied: nothing these tools produce goes to a customer, a supplier or an administration without being reviewed by a named person.
The word "named" does the work. A rule saying "it must be reviewed" is applied by nobody once volume rises; a rule saying who reviews what survives the load.
Review is not copy-editing. It covers three things only: checkable facts, commitments made in the company’s name, and what the text implicitly promises. The rest can stay imperfect.
You also have to accept the economic corollary, which is unwelcome: the time saved is real and smaller than advertised, because review consumes part of it. A tool that writes in thirty seconds what takes ten minutes to verify is still useful, and it does not save ten minutes.
Where the gain is greatest is precisely on tasks that are quick to verify: sorting, extracting, rewording a text you know. Where it is illusory is on producing content nobody in the company is able to check.
Language changes everything
This is the most important point in the article for an Algerian business, and the one no presentation mentions: the quality of these tools is not the same in every language.
In French and English the level is high and consistent. In standard Arabic it is adequate but noticeably less safe, with stiff constructions and agreement errors a speaker spots immediately.
In Algerian dialect it is frankly poor. The tools produce something that resembles Algerian Arabic to somebody who does not speak it, and rings false to everybody who does. Publishing that does more damage than publishing nothing.
The operational consequence is clear. These tools are usable for producing French to be reviewed, acceptable for roughing out standard Arabic that an Arabic speaker will seriously rework, and to be avoided for anything that has to sound Algerian.
There is a useful exception in the other direction: they are good at understanding. Summarising customer messages written in dialect, sorting them by subject, extracting a request — that is reading rather than writing, and errors there are visible.
What you must not put into them
Everything you type into an online service leaves your company. It is a banal sentence and it is systematically forgotten the moment the tool is convenient.
The list of what must not go in is short and not negotiable: your customers’ personal data, documents covered by professional confidentiality, health information, a third party’s contractual material, and anything identifying a person in a situation.
The commonest case is also the quietest: pasting a table of orders to ask for an analysis. That table contains names, addresses and numbers, and it has just left your company without any decision having been taken.
The fix is not to ban the tool but to write the rule and give the alternative: anonymise before pasting, or work on an extract with no identifiers. It takes a minute and it turns a potential incident into a habit.
Two questions for any supplier installing a tool of this kind: where is the data processed, and is it used to improve the service. The second answer has to be in writing, because it decides what becomes of what you typed.
Automated customer service: where it helps and where it enrages
This is the most-sold and worst-calibrated use. An automated assistant genuinely helps when it answers questions whose answer is written somewhere and stable: opening times, address, prices, what to bring, how a procedure runs.
It becomes an obstacle the moment it stands between a customer and a person on a subject requiring a decision. A customer with a problem who is answered beside the point three times does not conclude that the technology is improvable: they conclude that you do not want to talk to them.
The rule that makes the tool bearable is simple and rarely applied: a visible way out from the first message. A means of reaching somebody, written down, without having to ask twice.
The second rule is forbidding it to invent. An assistant answering from your own texts is useful; the same assistant allowed to answer in general will become the source of a commitment you never made.
And you have to decide what it does when it does not know. "I do not know, here is how to reach somebody" is an acceptable and professional answer. An approximate answer is not, and it is the default behaviour of most of these tools.
Your customers are already searching for you through one
Some of the questions that once led to a search engine now go through an assistant, which answers directly instead of offering links.
That changes what matters for being found. It is no longer only appearing in a list, it is being the source a system cites when it composes an answer — a subject distinct enough that we have given it a whole article elsewhere in this blog.
The useful point here is narrower and concerns how the site is built: these systems pick up what is written as text, structured and checkable. They do not pick up what is locked inside an image, nor what is never explicitly stated anywhere.
That reinforces, for a new reason, advice this series has been giving for twenty articles: write opening times, prices, specifications and limits as text rather than as images, and date them.
And it adds a monitoring duty: what these systems say about your business is not controlled by you. The minimum is to ask, yourself, the three questions your customers ask, and look at what comes out.
What can be measured, and why this article quotes no figures
You have read this far without meeting a single statistic, on the subject where they are quoted most. That is a choice, and it deserves explaining where it is most visible.
The figures circulating on this subject belong to three families and none is citable. Surveys published by the vendors of these tools, who sell the result they are measuring. Projections, which describe a future rather than a fact. And global productivity-gain averages, aggregated across occupations with nothing in common.
We know of no dated, checkable series on the adoption or effect of these tools in Algeria. Publishing a foreign average and presenting it as applicable here would be exactly what this article criticises the tool for: a plausible, unverifiable assertion.
What can be measured, by contrast, is measured at your premises and without tools. Time spent on a task before, time spent after including review, and the number of corrections needed across ten outputs.
Those three figures, recorded over two weeks, answer the only question that concerns you: does this save time on that task, in your business. No global average can answer that one.
The real cost
The advertised price of these tools is low, and that is what makes their real cost hard to see.
