Our experience: where AI makes the difference.

Each purchase, of around 300 stones, could take up to fifteen hours: download a file, rename it by hand, put it in the right folder, and start again. About three minutes per stone, several hundred times over. This work is now automated from a single Excel file. The time freed up allows the person who handled it full time to devote themselves to sales.

Before

For each purchase, around 300 diamonds had to be handled from five Excel files. For each stone, the 360° video and the laboratory certificate had to be downloaded, renamed according to the internal naming convention, filed across several folder structures and given sharing links. About three minutes per stone, or some fifteen hours of work per batch. The slightest mistake could have disproportionate consequences: a single misnamed file was enough to make a diamond impossible to find.

After

The person in charge simply drops the Excel file onto a web page and starts the run. In three to five minutes, the files are downloaded, renamed and filed in the right places, and the sharing links are generated.

Result

Fifteen hours of work come down to about twenty-five minutes, saving nearly 175 hours a year. The person originally hired for this task has stayed with the company, but now spends their time selling rather than filing.

Role for AI: not needed. This is mainly about automating a sequence of structured steps that follow precise rules.

At a B2B trade show, each order required dozens of manual entries in front of the customer, then a second round of data entry into the ERP at the end of the day. The sales representative can now complete the order form in a few minutes, simply by describing the order to a conversational assistant. Nearly 185 hours of work were eliminated across 248 orders.

Before

An order taken at a B2B trade show could involve dozens of fields to fill in: reference, model, colour, size, quantity, price, discount and total. These fields had to be filled in by hand while other customers waited. After the show, the same data was entered again into the ERP, with the risk of errors and the delays that this double handling entailed.

After

The sales representative can now complete the order form in a few minutes. They can use a form, or describe the customer’s request directly to an AI conversational assistant that knows the product catalogue. The assistant identifies the references, suggests the sizes and colours available and builds the order as the conversation goes on. The VAT number is validated automatically, the address is found and prices are calculated according to the applicable rules. The order is then confirmed to the customer and sent straight into the ERP.

The role of AI

AI makes it possible to understand an order expressed in everyday language and to translate it into product references. It can also identify certain information from the data available. VAT validation, registry searches and price calculation, by contrast, rest on connections to databases and on business rules.

Result

Across 248 orders, about 103 hours were saved at order taking and 83 hours at re-entry, nearly 185 hours in total. The salesperson no longer spends their time filling in a form: they can focus on customers and on selling.

After a trade show, 200 business cards had to be entered, qualified and followed up in several languages. The whole workflow now runs in a single operation: each contact is added to the CRM and receives a message suited to their profile and language. The decisive gain is not only the time saved: it is the certainty that every contact is actually followed up.

Before

A trade show could generate around 200 business cards in three days. Adding them to the CRM already took several hours of data entry, but the real problem came afterwards: qualifying each contact, determining their profile, tailoring the message to their business and writing to them in their own language.

Some contacts were prospective customers, others were looking for a commercial partnership. They expected neither the same information nor the same message. In practice, for lack of time, they often all received a generic follow-up, or none quickly enough, even though the value of a trade show is largely decided in the days that follow.

After

The cards are photographed and handled in a single pass. Each contact is created in the CRM, classified by profile and receives an HTML email (designed in the brand’s colours) suited to their category and written in their language. Handling all 200 contacts takes less than ten minutes.

The role of AI

AI extracts the information on the cards, helps qualify contacts based on their company and job title, identifies the right language and writes the corresponding message. The layout relies on HTML templates to keep the result under control.

Result

Data entry alone took about three minutes per contact. The whole workflow, including qualification and follow-up, now takes less than ten minutes for all 200 contacts. The main benefit lies less in the time saved than in no longer letting the opportunities generated by the show slip away.

Role for AI: particularly relevant for extracting information, qualifying contacts and personalising messages.

Accounting documents arrived through five separate channels and had to be gathered and reconciled by hand. A single journal now brings them together, whatever their origin, and handles them automatically. Nearly 40% of the time spent on this function has been saved.

