A technician is at a customer site, in front of a leaking hydraulic cylinder. The job needs the nominal pressure and the part number of the seal kit. The manual is 400 pages long, it sits on the office server, and the printed copy is in a colleague's van. So the technician calls the office. At the office, the person who knows this product range is on leave.
At the same moment, a salesperson is looking for what was delivered to this customer in 2023, to offer the right replacement part. The information is somewhere: in an email, on a delivery note, in someone's memory. It exists, but it cannot be found quickly. This is one of the problems we address in our AI and automation for industry offer.
Why this search time is expensive
The problem is not specific to industry, and it has been measured for a long time. The McKinsey Global Institute (2012) estimated that office workers spend close to 20% of their week looking for internal information or for the right person to ask. IDC (2007) put information search at 9.5 hours a week, 3.5 of them without finding what people were looking for.
These figures have limits worth stating: they are old, they cover office work in general, and they do not come from Tunisia. To our knowledge there is no recent measurement of time spent per task in Tunisian industrial companies. So we read them as an order of magnitude, not as a promise.
In a company that sells, installs and services equipment, the cost shows up somewhere other than a timesheet: a quote that goes out late, an intervention that takes longer than planned, a wrong part number, and a dependency on two or three people who "know where it is".
What an assistant connected to your documents can do
The technique is often called RAG (retrieval-augmented generation). The idea is simple: the assistant does not answer from memory. It first searches your documents, then writes its answer from what it found. We covered it in detail in our article on RAG for businesses.
In practice, for an industrial company:
- Gather the sources. Supplier catalogues, manuals, instructions, datasheets, internal procedures and, if you wish, the history of deliveries or interventions.
- Make the documents readable. Native PDFs are read directly. Scanned documents go through a text recognition step, which we test on a sample before going further. Tables and diagrams need particular care, and this is where you quickly see whether a document is usable. Our article on document AI explains what current models read well, and less well.
- Ask the question in plain language. "What is the nominal pressure of the series X cylinder?", "Which seal kit for this model?", "What did we deliver to this customer in 2023?". In French, Arabic or English, whatever the language of the source document.
- Get a sourced answer. Every answer shows the document and page it comes from. The technician or salesperson can check it in one click.
- Use the tool where people already work. A web page, a phone app, or a messaging channel your teams already use. The right channel depends on your teams, not on the technology.
The same foundation serves other problems in this series: finding a part number and its equivalents to prepare a quote, or comparing tender specifications with your product sheets for a TUNEPS tender.
What the assistant does not do
- It does not replace human checking. An assistant can be wrong, especially on a badly scanned table. That is why it cites its source, and why a person always validates what goes to a customer.
- It only knows what your documents contain. If a piece of information is written nowhere, the assistant should say so instead of inventing. We set that rule at design time and test it.
- It does not tidy your documents for you. Two contradictory versions of the same manual will give contradictory answers. The project often starts by deciding which version is authoritative.
- It does not replace your ERP. For customer history, we connect to what you already have, through its API or its exports. We explain why in the ERP trap.
Data also matters. Depending on the solution chosen, documents can be processed by a model hosted abroad, or by a model hosted locally or with a provider of your choice. Before we start, we tell you where the data is hosted and which AI models are used.
How to start small
There is no need to digitise all of the company's documentation to find out whether the tool is useful. A small first scope is enough:
- One set of documents: for example the catalogue and manuals of a product range that generates many questions.
- One team: after-sales service, or the salespeople of one region.
- A list of real questions: the ones your teams ask every week, with the known right answer. It is used to test the assistant before it goes live.
- One simple indicator, set at the start: for example the number of questions escalated to the office or to the range expert, or the time it takes to find a piece of information.
If the tool brings something on that scope, it is extended to other ranges and other teams. If not, you know quickly, at small cost.
If your technicians call the office for information that is already written down somewhere, let's talk.
Do you have this problem? Describe it in a few lines: what happens today, who is involved, what it costs you in time. We get back to you quickly, usually within 48 working hours, for a short discussion: to understand your project and see whether we are the right partner.
Describe your problem · WhatsApp · All our solutions for industry
Sources
- McKinsey Global Institute (2012) and IDC (2007), time spent searching for information at work: summary and origin of the figures