[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"blog-doc-en-intelligenza-artificiale-in-azienda":3,"site-version":303},{"id":4,"title":5,"body":6,"category":262,"credit":263,"date":266,"description":267,"extension":268,"faq":269,"image":282,"imageAlt":283,"keyword":284,"meta":285,"navigation":286,"path":287,"related":288,"seo":292,"seoTitle":293,"service":294,"stem":295,"tags":296,"__hash__":302},"blog\u002Fen\u002Fblog\u002Fintelligenza-artificiale-in-azienda.md","AI for businesses and professionals: use cases that hold up and projects to avoid",{"type":7,"value":8,"toc":242},"minimark",[9,13,18,23,26,30,33,37,40,44,47,51,54,58,93,97,136,140,143,147,150,154,157,195,198,202,205,208,212,215,218,221,225,228],[10,11,12],"p",{},"Language models are powerful tools with a precise profile: they excel at reading, classifying, extracting and rewording text, and they are unreliable when asked to remember facts, do arithmetic or make decisions without context. Projects that work start from this profile. Projects that fail start from the technology and look for a problem to solve.",[14,15,17],"h2",{"id":16},"use-cases-that-hold-up","Use cases that hold up",[19,20,22],"h3",{"id":21},"assistants-grounded-in-a-knowledge-base","Assistants grounded in a knowledge base",[10,24,25],{},"They answer repetitive questions from customers or colleagues using only verified content: terms, procedures, manuals. They work when the knowledge base is curated and the assistant can say \"I don't know\".",[19,27,29],{"id":28},"data-extraction-from-documents","Data extraction from documents",[10,31,32],{},"Supplier invoices, orders by email, hand-filled forms, contracts: the model extracts the fields into a structured, constrained format, the system validates them, and a person checks only the doubtful cases. It is often the project with the fastest return.",[19,34,36],{"id":35},"triage-and-priority","Triage and priority",[10,38,39],{},"Incoming requests classified by type, urgency and recipient, with the key data already extracted. The operator does not read everything: they work from an already sorted queue.",[19,41,43],{"id":42},"drafts-to-review","Drafts to review",[10,45,46],{},"Replies to frequent requests, product descriptions, meeting summaries, first drafts of documents from structured data. The person remains the author; the model removes the blank page.",[19,48,50],{"id":49},"semantic-search","Semantic search",[10,52,53],{},"Finding documents, cases or products similar in meaning, not just by keyword, in internal archives or the site's catalogue.",[14,55,57],{"id":56},"projects-to-avoid","Projects to avoid",[59,60,61,69,75,81,87],"ul",{},[62,63,64,68],"li",{},[65,66,67],"strong",{},"Starting from the model",": \"let's use AI for something\" produces prototypes nobody uses.",[62,70,71,74],{},[65,72,73],{},"Automating decisions without oversight",": quotes, refunds, sensitive messages sent without human review.",[62,76,77,80],{},[65,78,79],{},"Using the model as a database",": asking it for prices, deadlines or data it should read from a system.",[62,82,83,86],{},[65,84,85],{},"General-purpose chatbots"," on the site, with no knowledge base and no limits.",[62,88,89,92],{},[65,90,91],{},"Projects without measurement",", where success is an impression.",[14,94,96],{"id":95},"how-it-is-designed","How it is designed",[98,99,100,106,112,118,124,130],"ol",{},[62,101,102,105],{},[65,103,104],{},"One process, one metric",": time per case, errors, volume handled.",[62,107,108,111],{},[65,109,110],{},"A set of real examples"," with the correct answer, gathered before writing code.",[62,113,114,117],{},[65,115,116],{},"Structured, validated output",": the model returns data in a schema, the code checks it before using it.",[62,119,120,123],{},[65,121,122],{},"Verified escalation criteria",": failed checks, missing sources or out-of-distribution cases go to a person. A safety score declared by the model is not a calibrated probability.",[62,125,126,129],{},[65,127,128],{},"A model sized to the task",": often a smaller, faster, cheaper model is enough.",[62,131,132,135],{},[65,133,134],{},"Release alongside"," the existing process, comparing results.",[14,137,139],{"id":138},"costs-and-data","Costs and data",[10,141,142],{},"Running cost depends on volume and text length; with short answers, caching and the right model it stays contained, but it must be estimated upfront and monitored. On data: providers whose terms exclude training, minimisation of personal data, limited logs, a GDPR assessment. The EU AI Act adds transparency obligations when a system interacts with people.",[14,144,146],{"id":145},"the-people","The people",[10,148,149],{},"Projects that work involve whoever does the job today: they know the exceptions, they know where the model will get things wrong, and they are the ones who will use it. The aim is not to replace them but to take the mechanical part off their plate.",[14,151,153],{"id":152},"a-typical-example-start-to-finish","A typical example, start to finish",[10,155,156],{},"A practice receives dozens of documents from clients every day (invoices, receipts, forms) and someone transcribes the data into the management system.",[98,158,159,165,171,177,183,189],{},[62,160,161,164],{},[65,162,163],{},"Baseline",": how many documents a day, how long each takes, how many transcription errors.",[62,166,167,170],{},[65,168,169],{},"Test set",": about a hundred real documents, with the correct data filled in by hand.",[62,172,173,176],{},[65,174,175],{},"Extraction",": the model returns the fields in a defined schema; the code checks formats, totals and date consistency.",[62,178,179,182],{},[65,180,181],{},"Threshold",": if evidence is missing, the case is not covered by the sources or a check fails, the document goes to a review queue with the fields already filled in.",[62,184,185,188],{},[65,186,187],{},"Shadow run",": for