[{"data":1,"prerenderedAt":759},["ShallowReactive",2],{"navigation-en":3,"uslugi-slug-en-ai-implementation":87,"language-switcher-\u002Fen\u002Fservices\u002Fai-implementation":754},{"services":4,"industries":48,"caseStudies":69},[5,10,14,19,23,27,31,35,39,43],{"slug":6,"navLabel":7,"navGroup":8,"heroHeadline":9},"web-applications","Web applications","build","A web application built around your process",{"slug":11,"navLabel":12,"navGroup":8,"heroHeadline":13},"erp-systems","ERP systems","An ERP system built for your company, not the other way round",{"slug":15,"navLabel":16,"navGroup":17,"heroHeadline":18},"integrations-and-api","Integrations and APIs","improve","Your systems stop being islands",{"slug":20,"navLabel":21,"navGroup":17,"heroHeadline":22},"ai-implementation","AI implementation","AI deployed where it actually saves time",{"slug":24,"navLabel":25,"navGroup":17,"heroHeadline":26},"process-automation","Process automation","Work no one should be doing by hand",{"slug":28,"navLabel":29,"navGroup":8,"heroHeadline":30},"websites","Websites","A site that loads instantly and brings in inquiries",{"slug":32,"navLabel":33,"navGroup":8,"heroHeadline":34},"mobile-applications","Mobile applications","One codebase, both platforms",{"slug":36,"navLabel":37,"navGroup":17,"heroHeadline":38},"code-audit-and-project-takeover","Code audit and takeover","The project stalled. We'll take it over and tell you the truth about its condition",{"slug":40,"navLabel":41,"navGroup":17,"heroHeadline":42},"maintenance-and-support","Maintenance and support","A system needs someone watching it after launch",{"slug":44,"navLabel":45,"navGroup":46,"heroHeadline":47},"discovery-workshop","Discovery workshop","start","Before you decide to build, find out exactly what you're building",[49,53,57,61,65],{"slug":50,"navLabel":51,"heroHeadline":52},"manufacturing","Manufacturing","The works order reaches the floor as a printout from an email",{"slug":54,"navLabel":55,"heroHeadline":56},"automotive","Automotive","The customer exists in three places at once",{"slug":58,"navLabel":59,"heroHeadline":60},"industrial-automation","Industrial automation","You sell competence that doesn't photograph",{"slug":62,"navLabel":63,"heroHeadline":64},"e-commerce","E-commerce","The shop works. What sits behind it comes apart",{"slug":66,"navLabel":67,"heroHeadline":68},"logistics","Transport and logistics","The job arrives by email, the status comes back by phone",[70,77,82],{"slug":71,"client":72,"tag":73,"year":74,"summary":75,"image":76},"dss-by-caredo","Caredo","Product",2025,"We build the whole DSS product together with the Caredo team — from the CRM and planner to the software behind the key lockers and totems.","\u002Fimg\u002Fcase-studies\u002Fdss\u002Fkluczykomat\u002Fframe_0169.webp",{"slug":78,"client":72,"tag":79,"year":74,"summary":80,"image":81},"dss-by-caredo-website","Product website","We built the DSS product website end to end — from design to launch — on Caredo's visual identity.","\u002Fimg\u002Fscreenshots\u002Fdss-desktop.jpg",{"slug":83,"client":83,"tag":84,"year":74,"summary":85,"image":86},"2motion","Brand and website","A whole brand from scratch — logo, colours, business cards — a company website and blog posts that position 2motion in search.","\u002Fimg\u002Fscreenshots\u002F2motion-desktop.jpg",{"service":88,"relatedCase":199,"relatedPosts":315,"i18nSlugs":550,"subServices":552},{"id":89,"title":90,"audience":91,"body":95,"description":124,"extension":125,"faq":126,"heroHeadline":22,"heroSubline":139,"keyword":90,"meta":140,"navGroup":17,"navLabel":21,"navigation":141,"order":142,"path":143,"pricingModels":144,"process":154,"relatedCase":71,"relatedPosts":171,"scope":173,"seo":186,"sitemap":187,"slug":20,"stem":197,"__hash__":198},"servicesEn\u002Fservices\u002Fai-implementation.md","AI implementation for business",[92,93,94],"Teams drowning in documents to read, classify and retype","Support teams answering the same question for the thousandth time","Companies with a lot of internal knowledge nobody can find",{"type":96,"value":97,"toc":117},"minimark",[98,103,107,110,114],[99,100,102],"h2",{"id":101},"why-ai-pilots-dont-reach-production","Why AI pilots don't reach production",[104,105,106],"p",{},"A demo shows a model handling ten examples chosen by the person who prepared them. Production is ten thousand examples nobody chose, including scans crooked on the glass and documents in a format that wasn't in the test set.",[104,108,109],{},"The difference between the two isn't model quality. It's handling the cases where the model isn't sure — and that handling is the real substance of a rollout.",[99,111,113],{"id":112},"what-we-measure","What we measure",[104,115,116],{},"Before we start, we establish how long the work takes today and what share of cases the automation needs to handle for the rollout to make sense. After launch, the same metric tells us whether it's working. Without it, \"AI in the company\" is a cost with no verifiable