The first invisible line is the review described above. It is the time of a competent person — the most expensive time in your company — and it is proportional to the volume produced.
The second is integration. A tool used by hand in a window costs its subscription; the same tool wired into your data, your orders or your mail becomes a development project, with the ownership and maintenance questions described in the other articles in this series.
The third is dependency. A process built around an external service inherits its price changes, its evolutions and its outages. That is not a reason to do without it, it is a reason to know what happens if the service becomes three times more expensive.
The resulting decision rule is cautious and sufficient: start with manual use, on one precise task, integrating nothing. If the gain is real after two weeks of measurement, then and only then discuss an integration.
What a small business can do right now
There is an honest answer to "where do we start", and it goes through no supplier at all.
Take the most repetitive writing task of your week — a standard reply, a product description, a report — have it done, then review and correct. Record the total time, correction included.
Do that ten times over two weeks. You will then know, for your business rather than for an average, whether the gain exists and whether it is ten per cent or half.
Then take the most repetitive reading task — sorting incoming messages, extracting information from a batch of documents — and do the same. That is usually where the gain is clearest, and it is the use least talked about.
What not to start with: automating a reply to a customer. It is the most visible and most-sold use, and the one whose failure happens in front of the person you were trying to serve.
What to check before signing
The first question is about verification: ask who reviews, and what becomes of an unreviewed output. A supplier with no answer has built a system whose errors you will carry.
The second is about data: where is it processed, is it used to improve the service, and how long is it kept. All three answers have to be written, and the second is the one that counts.
The third is about the source: does the assistant answer only from your documents, or can it answer generally? The second option will one day produce a commitment you did not make.
The fourth is about the way out: how does a customer reach a person, and is that visible from the first message? If the answer is "after three exchanges", you are building an obstacle.
The fifth is a simple test: put to the demonstration a precise question about your trade whose exact answer you know. You will see immediately whether the tool says it does not know or improvises, and that is the only thing worth knowing.
What we do, and what we will refuse to do
What we will refuse: delivering an assistant allowed to answer outside your documents. It will look cleverer in the demonstration and will eventually make a commitment on your behalf, which you will discover either by honouring it or by denying it.
We will also refuse to produce content for publication in Algerian dialect with these tools. The quality is not there, it is audible, and text that rings false in your customers’ language costs more than no text.
And we will refuse to sell you an integration project before you have measured a gain by hand over two weeks. It is a sale we lose regularly, and we lose it willingly rather than wire a tool into your data for a benefit nobody has verified.
A limit of competence, stated plainly: we do not know what these tools will do in two years, and we are wary of anybody who claims to. We work on what they do today, on tasks whose results you can check.
What we do: the reviewer rule written with a name, the list of what does not leave the company, an assistant that answers from your texts and says when it does not know, a visible way out from the first message, and a before-and-after measurement on your own task.
And what you should do without us this week: take your most repetitive writing task, have it done ten times, correct it, and time the total including correction. That figure is yours, it is free, and it is worth more than any statistic we could have put in this article.
Frequently asked questions
Can these tools write our content?
They write good first versions of repetitive texts. They do not produce the published version: their failure mode is confident error, and it takes a named person to review the facts, the commitments and what the text implicitly promises.
Can we use them in Arabic and in dialect?
In standard Arabic, to rough out text an Arabic speaker will seriously rework. In dialect, for understanding — summarising, sorting, extracting — but not for publishing: the result rings false to everybody who speaks it.
What must never be typed into them?
Your customers’ personal data, documents covered by confidentiality, health information, and anything identifying a person in a situation. The commonest case is pasting a table of orders to ask for an analysis.
Should we put an automated assistant on our site?
Only if it answers from your own texts, says it does not know rather than improvising, and shows how to reach a person from the first message. Without those three it produces anger, not service.
How much time does this actually save?
We quote no figure: the available statistics come from vendors, from projections, or from global averages aggregating unrelated occupations. Measure your own across ten outputs over two weeks, review included.
Where should we start?
With manual use on a single repetitive task, integrating nothing, for two weeks. And not with automating a customer reply, which is the most-sold use and the one whose failure happens in front of the person you meant to serve.
Where we come in
Timing two tasks, one of writing and one of reading, produces the only figure worth having here. It does not say which of the two is worth pursuing.
- We compare your two measurements and keep one, never both.
- We set down in writing what is not allowed to leave your walls.
- We name the reviewer, because output nobody checks ends up published.
Nobody here knows what these tools will be capable of tomorrow, and we will build nothing that assumes otherwise.
Read next
Training a model: the three cases where it is worth it
Training changes the shape of answers, almost never their content. What you must supply, and the three cases where it is the right tool.The impact assessment, before connecting an assistant to your customer data
It is done before go-live, and that alone is what distinguishes it from a report. Afterwards it describes a risk that has already either materialised or not.The AI your staff already use: what leaves the business, and the one page of rules that is enough
You bought nothing and it is already in service, on personal accounts. The problem is not the tool — it is that nobody knows what goes through it.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.