Before

The gallery received its accounting documents through five separate routes: invoices by email, photos of till receipts, PayPal transactions, Spendesk card exports and expense claim spreadsheets. None of these channels communicated with the others. The person in charge of reconciliation worked two days a week and spent most of that time gathering, entering and checking information.

After

A single journal is now fed through whichever channel best suits each use. A receipt can be sent as a photo via a messaging app (WhatsApp, Telegram), an invoice dropped as a PDF into a conversation, and PayPal transactions are collected automatically. Each document is then read, classified and matched against the bank statement in the gallery’s Drive.

The role of AI and other automation

AI reads the documents and extracts the useful information. Classification rests on 341 supplier rules and a scheme of 52 categories. Bank reconciliation uses a four-pass algorithm over a forty-day window. The system also assesses its own level of confidence and, when it is too low, sends the document for human review rather than handling it automatically without a check.

Result

A time saving of more than 6 hours every week, nearly 40% of the time spent on this function.

The documents were never missing. What was missing was a way to bring them together.

Role for AI and other tools: AI is used mainly to read and interpret the documents. The flow of information, classification and reconciliation rest on structured rules and automation.

Orders and documents moved between several tools that did not communicate, generating double data entry, errors and administrative tasks. The systems were connected without replacing any of them. The time needed to handle an order fell from 1 hour 45 minutes to 10 to 20 minutes, and an additional accounting hire was avoided.

Before

A brand sold both direct, through its e-commerce site, and via an order-taking app for its resellers. These two channels fed three independent systems: the ERP for stock and logistics, the invoicing tool for invoices and VAT, and the CRM for customer follow-up.

Each order was entered at the time of sale, then copied by hand into the different systems. Handling could take up to an hour and forty-five minutes per order.

The same problem existed for supplier invoices. Six people received more than twenty subscription invoices a month, each through different channels. Some documents could not be found until the moment they were requested. The monthly search for invoices in a WhatsApp group took about an hour. Bank reconciliation was then done by hand, line by line, against the general ledger.

After

The different tools were connected without being replaced.

Sales from the e-commerce site and the ordering app now feed the ERP, the invoicing tool and the CRM directly. Invoices are generated without re-entry, in the right currency and with the appropriate VAT treatment, for sales spread across more than thirty countries and eight languages.

Supplier invoices are centralised through a single address. They are read, the key information is extracted, the document is archived and the data is sent to the relevant systems.

Paper documents can be photographed and sent through a messaging app. The system identifies the information, asks for any details it needs, displays a summary and waits for approval before passing the data on to the different tools.

Finally, the bank check can be run from a statement sent to the system. It extracts the lines, matches them against the general ledger and produces a report showing which transactions are already recorded in the invoicing tool and which are not.

The role of AI and other automation

AI reads the invoices, receipts and statements, and interprets corrections expressed in everyday language. Everything else relies on explicit rules: reconciliation, data flow between the tools, VAT logic and reminders.

Result

Handling an order fell from about 1 hour 45 minutes to 10 to 20 minutes, saving nearly 375 hours across 250 orders. The bank check saves about 24 more hours a year.

The growth of the business would have required hiring an additional accounting employee, estimated at €38,400 gross a year. That hire was not needed.

The aim was not to overhaul the information system, but to make information flow between its parts. No tool was replaced, and no new software had to be learnt.

The data already existed: it simply never reached its destination.

Role for AI and other tools: AI is used where documents and everyday language have to be read, understood or interpreted. Integration between the systems, the calculations and the handling rules rest on conventional automation.

To respond to a tender, the teams spent several hours looking for prices they already knew, scattered across old quotes and exchanges with suppliers. The data has been brought together in a single catalogue, with its full price history, accessible directly from the pricing tool. Search time has been halved, and the prices applied are now consistent from one bid to the next.

Before

The company regularly had to respond to tenders with several hundred lines to price. For each line, it had to find the purchase cost of the material and the time needed to install it.