return.",{"title":118,"searchDepth":119,"depth":119,"links":120},"",3,[121,123],{"id":101,"depth":122,"text":102},2,{"id":112,"depth":122,"text":113},"We deploy AI where it genuinely shortens real work — document classification, search over internal knowledge, support assistants. No pilots that never reach production.","md",[127,130,133,136],{"q":128,"a":129},"Where do we start with AI in our company?","With one process that's repetitive, has a measurable time cost, and tolerates the occasional mistake. Classifying incoming documents usually meets all three. Starting with \"an assistant that knows everything\" is the most common way to get a project with no end.",{"q":131,"a":132},"Will my data go to a public model?","Not unless you want it to. We match the solution to your requirements — from hosted models under a no-training-on-data agreement, to models run on your own infrastructure.",{"q":134,"a":135},"What if the model gets it wrong?","We design assuming it will. Every rollout has a confidence threshold — above it, the automation runs; below it, the case goes to a human. Every model decision is logged and traceable.",{"q":137,"a":138},"What does it cost?","A prototype is usually PLN 15–30k net. A production rollout is PLN 60–200k, depending on the number of integrations. On top of that, the cost of running the models — typically hundreds to a few thousand zloty a month.","Most corporate AI projects end at the demo. We start by asking which repetitive work you want gone — and whether its disappearance can be measured.",{},true,4,"\u002Fservices\u002Fai-implementation",[145,148,151],{"name":146,"body":147},"Fixed-price prototype","Quoted separately and deliberately small. If accuracy comes in below the threshold, we stop — and that's a good outcome too, just a cheaper one than finding out a year in.",{"name":149,"body":150},"Fixed-price rollout","Once the prototype exists the scope is known, so we can give an amount and a deadline.",{"name":152,"body":153},"Maintenance retainer","Accuracy monitoring, tuning and model updates.",[155,159,163,167],{"step":156,"duration":157,"body":158},"Choosing the use case","1 week","We review the candidates and pick the one with the best ratio of time saved to risk. We define the success metric before starting.",{"step":160,"duration":161,"body":162},"Prototype on your data","2–3 weeks","A working prototype on real documents, with measured accuracy. This is the decision point on whether to continue.",{"step":164,"duration":165,"body":166},"Production rollout","4–10 weeks","Integration with your systems, handling of uncertain cases, permissions and a model decision log.",{"step":168,"duration":169,"body":170},"Measurement and tuning","ongoing","Accuracy measured continuously, because data changes — a model that worked in March may not work in November.",[172],"wdrozenie-ai-w-firmie-od-czego-zaczac",[174,177,180,183],{"title":175,"body":176},"Document classification and extraction","Invoices, contracts, orders and correspondence read automatically, with fields extracted into your system and a queue of uncertain cases for a human to decide.",{"title":178,"body":179},"Search over company knowledge","A question asked in plain language, an answer with a link to the source document. No making things up — if the answer isn't in the knowledge base, the system says so.",{"title":181,"body":182},"Customer support assistants","Answers to repetitive questions, handed off to a human the moment confidence drops.",{"title":184,"body":185},"Automation with a model in the loop","The model as one step in a process, not the whole process — with the output validated before it's written to the system.",{"title":90,"description":124},{"loc":188,"alternatives":189},"\u002Fen\u002Fservices\u002Fai-implementation",[190,193,195],{"hreflang":191,"href":192},"x-default","\u002Fuslugi\u002Fwdrozenia-ai",{"hreflang":194,"href":192},"pl-PL",{"hreflang":196,"href":188},"en-US","services\u002Fai-implementation","JRpvH1T996kqI1_eb4j8O7Zhl72twBMjnOeKMf2EEUU",{"id":200,"title":201,"body":202,"challenge":267,"client":72,"description":268,"extension":125,"features":269,"meta":276,"navigation":141,"order":277,"path":278,"relatedServices":279,"results":280,"screenshotDesktop":76,"screenshotMobile":284,"seo":285,"sequence":286,"sitemap":289,"slug":71,"solution":296,"stack":297,"stem":302,"summary":75,"tag":73,"url":303,"usedBy":304,"year":74,"__hash__":314},"caseStudiesEn\u002Fcase-studies\u002Fdss-by-caredo.md","DSS by Caredo — a product we build together with the Caredo team",{"type":96,"value":203,"toc":261},[204,208,211,214,218,234,241,245,248,252],[99,205,207],{"id":206},"starting-point","Starting point",[104,209,210],{},"DSS (Digital Service Station) is a system for car dealerships and service centres that takes the customer from the moment they pull into the car park to the moment they collect their car. A totem greets them and shows them where to park, a key locker takes the key without the front desk, a parts dispenser hands a mechanic the parts for a specific work order.",[104,212,213],{},"Each of those devices