What We Will Not Let AI Do

Every agency in this market now says it uses AI. Almost none of them will tell you where they stop. That second half is the part that matters, because the tools are the same for everybody — the same models, the same prompts, the same ability to produce a month of content in an afternoon. What separates one agency from the next is no longer what its software can do. It is what it refuses to do with it. So here is our list, and the reason under each line.

Competence became free, and that changed what you are buying

For most of this industry's history you paid an agency for competence. A clean paragraph. A decent ad. A tidy monthly report. Those things took skill and time, and the people who had the skill charged for the time. They do not take either any more. Anyone with a browser can produce a passable landing page, forty ad variations and a quarter of social captions before lunch. The floor rose, and when the floor rises, everything standing on the floor loses its price.

What stayed scarce is narrower and harder. Judgment — knowing which of twenty good options is the right one for this business, this quarter, in this market. Taste — knowing when something is technically correct and still wrong. Accountability — a name on the work when it goes out, and a person who is still there when the numbers come in. Those three are what a client is actually buying now, and not one of them can be generated. Everything below is us protecting them from our own tools.

AI never publishes unread

The rule is plain. Nothing leaves this agency that a human has not read end to end. Not a caption, not a meta description, not a WhatsApp reply template, not a line inside a proposal. We use AI to draft, never to publish unread.

That sounds like an obvious standard and it is broken constantly, because the output is good enough to skim and skimming feels like reading. A model will give you nine accurate sentences and one confident invention, and the invention reads exactly like its neighbours. There is no tell — no clumsy grammar, no hedging, nothing that catches the eye. So the only defence is reading with the assumption that something in there is wrong. That is a different activity to proofreading, it is slower, and it is the part most agencies quietly dropped the month the tools got good.

AI never invents a number, a result, or a testimonial

This is the line we hold hardest, because it is the easiest one to cross by accident. Ask any model for an article about marketing and it will hand you a statistic with a percentage, a year and a plausible-sounding source. Some of those are real. Some are not, and you cannot tell which by looking at them. So we do not use them. If we cannot say where a number came from, the sentence gets written without the number — which usually makes it a better sentence, because a claim you can defend in a meeting is worth more than a decoration.

The same rule governs work and praise. We do not invent case studies, we do not round a result upward, and we never draft a testimonial for a client to approve. What we point at is what we did: a hotel and residences site for Matarma Beach Residence with a booking engine, an AI concierge and a gallery of a hundred and twenty-nine real images; the site and CMS behind G Developments; a lead-generation ad account with forty-two creatives that actually ran; a cinematic film called DUNE. A short true list beats a long impressive one that cannot survive a question.

It extends to pictures. Generative tools are genuinely useful in Video Production for boards, previs and treatment films, and we use them there. We do not use them to show a client's property in a state it is not in, or a finish that was never installed. A generated image of a building that does not look like the building is a lie told in a nicer medium, and the buyer finds out on the site visit.

AI never speaks as your brand without a human approving the words

Automation makes it trivial to let a machine answer your customers, and in Egypt that risk is concentrated in one place. WhatsApp is where the conversation actually happens, and the first thing a serious buyer asks about is the payment plan. A model improvising that answer will improvise something reasonable. Reasonable is not the same as correct, and a wrong number in a WhatsApp reply is not a typo — it is a commitment your customer will hold you to, screenshot and all. Automation should route the conversation, remember the contact, and follow up so nothing leaks. It should hand a human the moment it matters, and everything it says on your behalf should run on words a person wrote first.

Arabic is the second place this line matters. A model will turn your English copy into Arabic that is grammatically clean and completely lifeless, or formal in a register nobody actually speaks. Bilingual content is not a translation job; it is two pieces of writing that happen to make the same promise. That is how we treat it in Content Creation, and it is the difference between copy that reads like a form and copy that reads like your business talking.

AI never makes the strategic call

Ask a model for a strategy and you will get one. It will be structured, sensible and average, because it is assembled from what everyone else has already published. Average is the one thing a business genuinely cannot afford, since the whole point of a position is that it is not the position next door. Deciding what your business should say — what it stands on, what it refuses to compete on, which offer it leads with — is a bet with money behind it. Bets need an owner.

Machines are excellent at the work that feeds the call. Reading competitors, sorting demand signals, summarising a market in a morning that used to take a fortnight — that is Market Research, and it is better with AI than without it. But the sentence at the end of the research, the one that says this is the message and this is the market we go at first, gets written by a person who will be in the room when it is measured.

Is AI content bad for SEO? That is the wrong question

Search engines do not punish text for being machine-written. They demote pages that are thin, duplicated, and add nothing a hundred other pages have not already said. Almost all careless AI content is exactly that, which is why the correlation looks like a penalty and gets talked about as one.

The real risk is not detection. It is sameness. If your page is a fluent summary of the consensus, it competes with every other page assembled from the same consensus, and it gives a reader no reason to trust you over any of them. What ranks and keeps ranking is what was never available to the model: your data, your work, your argument, your market. Our free Growth Scan reads a whole site and scores it out of a hundred across ads, speed, search and content — and the content pages that score badly are almost never badly written. They are interchangeable, which on the web is the more expensive problem.

Verification is a habit, not a policy

None of this holds because it is written in a document. It holds because it is a routine somebody performs every day, on work they are tired of, when the deadline is close. In practice it looks like this:

• Every number is either traced to a source or deleted. There is no third option.
• Every client-facing word is read end to end by a person who can be held to it.
• Every claim about our own results points at work that exists and can be opened in a browser.
• Every generated visual is checked against the real building, room or product it depicts.
• Anything a machine sends on a client's behalf runs on words a human wrote and approved.

That list is not glamorous and it is not fast. It is also most of the job now. The tools took the effort out of producing work and left all of it in deciding what deserves to exist, then proving it is true.

AI made competent work free. Everything you are still paying an agency for sits on the far side of the line it is not allowed to cross.

If an agency cannot tell you what it will not let its tools do, it has not decided yet — and you will find out where its line was the day something goes out with your name on it. Ours is written down, and we are glad to be asked about it.