a few weeks the workflow runs alongside the manual work, and results are compared.",[62,190,191,194],{},[65,192,193],{},"Release",": the person no longer transcribes, they only check the doubtful cases.",[10,196,197],{},"The result is measured with the same indicators as the baseline. If it does not improve, the project stops: that is a useful outcome too.",[14,199,201],{"id":200},"assessing-errors-and-access-to-sources","Assessing errors and access to sources",[10,203,204],{},"For an extractor I measure correctness per field, the number of documents sent to review and the errors that slip past the checks. I keep the examples used to develop the system separate from the evaluation ones. Passing a JSON schema shows that the output has the expected structure, not that amounts, names or dates are true.",[10,206,207],{},"For a document assistant I apply permissions before retrieving the sources: a document the user is not authorised to see must not enter the model's context. I test cases with missing or contradictory sources and document content that tries to alter the assistant's behaviour. Actions with external effects go through deterministic checks and the approval steps the process requires. A model's self-assessment is not enough to decide that a result is reliable.",[14,209,211],{"id":210},"my-take","My take",[10,213,214],{},"I use artificial intelligence every day and integrate it into projects, but with a clear idea: it is a component, not the product. The Altura Vela management app has AI features that take repetitive work off whoever organises courses and outings, inside an application with its own data model and checks. In Fix and Sail's Captain, the assistant answers based on the manuals of the accessories on that boat. The chat on this site is configured to answer from a knowledge file and to flag requests outside its scope; this behaviour must be verified and does not guarantee the absence of errors.",[10,216,217],{},"In all three cases the value is not in the model, which can be swapped, but in what surrounds it: the right data, the limits, the code that checks the answers.",[10,219,220],{},"There is another aspect I consider decisive: reversibility. Models improve and change price every few months, and a project built around a single provider's quirks ages quickly. That is why I always separate the application logic from the model: instructions, output schemas, test set and checks live in the project's code, and the model is a replaceable component. When a better-suited model comes out, the test set is re-run and the decision is made on numbers, not promises. It is the same principle I use for every other dependency: few, replaceable, under control.",[14,222,224],{"id":223},"why-work-with-me","Why work with me",[10,226,227],{},"I have delivered more than 300 projects since 1999, many for agencies, and today AI is a tool of my work, not a label to sell. I integrate language models into websites, management systems and workflows with validated output, measurement and care for data.",[10,229,230,231,236,237,241],{},"If you want to understand where AI could really help your business, ",[232,233,235],"a",{"href":234},"\u002Fen\u002Fcontact","get in touch",": in a call of about 20 minutes we pick a process and a metric. The service is in ",[232,238,240],{"href":239},"\u002Fen\u002Fservices\u002Fai","AI and automation",".",{"title":243,"searchDepth":244,"depth":244,"links":245},"",2,[246,254,255,256,257,258,259,260,261],{"id":16,"depth":244,"text":17,"children":247},[248,250,251,252,253],{"id":21,"depth":249,"text":22},3,{"id":28,"depth":249,"text":29},{"id":35,"depth":249,"text":36},{"id":42,"depth":249,"text":43},{"id":49,"depth":249,"text":50},{"id":56,"depth":244,"text":57},{"id":95,"depth":244,"text":96},{"id":138,"depth":244,"text":139},{"id":145,"depth":244,"text":146},{"id":152,"depth":244,"text":153},{"id":200,"depth":244,"text":201},{"id":210,"depth":244,"text":211},{"id":223,"depth":244,"text":224},"Artificial intelligence",{"name":264,"url":265},"drmakete lab","https:\u002F\u002Funsplash.com\u002F@drmakete","2026-05-21","Where AI brings measurable value to a business and where it does not: grounded assistants, data extraction, triage, drafts. With selection criteria, evaluation, costs and privacy.","md",[270,273,276,279],{"q":271,"a":272},"Where is the best place to start?","With a frequent, measurable task with a low cost of error, where today someone reads text and transcribes or routes it. That is where language models deliver most with least risk.",{"q":274,"a":275},"Does data sent to models stay confidential?","It depends on the provider and the contract. You use APIs whose terms exclude training on submitted data, minimise personal data and assess GDPR compliance before starting.",{"q":277,"a":278},"Do you need a specially trained model?","Almost never. Most cases need a good general model, precise instructions, schema-constrained output and, where specific knowledge is required, a system that supplies the right documents with each request.",{"q":280,"a":281},"How do you know it works?","With a set of real examples and the correct answers, measured before and after every change, and with a process metric: time saved, errors, cases handled.","\u002Fimg\u002Fblog\u002Fintelligenza-artificiale-in-azienda.jpg","Hexagonal tunnel of mirrors reflecting a mountain range","artificial intelligence for business",{},true,"\u002Fen\u002Fblog\u002Fintelligenza-artificiale-in-azienda",[289,290,291],"gestionale-su-misura-o-excel","brief-per-sviluppatore","da-google-analytics-a-cloudflare-web-analytics",{"title":5,"description":267},"AI for businesses: real use cases and what to avoid","ai","en\u002Fblog\u002Fintelligenza-artificiale-in-azienda",[297,298,299,300,301],"artificial intelligence","ai for business","automation","ai assistant","llm","e_HDMTyBFYIY5HUtTy2zOkZUY2L7-rS0i3Pq3QqZH2Y","2.1.0"]