only makes sense if it knows who the customer is and which work order is in progress. That is the software's job — and that is the part we build together with Caredo.",[99,215,217],{"id":216},"what-we-build-together","What we build together",[104,219,220,221,225,226,229,230,233],{},"A ",[222,223,224],"strong",{},"CRM"," where the customer, their history and their work orders are one record. A ",[222,227,228],{},"planner"," where the schedule of visits and work orders is made. A ",[222,231,232],{},"key handover system",": the customer drops off or collects a key at any hour, and the system assigns it to the work order and logs every operation on its own.",[104,235,236,237,240],{},"On top of that, ",[222,238,239],{},"device integrations"," — external and internal key lockers, the parts dispenser, the tool dispenser and the parking totem — so that each one is part of a single process rather than a separate machine with its own database.",[99,242,244],{"id":243},"device-interfaces-for-brands","Device interfaces for brands",[104,246,247],{},"A Porsche customer shouldn't see the same key-locker screen as a Kia customer. For Mercedes-Benz, Porsche, Toyota and Kia dealers we build device interfaces in each brand's identity — screens the customer uses alone, with no member of staff beside them, so they have to make sense from the first touch.",[99,249,251],{"id":250},"how-we-work-together","How we work together",[104,253,254,255,260],{},"We're not the vendor of a single module. We work with the Caredo team on the whole product: when a new device is designed, its software and its integration with the rest of the system are built alongside it, not after the fact. The product website, dssbycaredo.com, we built separately — it's a ",[256,257,259],"a",{"href":258},"\u002Fen\u002Fcase-studies\u002Fdss-by-caredo-website","separate case study",".",{"title":118,"searchDepth":119,"depth":119,"links":262},[263,264,265,266],{"id":206,"depth":122,"text":207},{"id":216,"depth":122,"text":217},{"id":243,"depth":122,"text":244},{"id":250,"depth":122,"text":251},"Caredo designs devices for dealerships and service centres: key lockers, parts dispensers, parking totems. The hardware is half the product. The other half is the software that connects each device to the customer, the work order and the service team — and that has to look like the brand's showroom, not like a technical panel.","Together with the Caredo team we build the whole DSS product: the CRM, the planner, the key handover system, the integrations of the devices on site and the device interfaces for Mercedes-Benz, Porsche, Toyota and Kia dealers.",[270,271,272,273,274,275],"A CRM with customer history, work orders and permissions per branch and role","A planner for visits and service orders","A 24\u002F7 key drop-off and pickup system that assigns each key to its work order","Integrations for the key lockers, parts and tool dispensers and the parking totem","Customer-facing device interfaces in Mercedes-Benz, Porsche, Toyota and Kia identity","An analytics dashboard with real-time data",{},1,"\u002Fcase-studies\u002Fdss-by-caredo",[6,15,24],[281],{"metric":282,"label":283},"100%","System uptime",null,{"title":201,"description":268},{"path":287,"frames":288,"poster":76},"\u002Fimg\u002Fcase-studies\u002Fdss\u002Fkluczykomat\u002Fframe_",169,{"loc":290,"alternatives":291},"\u002Fen\u002Fcase-studies\u002Fdss-by-caredo",[292,294,295],{"hreflang":191,"href":293},"\u002Frealizacje\u002Fdss-by-caredo",{"hreflang":194,"href":293},{"hreflang":196,"href":290},"We work as part of the Caredo team: together we build the CRM, the planner, the key drop-off and pickup system and the integrations that bring every device into one system. For Mercedes-Benz, Porsche, Toyota and Kia dealers we design device interfaces in each brand's identity.",[298,299,300,301],"TypeScript","Nuxt","Node.js","PostgreSQL","case-studies\u002Fdss-by-caredo","https:\u002F\u002Fdssbycaredo.com",[305,306,307,308,309,310,311,312,313],"Mercedes-Benz","Porsche","Toyota","Kia","Audi","Cupra","Ford","Škoda","Volkswagen","MS9XoW0yDSBUBi4JOZJ4VQn87W5x7rIKYo99PuI8oSg",[316],{"id":317,"title":318,"author":319,"body":320,"category":534,"date":535,"description":536,"draft":537,"editorialNote":284,"extension":125,"image":538,"meta":539,"navigation":141,"path":540,"rawbody":541,"seo":542,"sitemap":543,"stem":544,"tags":545,"translationKey":548,"updated":284,"__hash__":549},"blogEn\u002Fblog\u002Fwdrozenie-ai-w-firmie-od-czego-zaczac.md","AI adoption in your company — where to start","zespol",{"type":96,"value":321,"toc":524},[322,337,341,344,351,355,358,364,370,376,379,383,390,393,397,400,406,412,418,424,428,431,453,456,460,486,490,496,502,508,514,518,521],[104,323,324,325,328,329,332,333,336],{},"Start with a process that meets three conditions at once: it's ",[222,326,327],{},"repetitive",", has a ",[222,330,331],{},"countable time cost",", and ",[222,334,335],{},"tolerates a mistake",". Classifying incoming documents usually meets all three. \"An assistant that knows the whole