This information already existed within the company, but it was scattered across previous quotes and exchanges with suppliers. The estimator therefore spent a large part of his time looking for data he should already have had at hand. Preparing each tender took about eight hours, at a rate of roughly four a month.

Beyond the time lost, nothing guaranteed that the same product would be priced the same way from one bid to the next.

After

Purchase costs are now gathered in a single catalogue. Every change is recorded automatically: an old price is never overwritten, and the cost of a product on any given date can still be found.

Supplier invoices gradually feed this database. Signed bids also record the selling prices that were accepted, without changing the purchase costs. The catalogue is synchronised with the spreadsheet the estimator uses, so he keeps his usual working environment.

The role of AI and other automation

AI reads supplier invoices and suggests which catalogue product each line belongs to. No data is added automatically without human approval, and the reviewer can compare the suggestion with the original document.

The price history and data synchronisation rely on the database and on automation.

Result

The time spent pricing fell from about eight hours to four per tender, nearly 192 hours a year. Above all, the company now has a common reference for applying the same prices to the same products.

A price you have to go looking for is not data. It is a rumour.

Role for AI and other tools: AI to read documents and suggest matches; a database and automation to centralise the information, keep its history and share it.

Every day, the gardeners described their work in each site’s chat, but turning those messages into a monthly report took one to two hours per client: out of 160 sites, only a few reports went out, and late. The chats are now read automatically, following rules agreed with the gardeners, and each report is laid out with its photos. The office reviews it in a few minutes and sends it in one click: every client receives their report, every month.

Before

The same company works on 160 client sites. Its contract provides for a monthly site report: what was done, on which date, by whom, with photos to back it up.

This information already existed. It lived in each site’s chat channel, where the gardeners describe their visits and post their photos, among the exchanges between colleagues. To produce a report, someone from the office had to reread the month’s channel, sort out what could be shown to the client, export the photos, write a document, insert the photos and rewrite the messages in presentable French. Each report took one to two hours.

Across 160 sites, the task was impossible: only a few reports went out, often late.

After

The app reads the channels of all 160 sites. Rules agreed with the gardeners decide what goes into the report: a message marked (OFF) stays private until an (ON) or until the next day, ‘before’ and ‘after’ photos are paired, a visit report posted late is linked to the right visit in the schedule, and plant replacements stay internal.

On the last day of the month, a report is ready for each site: one paragraph per visit, the list of work carried out, the photos and the before and after shots, laid out as a PDF. The office reviews and corrects it directly in the app, then sends the approved reports in one click. No report goes out without approval.

The role of AI and other automation

AI comes in at two points. First, it tells visit reports apart from ordinary exchanges between colleagues. Then, for each visit, it writes a paragraph in professional French based solely on the gardener’s messages: it states what was done, recommends nothing and invents nothing. An automatic check blocks any recommendation or any paragraph without a source, and the visit then goes back to the office. When the office rewrites a paragraph itself, a second, lighter model corrects the spelling.

Confidentiality does not depend on AI: the (OFF)/(ON) filter applies before it, and no private message is ever passed to it. Linking to the schedule, choosing the photos, the layout, reminders to the gardeners and sending all rest on written rules and automation.

Result

Each report now takes two to five minutes of review and adjustment, instead of one to two hours of writing. Above all, all 160 sites receive a complete, well-presented report every month, which was impossible until now. The company keeps its contractual commitment, and the quality of its work becomes visible to its clients.

Work the client never sees is not a service. It is a promise.

Role for AI and other tools: AI to sort the messages and write the reports; written rules and automation for confidentiality, linking to the schedule, photos, layout and sending.

A catalogue of nearly 5,000 visuals would have been economically impossible to produce through traditional photography: about €150,000 at market rates. An app now generates every variant from a single photo of the product, in 4K, on a white background and with a strictly identical rendering from one visual to the next. The unit cost falls from about €30 to €0.10, but what matters most is making possible a catalogue that was not possible before.

