AI in an organisation that isn't ready for it
On AI applications that stay and AI applications that are switched off again after three months, with the criteria to tell the difference beforehand.
5 articles
What AI really does today in the work of an agency and of our clients. About automation, agents and where the limits lie.
On AI applications that stay and AI applications that are switched off again after three months, with the criteria to tell the difference beforehand.
In recent years, artificial intelligence (AI) has developed at breakneck speed and found its way into various sectors, including web development.
One example from the documentary illustrated this well: a young man, freshly graduated, was not a fan of video job applications.
At Mediabirds we keep moving with what is happening in AI, so that our clients get the most out of it.
One of the most important features of ChatGPT is its capacity to understand context.
The useful part of AI is fairly boring at the moment: work that occurs often, has a clear input and whose result someone still checks. Preparing quote texts, sorting incoming email, putting product texts into several languages, writing a first version of a report. That genuinely saves hours. The spectacular part, a system that takes decisions on its own, gets stuck on the same point: you cannot check what you cannot see.
Where it goes wrong is almost never the model. It goes wrong on the data around it: an assistant that cannot reach the right source gets its answer from its general knowledge and sounds just as convinced doing so. That is why we always build AI on a defined source, have that source named in the answer, and make sure it is visible when it does not know something.
On the visibility side something is changing at the same time. A growing share of your potential customers put their question to an assistant instead of a search engine, and get one answer with a few sources instead of ten links. You get into that with content that states a question explicitly and answers it briefly, with structured data around it and with facts that can be checked.
This category holds the articles in which we work that out, including about our own site. The assistant in the header and the search function in this archive were built on these principles.
The questions we get asked most about this. Yours not among them? Ask our assistant or just give us a call.
Work that repeats itself, has a clear input and where a person can still check the result. Think of classifying and forwarding incoming enquiries, translating or reusing texts, summarising documents, and making a first version of something that someone then finishes. What does not yet work reliably is work where a mistake only shows up later and nobody is watching in between.
By publishing content that states a question word for word and answers it directly, and by making that content machine-readable with structured data. An assistant prefers to cite sources that can be checked: an author with a name and a profile, a date, figures with their origin. What does not help is repeating keywords. These systems do not steer on that.
That depends on the setup, and hardly on the model. The questions to ask: does your data go to a party that is allowed to keep it, is it used for training, and where are the servers. With the business versions of the major providers, training on your input is switched off by default and a data processing agreement is available. With a free consumer account that is not the case, and that is where most company data leaks away.
With one task that costs a lot of time, occurs often and carries little risk if it goes wrong once. First measure how long that task takes now, then automate it, and measure again. That gives you a number with which you can defend the next step. Do not start with the task that makes the most impression; that is almost always also the hardest to check.
Not in the form we see now. What you do see is that the executing bottom end is getting cheaper: a first version, a translation, a variant on an ad. As a result the value shifts to the questions that come before it: what do we want to say, to whom, and why would anyone believe it. Teams that only produced get squeezed; teams that choose do not.
Questions about ai?
Reading is one thing. Knowing what it means for you is another.
Tell us what you are running into. In half an hour you hear what we would do, and why we think so.
Free check·30 sec
How visible is your website?
Over twenty checks on SEO, technology and whether an AI model can quote you.