company\" meets none of them — which is why so many rollouts that start there never finish.",[99,338,340],{"id":339},"why-pilots-dont-reach-production","Why pilots don't reach production",[104,342,343],{},"A demo shows a model handling a dozen or so examples chosen by the person who prepared them. Production is tens of thousands of examples nobody chose: scans crooked on the glass, documents in a format outside the test set, invoices with a handwritten note in the margin.",[104,345,346,347,350],{},"The difference between the two isn't model quality. It's ",[222,348,349],{},"handling the cases where the model isn't sure"," — and that's exactly what a rollout is actually made of. A pilot that skips it isn't an earlier version of the rollout. It's a different thing entirely.",[99,352,354],{"id":353},"how-to-choose-your-first-use-case","How to choose your first use case",[104,356,357],{},"Make a list of candidates and score each on three dimensions.",[104,359,360,363],{},[222,361,362],{},"Repeatability."," How many times a month does someone do this? Below a hundred, the savings rarely justify a rollout.",[104,365,366,369],{},[222,367,368],{},"Time cost."," How many minutes does one instance take? Multiply by repeatability. That number sets the upper bound on a sensible budget.",[104,371,372,375],{},[222,373,374],{},"Tolerance for error."," What happens if the model gets it wrong? If the answer is \"someone catches it at review\" — good. If it's \"money leaves an account\" — you need a human approval step, which changes the maths.",[104,377,378],{},"The best candidates in a typical company: classifying and routing incoming documents, extracting data from invoices and orders, drafting first responses to repetitive support questions, searching internal documentation.",[99,380,382],{"id":381},"set-a-success-metric-before-you-start","Set a success metric before you start",[104,384,385,386],{},"Before a prototype exists, write down one sentence: ",[387,388,389],"em",{},"\"This rollout makes sense if the model handles at least X% of cases on its own, at accuracy of at least Y%.\"",[104,391,392],{},"Without that number, there's no way to later decide whether the project succeeded. With it, the prototype ends in a clear decision: proceed, or stop. Stopping is also a good outcome — just a cheaper one than finding out a year in.",[99,394,396],{"id":395},"an-architecture-that-survives-production","An architecture that survives production",[104,398,399],{},"A rollout that lasts has four elements a demo doesn't.",[104,401,402,405],{},[222,403,404],{},"A confidence threshold."," The model returns a result along with its own confidence score. Above the threshold, automation runs; below it, the case goes to a human. Setting that threshold is a business decision, not a technical one.",[104,407,408,411],{},[222,409,410],{},"A queue for uncertain cases."," The place where a human resolves cases below the threshold — and its size is the real measure of how well the system is working.",[104,413,414,417],{},[222,415,416],{},"A decision log."," A record of what the model received, what it returned, and with what confidence. Without it you can't diagnose degradation or answer an auditor's question.",[104,419,420,423],{},[222,421,422],{},"Continuous measurement."," Data changes. A model that was 94% accurate in March can be at 80% by November — not because it broke, but because the input stream changed.",[99,425,427],{"id":426},"where-your-data-goes","Where your data goes",[104,429,430],{},"The question that comes up in every first meeting, rightly so. You have three options:",[432,433,434,441,447],"ul",{},[435,436,437,440],"li",{},[222,438,439],{},"A hosted model, no-training agreement."," Data leaves the company, but the provider commits not to use it for training. Cheapest and fastest, sufficient for most use cases.",[435,442,443,446],{},[222,444,445],{},"A hosted model in your own private cloud."," Data never leaves your infrastructure. More expensive, justified for sensitive data.",[435,448,449,452],{},[222,450,451],{},"A model on your own hardware."," Full control, the highest entry and running cost. Makes sense under regulatory requirements that rule out the other options.",[104,454,455],{},"The choice is a decision about risk, not technology. You make it once, and it affects cost for the entire lifetime of the solution.",[99,457,459],{"id":458},"what-it-costs","What it costs",[432,461,462,468,474,480],{},[435,463,464,467],{},[222,465,466],{},"Prototype on your data:"," PLN 15–30k net, 2–3 weeks. Ends with a number you can base a decision on.",[435,469,470,473],{},[222,471,472],{},"Production rollout:"," PLN 60–200k net, depending on the number of integrations and approval requirements.",[435,475,476,479],{},[222,477,478],{},"Running the models:"," from a few hundred to a few thousand zloty a month, roughly proportional to volume.",[435,481,482,485],{},[222,483,484],{},"Maintenance and tuning:"," usually 15% of the rollout's value per year.",[99,487,489],{"id":488},"four-mistakes-that-come-up-most-often","Four