Before

Nine collections, forty cord colours, three metals and several angles: the catalogue to be produced amounted to nearly 5,000 visuals.

There were two paths, both impractical. The first was the photographer. Published rates for product packshots range from €20 to €36 excluding VAT per image depending on volume, plus a set-up fee per session, and studios point out that jewellery shoots, which require specialised lighting and retouching, are charged above that range. In Belgium, advertised rates for jewellery start at €35 per image. At €30 an image, a low estimate, the 5,000 visuals came to about €150,000.

Pricing sources: packshot-paris.fr (€19 per image standard, €29 premium, excluding the set-up fee); atelierb9.com (€36 excl. VAT per image for 1 to 6 products, €20 excl. VAT above 40, plus €150 for set-up, and more for jewellery); studio-360.be (from €35 per image for jewellery, with volume discounts).

The second path was to produce the images by hand with a generative AI tool. That meant about an hour per visual, with instructions to rewrite and adjust at every attempt. It was no longer a question of money, but of time: the catalogue would never have been finished.

After

A single photo taken with a smartphone is enough to start the whole job. The app identifies the product and suggests six framings, and the operator picks the best angle. From that one choice, it generates the twenty cord colours in the three metals with a single button, in 4K, on a white background and in the expected views: three-quarter, close-up and worn.

The difficulty is not producing one beautiful image. It is producing five thousand that look alike. The same framing, the same background, the same light, the same resolution and, above all, a cord colour that stays exactly the same from one product and one collection to the next. This is precisely what a generative model does not do on its own, and it is where most of the work lies.

The background then adapts to where the visual will be used: white for product pages and marketplaces, specific settings for resellers and marketing materials. Patent photographs are produced by the same set-up.

Visuals can be regenerated one by one, or corrected with a simple written instruction. An automatic check verifies each packshot before export, and the generation history makes it possible to go back to an earlier version. The files are then exported with automatic naming.

The role of AI and other technologies

Generative AI produces the images and is the core of the set-up. On its own, however, it cannot guarantee consistency. Cutting out and recolouring the cords rely on programmatic image editing, which is more reliable for this precise operation, and the colour library is managed outside the model so that a colour stays identical whatever the product. An automatic check rejects non-compliant visuals before they enter the catalogue.

Result

The unit cost falls from about €30 to about €0.10, nearly €150,000 of spending avoided across the full catalogue. Production time falls from about an hour per visual to a few minutes for an entire collection.

In reality, the essential result lies elsewhere: it makes possible a catalogue that, until then, could not be considered economically.

The gain therefore lies not in a mere reduction in spending, but in the ability to achieve what could not be achieved before.

Role for AI and other tools: generative AI produces the visuals; programmatic image editing guarantees the reliability of repetitive, precisely defined operations.

Every month, statements from six different sources (BNP Paribas Fortis, Revolut, N26, Stripe, SumUp and Shopify payouts) were compared by hand with the Belgian invoicing tool, line by line, to spot missing invoices. This check took about two hours and was not seen as a problem. It is now done in a few minutes from a statement sent through a messaging app, and missing invoices are requested automatically from the people concerned.

Before

Money came in and went out through six channels: a traditional bank (BNP Paribas Fortis), two neobanks (Revolut, N26), two payment providers (Stripe, SumUp) and the online shop’s payouts (Shopify). Each produces its own statement, in its own format, with its own descriptions and sometimes its own currency. Everything had to be reconciled with the Belgian invoicing tool, which recognises none of these formats.

Every month, the accounts team downloaded the statements, opened them side by side with the invoicing tool and checked line by line whether each transaction had its invoice. Lines without an invoice were highlighted, copied into a list and sent to the people who had made the expense. Everything then had to be checked a second time on receipt, and whatever was missing entered by hand.

Two hours a month, every month. Nobody questioned this way of working: with six sources, checking by hand seemed the only way to do it.

