The process, the technical decisions and the numbers. No percentage claims we cannot show you the source of.
Most of these projects were sized for an SME. They start from files exported out of the systems that already exist — purchasing, stock, invoicing, an inbox — and answer one narrow question before anyone mentions a platform. The technology grows after the value is demonstrated, not before. That is why almost none of these cases began by replacing software.
Hotels, restaurants and event venues lose business in the same two places: the enquiry that takes days to answer, and the customer asking a simple question out of hours — opening times, allergens, availability. We work on winning and qualifying MICE enquiries, on room visualisation by layout, on agents that accompany the guest before, during and after the stay, and on automating the answers to the questions that repeat every single day — always on the channels the customer already uses, without asking them to install anything.
A group with 8 properties across 3 regions replaced its email quote request with a seven-step journey that returns a proposal to the client and a qualified lead to the sales team. 135 event spaces, 269 catering items, six back offices.
Delivered project. Catalogue figures, September 2026.
Read the caseFrom booking to after checkout, over email: welcome, trip preparation, pre-check-in in a portal, upselling that grows out of what the guest just said, and free-text feedback turned into themes. The booking starts the conversation, not the guest.
Architecture description, not a delivered project.
Read the caseHow an event venue wins more enquiries: calculated floor plans of every room in every layout at no cost per view, data AI assistants can cite, and automatic triage of the enquiries that arrive.
Architecture description, not a delivered project.
Read the caseA traditional Lisbon pastry house of more than eighty years answers questions about hours, products and allergens through conversation — and takes orders — at any hour, with a rule that says «I cannot confirm» rather than guessing.
Built and ready, not yet in production.
Read the caseA proof of concept for a hotel group’s own gift voucher system: 30 vouchers across 6 categories, 5 languages, a purchase microsite and a management back office — with the payment simulated on purpose, and the corporate incentives channel identified as the next step.
Proof of concept delivered. Catalogue figures, March 2026.
Read the caseMunicipalities and parish councils receive more communications each day than they can process, and most of the administrative time goes into interpreting and routing rather than resolving. We apply automatic reading, classification by department, geolocation of the incident and drafted replies, with human supervision matched to the risk of each type of case.
An architecture for councils and parish authorities: an agent that reads the complaint, summarises it, extracts the location, classifies the responsible service, geolocates the incident and drafts a reply — with three levels of human supervision, chosen by risk.
Architecture description, not a delivered project.
Read the caseIn healthcare the cost of a stockout is not financial, it is clinical. We audit the data that already exists — consumption, invoicing, expiry dates, suppliers — and forecast needs from it. We do not touch clinical decisions: diagnosis, treatment and prescription validation always belong to qualified professionals.
Purchases, invoicing and consumption combined with seasonality and weather to anticipate demand in clinics and pharmacies — without personal clinical data, because it adds nothing to a stock forecast.
Delivered project. No figures published.
Read the caseIn insurance the hard part is not the calculation, it is the explanation. We build assistants that run a simulation, ask the questions in the right order and return a proposal the customer actually understands — with underwriting, advice and risk decisions staying human.
An insurer replaced its «Call Me» button with an assistant that talks, gathers what the quote needs and returns a result in the same interaction — at 11pm or on a Sunday, without waiting for office hours.
Delivered project. No conversion figures published.
Read the caseIn real estate, the speed of the first reply decides who keeps the client. We work on assistants that read the enquiry as it arrives by email, match it against available stock and draft the response — leaving the relationship and the viewing with the agent, which is where deals close.
A property development that answers by itself over email, with AWS Lambda and an LLM using the project's real knowledge — and the separation between instructions, knowledge and email content that stops a message from becoming a rule.
Delivered project. No conversion figures published.
Read the caseOn a building site, buying late stops a crew and buying early ties up cash, fills the warehouse and damages material. We combine the bill of quantities, the programme, real progress, supplier lead times and weather to forecast what will run short — by site and by week, not in general.
Forecasting what each site needs and in which week, combining the plan, actual progress, supplier lead times and the weather. Buy early only what really needs buying early.
Architecture description, not a delivered project.
Read the caseSpecialist retailers now compete with international operators who have turned the individual product into a commodity. The way out is rarely selling cheaper: it is to stop selling pieces and start selling decisions — professional judgement, integrating brands into one coherent solution, and taking responsibility for the result in the client’s actual space. We work on positioning, on the role of physical stores, and on a digital architecture legible to both people and AI systems.
A high-end furniture and interior design SME in Lisbon stopped competing on product and started competing on decisions: new positioning, a new role for the stores, five hero products and an AI-first architecture.