mistakes that come up most often",[104,491,492,495],{},[222,493,494],{},"Starting with the most impressive use case."," A conversational assistant impresses the board and is the hardest possible starting point — unbounded scope, no good success metric, high expectations.",[104,497,498,501],{},[222,499,500],{},"No human in the loop."," A system with no path for uncertain cases either performs badly or needs such a high confidence threshold that it only automates a small share of cases.",[104,503,504,507],{},[222,505,506],{},"Measuring accuracy alone."," 95% accuracy sounds great until you work out that at a thousand documents a day, that's fifty errors someone has to catch.",[104,509,510,513],{},[222,511,512],{},"Treating the rollout as a closed project."," A model needs watching, like any other piece of production infrastructure. A rollout with no measurement plan ages quietly.",[99,515,517],{"id":516},"the-first-step-you-can-take-this-week","The first step you can take this week",[104,519,520],{},"Pick one process. Count how many times a month it happens and how many minutes it takes. Write down what happens when someone gets it wrong.",[104,522,523],{},"Those three numbers are enough to judge whether a prototype is worth building — and they're worth more than a month of conversations about AI's possibilities.",{"title":118,"searchDepth":119,"depth":119,"links":525},[526,527,528,529,530,531,532,533],{"id":339,"depth":122,"text":340},{"id":353,"depth":122,"text":354},{"id":381,"depth":122,"text":382},{"id":395,"depth":122,"text":396},{"id":426,"depth":122,"text":427},{"id":458,"depth":122,"text":459},{"id":488,"depth":122,"text":489},{"id":516,"depth":122,"text":517},"ai","2026-09-02","How to choose your first use case, what it costs, and why most corporate AI projects end at the demo instead of reaching production.",false,"\u002Fimg\u002Fblog\u002Fwdrozenie-ai-w-firmie.jpg",{},"\u002Fblog\u002Fwdrozenie-ai-w-firmie-od-czego-zaczac","---\ntitle: AI adoption in your company — where to start\ndescription: How to choose your first use case, what it costs, and why most corporate AI projects end at the demo instead of reaching production.\ndate: 2026-09-02\ncategory: ai\ntags:\n  - ai\n  - automation\n  - rollouts\nauthor: zespol\nimage: \u002Fimg\u002Fblog\u002Fwdrozenie-ai-w-firmie.jpg\ntranslationKey: ai-adoption-where-to-start\ndraft: false\n---\n\nStart with a process that meets three conditions at once: it's **repetitive**, has a **countable time cost**, and **tolerates a mistake**. Classifying incoming documents usually meets all three. \"An assistant that knows the whole company\" meets none of them — which is why so many rollouts that start there never finish.\n\n## Why pilots don't reach production\n\nA demo shows a model handling a dozen or so examples chosen by the person who prepared them. Production is tens of thousands of examples nobody chose: scans crooked on the glass, documents in a format outside the test set, invoices with a handwritten note in the margin.\n\nThe difference between the two isn't model quality. It's **handling the cases where the model isn't sure** — and that's exactly what a rollout is actually made of. A pilot that skips it isn't an earlier version of the rollout. It's a different thing entirely.\n\n## How to choose your first use case\n\nMake a list of candidates and score each on three dimensions.\n\n**Repeatability.** How many times a month does someone do this? Below a hundred, the savings rarely justify a rollout.\n\n**Time cost.** How many minutes does one instance take? Multiply by repeatability. That number sets the upper bound on a sensible budget.\n\n**Tolerance for error.** What happens if the model gets it wrong? If the answer is \"someone catches it at review\" — good. If it's \"money leaves an account\" — you need a human approval step, which changes the maths.\n\nThe best candidates in a typical company: classifying and routing incoming documents, extracting data from invoices and orders, drafting first responses to repetitive support questions, searching internal documentation.\n\n## Set a success metric before you start\n\nBefore a prototype exists, write down one sentence: *\"This rollout makes sense if the model handles at least X% of cases on its own, at accuracy of at least Y%.\"*\n\nWithout that number, there's no way to later decide whether the project succeeded. With it, the prototype ends in a clear decision: proceed, or stop. Stopping is also a good outcome — just a cheaper one than finding out a year in.\n\n## An architecture that survives production\n\nA rollout that lasts has four elements a demo doesn't.\n\n**A confidence threshold.** The model returns a result along with its own confidence score. Above the threshold, automation runs; below it, the case goes to a human. Setting that threshold is a business decision, not a technical one.\n\n**A queue for uncertain cases.** The place where a human resolves cases below the threshold — and its size is the real measure of how well the system is working.