After

The statement is sent as it is, as a PDF, to an assistant in the team’s messaging app. Whatever the format (bank, neobank, payment provider or online shop), the assistant reads the statement, extracts each transaction and compares it with what is recorded in the invoicing tool. Within minutes, it sends back a summary message and a report: the transactions whose invoice is present, those for which it is missing, and the amounts involved. The people behind an expense without an invoice receive the request directly, by email or in the group chat, in a single operation.

The role of AI and other automation

AI reads the statements, including scanned or annotated ones, and extracts the transactions. It then matches each line with the recorded invoices, taking into account differing descriptions, currencies and fees. The flow of information (reading the invoicing tool, producing the report, sending the requests) rests on automation.

Result

A two-hour monthly check comes down to a few minutes, saving about 24 hours a year. Above all, the company now knows, every month, which invoices are missing and whom to ask for them, which it did not know before.

What changed is not the time, but the question: the check used to ask whether the lines matched, not whether invoices were missing.

Role for AI and other tools: AI to read the statements and match the lines; automation to query the invoicing tool, produce the report and send the requests.

Every month, around 70 supporting documents (software and subscription invoices received by email, receipts for everyday expenses) were held by six people and had to be chased one by one, without anyone knowing who had what. This constant chasing was considered a normal part of the bookkeeping work. Now it is enough to forward the invoice to a dedicated address, or to photograph the receipt in the team’s Telegram group: the document is read, archived, recorded in the invoicing tool and, where applicable, attached to the Revolut transaction. Nearly 140 hours of searching and data entry are saved every year.

Before

Six employees incurred expenses for the company with two cards, Revolut and N26. The supporting documents were of two kinds: on the one hand, software and subscription invoices, received by email in each person’s inbox; on the other, everyday expenses, restaurants and routine purchases, whose only proof was a paper receipt. In all, around 70 documents a month, including more than twenty subscription invoices, had to reach the person in charge of the bookkeeping, who recorded them in the invoicing tool and attached the matching document to each Revolut transaction.

The documents arrived by WhatsApp or by email, when they arrived at all. Most often, they had to be chased, without knowing for sure who held which. Finding a document took about five minutes; recording it, five more.

The documents were not lost: someone had them. And since the person holding a document was never the one who needed it, it seemed obvious that the second should go and fetch it. This chasing was tedious, but it seemed inevitable: it was part of the bookkeeping work.

After

Each type of document now follows the path that matches what everyone already does. A software invoice received by email is simply forwarded to a dedicated address, collected automatically, and the assistant confirms the recording by email. A restaurant or everyday purchase receipt is photographed and sent to the Telegram group that brings the whole team together: the assistant recognises the sender, asks which card paid for the expense (two buttons: Revolut or N26), then confirms in the group what it has recorded.

In both cases, the document is read, recorded in the invoicing tool, archived in the Drive and, when the expense was paid with the Revolut card, attached to the matching transaction.

Each document appears in a single tracking table: supplier, amount, currency, card, date, person, source (email or Telegram), period covered for a subscription, and a link to the document in the Drive. Two tabs complete the table: a monthly summary and the list of subscriptions. Opening it is enough to see what has been received, from whom and through which channel.

The role of AI and other automation

AI reads the invoices and receipts, including those photographed on a phone, and extracts the supplier, amount, date, document number and, for a subscription, the period covered. Everything else rests on automation: collecting from the dedicated address, recognising the sender, recording in the invoicing tool, passing documents to Revolut, archiving and keeping the tracking table up to date.

Result

Ten minutes per document, 70 documents a month: nearly twelve hours a month, about 140 hours a year, are saved.

The most important change, however, lies in the order of things: the document is sent at the time of the expense, by the person who made it, instead of being chased afterwards by the person who has to book it. Each document is linked to the person who sent it; the question ‘who has what?’ no longer arises.

A document you have to chase is already a late document.

Role for AI and other tools: AI to read the invoices and receipts; automation to collect what is sent, feed the invoicing tool and Revolut, archive the documents and keep the tracking table up to date.