Delivered project. No conversion figures published.
Read the caseResearch projects spend most of their time on preparatory work: locating and appraising literature, organising data, structuring databases, handling geographic and photographic records. That is where AI speeds things up. Scientific method, field observation and expert validation are not delegated — and that is exactly why the boundary has to be written down.
An internal project on biodiversity in a major Portuguese estuary: literature review, comparison with estuaries at other latitudes, a species database and preparation for field observation — with a boat for the part you cannot do at a computer.
Internal applied research project.
Read the caseAn AI agent interprets untrusted language and produces probabilistic results. Treating it like a traditional application is the mistake that opens the door. We work on threat modelling before the first prompt, explicit trust boundaries, authorised retrieval, minimal tools and human confirmation proportional to impact. It applies across every sector above.
A cross-cutting case with the patterns applied in insurance, healthcare, public administration, hospitality and retail: five boundaries instead of one giant prompt, schema-validated output, RAG with access control, and enough logging to investigate an error.
Cross-cutting case. Verifiable architecture, no incident metrics.
Read the caseA newsroom produces the same story types every day — interest rates, inflation, fuel prices, results, markets. Individually they cost minutes; repeated several times a day they cost hours of editorial capacity. We automate what repeats, from approved sources with deterministic validation, with RAG retrieving the group’s own archive and the newsroom deciding what gets published.
A platform that monitors sources with a known cadence, validates the data, retrieves archive context through RAG and prepares recurring stories — with human oversight matched to risk, SEO and AEO, and the rule that monitoring everything is not publishing everything.
Proprietary technology at mature proof-of-concept stage. Implementation in 3 to 9 months.
Read the caseMore cases in preparation. We publish a project when there is technical substance worth explaining — not with every delivery.
BigLearn is a Portuguese artificial intelligence consultancy, founded in 2017 and based in Lisbon. We work in hospitality and tourism, HORECA, local government and public administration, insurance, real estate, healthcare, construction, manufacturing and professional services — with organisations that already have working processes and want to know where AI improves them.
The cases published here cover hospitality and tourism, HORECA, local government, healthcare, insurance, real estate, construction, retail and interiors, applied research, cybersecurity and media/editorial. The others do not — and until we have a project worth explaining, they will not. A sector listed without a case is a sector we work in, not one where we have public proof. It is the same distinction we draw between a delivered project and an architecture description.
What the cases share is not the technology, it is the shape of the problem: someone doing by hand what is already written down somewhere, and a wait between the intent of the person searching and the first useful answer. If that sounds familiar, the route starts with a proof of concept, with a written success criterion agreed before we begin.
Clients are not named. We describe the sector, the size and the problem with enough precision for the work to be understood, and we stop there. Naming a client requires written permission, and we would rather publish without a name than wait for it.
The numbers are catalogue figures, not results. We say how many rooms, how many items, how many languages — verifiable things, with a date attached. We do not publish revenue-increase or cost-reduction percentages we cannot show the source of.
We tell you the decisions, including the ones that ruled out AI. A case study where everything went well and the technology solved everything teaches the reader nothing. The interesting choices are the ones that had an alternative.
We say when it is architecture rather than delivery. Some of these describe completed projects; others describe what one of our architectures can do, written in the conditional. They are different things and they are labelled as such — in the list above and at the head of each page. A catalogue of capabilities presented as a track record is the fastest way to lose the reader.
BigLearn is a Portuguese artificial intelligence consultancy, founded in 2017 and based in Lisbon. It works in hospitality and tourism, local government and public administration, insurance, healthcare, manufacturing and professional services, on AI agents, business process automation and digital transformation.
Three: hospitality, with an event configurator for a group of 8 properties; local government, with a complaint-handling agent for councils; and insurance, with a life insurance quoting assistant. We work in other sectors, but a sector listed without a case is one we work in, not one where we have public proof.
Naming a client requires written permission, and we would rather publish without a name than wait for it. We describe the sector, the size and the problem with enough precision for the work to be understood, and we stop there.
No. We publish only verifiable, dated figures — how many rooms, how many items, how many languages, how many steps. We do not publish revenue or cost percentages we cannot show the source of.
A delivered project reports completed work for a real client, with verifiable catalogue figures. An architecture description is written in the conditional and says what one of our solutions can do, with no concrete deployment behind it. Each case says which it is, in the list and at the head of its own page.
With a proof of concept and a written success criterion agreed before we begin. We look at what people do today, what information is already written down and what the bottleneck costs, before discussing technology. If the proof does not pass, it was cheap to find out.