\n\n**A decision log.** A record of what the model received, what it returned, and with what confidence. Without it you can't diagnose degradation or answer an auditor's question.\n\n**Continuous measurement.** Data changes. A model that was 94% accurate in March can be at 80% by November — not because it broke, but because the input stream changed.\n\n## Where your data goes\n\nThe question that comes up in every first meeting, rightly so. You have three options:\n\n- **A hosted model, no-training agreement.** Data leaves the company, but the provider commits not to use it for training. Cheapest and fastest, sufficient for most use cases.\n- **A hosted model in your own private cloud.** Data never leaves your infrastructure. More expensive, justified for sensitive data.\n- **A model on your own hardware.** Full control, the highest entry and running cost. Makes sense under regulatory requirements that rule out the other options.\n\nThe choice is a decision about risk, not technology. You make it once, and it affects cost for the entire lifetime of the solution.\n\n## What it costs\n\n- **Prototype on your data:** PLN 15–30k net, 2–3 weeks. Ends with a number you can base a decision on.\n- **Production rollout:** PLN 60–200k net, depending on the number of integrations and approval requirements.\n- **Running the models:** from a few hundred to a few thousand zloty a month, roughly proportional to volume.\n- **Maintenance and tuning:** usually 15% of the rollout's value per year.\n\n## Four mistakes that come up most often\n\n**Starting with the most impressive use case.** A conversational assistant impresses the board and is the hardest possible starting point — unbounded scope, no good success metric, high expectations.\n\n**No human in the loop.** A system with no path for uncertain cases either performs badly or needs such a high confidence threshold that it only automates a small share of cases.\n\n**Measuring accuracy alone.** 95% accuracy sounds great until you work out that at a thousand documents a day, that's fifty errors someone has to catch.\n\n**Treating the rollout as a closed project.** A model needs watching, like any other piece of production infrastructure. A rollout with no measurement plan ages quietly.\n\n## The first step you can take this week\n\nPick one process. Count how many times a month it happens and how many minutes it takes. Write down what happens when someone gets it wrong.\n\nThose three numbers are enough to judge whether a prototype is worth building — and they're worth more than a month of conversations about AI's possibilities.\n",{"title":318,"description":536},{"loc":540},"blog\u002Fwdrozenie-ai-w-firmie-od-czego-zaczac",[534,546,547],"automation","rollouts","ai-adoption-where-to-start","3h3Uyc_r6v-R1oggm3sXVR3s8arJjyy0xQy78uryQSg",{"pl":551,"en":20},"wdrozenia-ai",[553,620,688],{"id":554,"title":555,"audience":556,"body":559,"description":579,"extension":125,"faq":580,"heroHeadline":590,"heroSubline":591,"keyword":592,"meta":593,"navigation":141,"order":594,"parent":20,"path":595,"relatedCase":284,"scope":596,"seo":609,"sitemap":610,"slug":617,"stem":618,"__hash__":619},"subServicesEn\u002Fservices\u002Fai-implementation\u002Fchatbots-and-assistants.md","AI chatbots and support assistants",[93,557,558],"Companies with extensive product documentation customers can't navigate","Teams where onboarding a new support person takes months",{"type":96,"value":560,"toc":575},[561,565,568,572],[99,562,564],{"id":563},"where-we-start","Where we start",[104,566,567],{},"By exporting the last few months of tickets and counting how many are variants\nof the same question. That figure decides whether the project makes sense — and\nit can be known before any quote.",[99,569,571],{"id":570},"when-we-advise-against-it","When we advise against it",[104,573,574],{},"When customers' questions are different every time, when the documentation is\nout of date, or when answering requires checking something in a system the\nassistant cannot reach. In all three cases it adds work rather than removing\nit.",{"title":118,"searchDepth":119,"depth":119,"links":576},[577,578],{"id":563,"depth":122,"text":564},{"id":570,"depth":122,"text":571},"An assistant that answers from your own knowledge, not a general model. With stated limits, a hand-off to a person, and a measurable effect.",[581,584,587],{"q":582,"a":583},"Can the assistant make an answer up?","A language model always can. That is why we build it to answer only from supplied sources and show them, and to have sensitive subjects cut off by a rule rather than by a request in a prompt. The risk drops to a level you can control — not to zero, and we say so.",{"q":585,"a":586},"How much content is needed for this to work?","Less than people assume, but it has to be current. Thirty good pages of documentation work better than three hundred out-of-date ones, because the assistant will reproduce the errors too.",{"q":588,"a":589},"Will it replace the support team?","No, and we don't sell it that way. The real effect is removing repeat questions so the team handles the cases that genuinely need a person.","An assistant that answers from your knowledge","A bot wired to a general model invents answers. An assistant wired to your documentation, price list and ticket history answers, or says it doesn't know — and that is the whole difference.","AI chatbot for business",{},9,"\u002Fservices\u002Fai-implementation\u002Fchatbots-and-assistants",[597,600,603,606],{"title":598,"body":599},"Answers grounded in your sources","The assistant answers from specific documents and shows which. Without that there is no way to verify an answer or fix it at source, and without the ability to fix it the rollout dies within a month.",{"title":601,"body":602},"Explicit limits","A list of subjects the assistant does not touch — individual pricing, contractual commitments, complaints. \"I don't know, passing this on\" is a correct answer and has to be a designed one.",{"title":604,"body":605},"Hand-off with context","When the assistant can't cope, the conversation reaches a person along with its history rather than starting again. Otherwise the bot adds work instead of removing it.",{"title":607,"body":608},"Measurement from day one","Share of conversations closed without a person, subjects handed off most often, answers marked wrong. Without those figures there is no way to say whether the rollout is worth it — and that question arrives after a quarter.",{"title":555,"description":579},{"loc":611,"alternatives":612},"\u002Fen\u002Fservices\u002Fai-implementation\u002Fchatbots-and-assistants",[613,615,616],{"hreflang":191,"href":614},"\u002Fuslugi\u002Fwdrozenia-ai\u002Fchatboty-i-asystenci",{"hreflang":194,"href":614},{"hreflang":196,"href":611},"chatbots-and-assistants","services\u002Fai-implementation\u002Fchatbots-and-assistants","AWCVJDmo88rcgm11f2YtB3YELFvGuWTPgt_Q7H9qYhc",{"id":621,"title":622,"audience":623,"body":627,"description":647,"extension":125,"faq":648,"heroHeadline":658,"heroSubline":659,"keyword":660,"meta":661,"navigation":141,"order":662,"parent":20,"path":663,"relatedCase":284,"scope":664,"seo":677,"sitemap":678,"slug":685,"stem":686,"__hash__":687},"subServicesEn\u002Fservices\u002Fai-implementation\u002Fsemantic-search.md","Semantic search over internal knowledge",[624,625,626],"Organisations with knowledge scattered across drives and systems","Teams where onboarding means asking colleagues","Companies running projects where the same thing has been solved before",{"type":96,"value":628,"toc":643},[629,633,636,640],[99,630,632],{"id":631},"how-to-check-this-before-a-project","How to check this before a project",[104,634,635],{},"Collect ten questions somebody asked in the company chat last month instead of\nlooking the answer up. If the answers exist in documents, the project makes\nsense and can be measured.",[99,637,639],{"id":638},"what-we-dont-promise","What we don't promise",[104,641,642],{},"We don't promise the tool will tidy up disorganised documentation. Semantic\nsearch softens the effects of dispersion, but contradictory procedures stay\ncontradictory — it just becomes easier to notice.",{"title":118,"searchDepth":119,"depth":119,"links":644},[645,646],{"id":631,"depth":122,"text":632},{"id":638,"depth":122,"text":639},"Search that understands the question rather than matching keywords. Documentation, procedures and project history you can finally find things in.",[649,652,655],{"q":650,"a":651},"Will our documents go to an external model?","That is the first thing we settle, and there are several options: a model running in your infrastructure, a provider contractually barred from training on your data, or an architecture where the model only ever sees the fragment needed to answer. The choice depends on how sensitive the content is.",{"q":653,"a":654},"How many documents does this need?","Volume isn't the deciding factor — dispersion is, along with how often somebody is looking for something. With five hundred documents in one well-named folder, ordinary search is enough, and we will say so.",{"q":656,"a":657},"Does it update itself as documents are added?","Yes, the index refreshes automatically. It is worth knowing that it will also surface out-of-date documents — which is why these rollouts usually expose a need to mark what still applies.","The knowledge exists. Finding it is the problem","In most companies we have looked at, the answer to an employee's question existed in a document from two years ago. Keyword search didn't find it, because the document called the same thing something else.","search over internal company knowledge",{},10,"\u002Fservices\u002Fai-implementation\u002Fsemantic-search",[665,668,671,674],{"title":666,"body":667},"Indexing what you already have","Documents, procedures, proposals, notes and tickets from the systems you already use. Without moving everything into a new tool — migrating knowledge into yet another place is the most common reason such projects die.",{"title":669,"body":670},"Search by meaning","A question in your own words finds a document that uses different terminology. That is the difference between \"there isn't one\" and \"there is, it's just called something else\".",{"title":672,"body":673},"Permissions respected in results","Search must not show somebody a document they have no access to. It sounds obvious and is the most commonly skipped requirement in this kind of rollout.",{"title":675,"body":676},"Answers with their source","A summary with a link to the document it came from. The employee has to be able to check — otherwise the tool is only a faster way to get an uncertain answer.",{"title":622,"description":647},{"loc":679,"alternatives":680},"\u002Fen\u002Fservices\u002Fai-implementation\u002Fsemantic-search",[681,683,684],{"hreflang":191,"href":682},"\u002Fuslugi\u002Fwdrozenia-ai\u002Fwyszukiwanie-semantyczne",{"hreflang":194,"href":682},{"hreflang":196,"href":679},"semantic-search","services\u002Fai-implementation\u002Fsemantic-search","AwHF2peyRbV-T0ANvN5nTOSf7mFGeTZInCdQkja7D0k",{"id":689,"title":690,"audience":691,"body":695,"description":713,"extension":125,"faq":714,"heroHeadline":724,"heroSubline":725,"keyword":726,"meta":727,"navigation":141,"order":728,"parent":20,"path":729,"relatedCase":284,"scope":730,"seo":743,"sitemap":744,"slug":751,"stem":752,"__hash__":753},"subServicesEn\u002Fservices\u002Fai-implementation\u002Fdocument-classification.md","Document classification and data extraction",[692,693,694],"Teams where somebody retypes invoices or orders into a system","Companies receiving documents in many formats from many suppliers","Departments buried in documents to read and classify",{"type":96,"value":696,"toc":709},[697,701,704,706],[99,698,700],{"id":699},"how-we-calculate-the-case","How we calculate the case",[104,702,703],{},"Documents per month times the time to retype one, minus the time spent\nreviewing cases below the threshold. That simple difference is usually enough\nto decide, and it can be worked out before the project starts.",[99,705,564],{"id":563},[104,707,708],{},"With a representative sample, not the tidiest documents. Rollouts like this\nbreak on edge cases, so the sample has to contain them — otherwise the accuracy\nmeasurement is worthless.",{"title":118,"searchDepth":119,"depth":119,"links":710},[711,712],{"id":699,"depth":122,"text":700},{"id":563,"depth":122,"text":564},"Invoices, orders and contracts read automatically instead of retyped. With a confidence threshold and a review path for anything unclear.",[715,718,721],{"q":716,"a":717},"How accurate is it?","It depends on the quality and consistency of the documents, and we don't give a figure before testing on yours. We usually start with a sample of a few hundred — measuring on those is cheaper than any claim made in advance.",{"q":719,"a":720},"What happens when it gets something wrong?","That is what the confidence threshold and the review screen are for. Automation without a path for doubtful cases pushes errors deeper into the process, where they cost many times more than at data entry.",{"q":722,"a":723},"Will it handle scans and photographs?","Yes, though input quality translates directly into the result. With documents photographed on a phone, the cheapest improvement is usually changing how they are received, not a stronger model.","Nobody should be retyping data out of a PDF","This is work that can be counted in hours per week and whose effect can be measured in errors avoided. One of the few AI use cases where the return shows up in the first month.","automatic document classification",{},11,"\u002Fservices\u002Fai-implementation\u002Fdocument-classification",[731,734,737,740],{"title":732,"body":733},"Recognising the document type","Invoice, order, report, contract — separated automatically on intake. Sorting at the door is usually a bigger saving than the extraction itself.",{"title":735,"body":736},"Field extraction with a confidence threshold","Number, dates, amounts, counterparty, line items. Every field with a confidence score — anything below the threshold goes to review rather than into the system.",{"title":738,"body":739},"A path for unclear cases","A review screen with the document beside the extracted data. This is the part whose absence turns automation into a source of errors harder to catch than the manual ones.",{"title":741,"body":742},"Wired into the target system","Data lands where it was going anyway — in the ERP or the document flow. Without that step the project ends in a CSV somebody has to import.",{"title":690,"description":713},{"loc":745,"alternatives":746},"\u002Fen\u002Fservices\u002Fai-implementation\u002Fdocument-classification",[747,749,750],{"hreflang":191,"href":748},"\u002Fuslugi\u002Fwdrozenia-ai\u002Fklasyfikacja-dokumentow",{"hreflang":194,"href":748},{"hreflang":196,"href":745},"document-classification","services\u002Fai-implementation\u002Fdocument-classification","99cepK35Ae3QFiqrocyMAmLUAlpCRLJ8_TAOOXhNfYU",[755],{"code":756,"label":757,"name":758,"to":192},"pl","PL","Polish",1790522544587]