# AI strategy, data and AI agents We get your data and processes ready for AI, then build the AI tools that run on them, inside the systems you already use. **AI Eval.** Test sets built from real cases on three of our projects. - 200+ curated questions to tune an assistant - ~100 acceptance cases, security tests included - 3,000 posts in a gold test set ## What we build ### Strategy & Data We get your data and processes ready for AI. - **AI Roadmap**: AI ideas everywhere, no clear priorities → A concrete plan. Where to start and why - [Data Integration](/case-studies#erp-kpi-platform): Same data, different numbers in every report → One trusted source of truth, ready for AI - **Knowledge base**: Know-how in people's heads and old files → Company knowledge AI can answer from ### AI Product Development We build AI tools that work inside your systems. - [Agentic Workflows](/case-studies#catalogue-sales-assistant): Routine work done manually → Agents run it for you in your systems - [AI Governance](/case-studies#site-report-agent): Every AI tool reaching your data its own way → One access layer with rules and policies - [AI Data Pipelines](/case-studies#social-signal-pipeline): Complex data that must be interpreted → Millions of records turned into signals ## Measured in production ### Assistant on a technical catalogue A large industrial manufacturer with an extensive technical catalogue. Pillar: Systems. - 252 users in the first 10 weeks - 5,873 questions in the first 10 weeks - 1,800+ technical documents indexed **Problem.** After a merger, the sales reps were working from two separate sets of technical documents, more than 1,800 between them, with no shared way to find the right product code. Answering a customer meant digging through datasheets, then looking up price, stock and lead time in the ERP. The company's own team had already built a tool for this and rejected it as unreliable. **What we built:** - An assistant that answers questions in plain language with the right product code and the sources behind the answer - Price, stock, minimum order and lead time read from the ERP as the signed-in user - Search across more than 1,800 technical documents, combining keywords, vectors and reranking - An evaluation of correctness, faithfulness to sources and recall; every change is tested on 50 real questions before release - Hosting on Azure, with sign-in through the company's own accounts ### One set of KPIs, rebuilt from the ERP A mid-sized wholesale distributor running SAP Business One. Pillar: Systems. - 21 KPIs reconciled to the cent - 30 source tables read every night **Problem.** Sales, finance and logistics each took their numbers from a different tool and defined the same KPIs in their own way, so two reports could give two answers to the same question. 21 of the KPIs management needed were not visible to it at all. Putting AI on top of that would only have scaled the confusion. **What we built:** - KPIs on sales, receivables, stock and logistics, rebuilt every night from 30 raw source tables and read-only on the ERP - Every KPI built twice, independently, and released only when both versions matched to the cent - A dashboard and an API behind the company's own sign-in, with roles taken from the company's own groups - A read-only connector so people can ask Claude for the KPIs; it acts as the user who asks and refuses anything but reads ### Social posts turned into weekly trend signals A data company that sells trend signals to the music industry. Pillar: Systems. - 17.9M+ posts collected from 4 platforms - 103,676 artists tracked - $0.0001 AI cost per post - 54% lower cost per post at equal quality **Problem.** The client sells weekly trend signals to record labels, and its product ran on social data bought from an outside supplier. When the supplier cut off the feed, the product had nothing left to run on. It needed a pipeline of its own, able to read millions of posts across four platforms at a cost that stays low at that scale. **What we built:** - Collection from YouTube, Reddit, TikTok and Instagram in parallel, where one platform failing does not stop the others - Rules that match genres and keywords, and a model that reads each post only for artists named in free text - A cache per prompt version, so re-running millions of posts costs no new model calls - A metered API with keys and quotas per customer, and a dashboard - Twelve release gates on a 3,000-post test set, with a test proving the gates fail when something breaks ## How a project runs The evaluation is how we show that a system works. The same four steps run on every project, with tools adapted to the problem. - **Agree what good means**: We agree with you, before we build, which answers count as right and how often they must be right. - **Build the test set from real cases**: Questions, documents and records taken from your own work. - **Measure every change against a baseline**: Each change is compared with the version before it, so a regression shows before release. - **Release on the result**: A change ships only when it meets the bar. ## Common questions ### What does AI data readiness mean? It means the data an AI will use is consistent and trusted, because if two systems disagree, no AI can tell which one is right. For a wholesale distributor on SAP Business One we rebuilt 21 KPIs from the raw ERP tables and reconciled each one to the cent before any AI used them. ### What can AI agents do for a business? They can run routine work inside the company's own systems, once they have clean data, integration, rules and human verification. An assistant we built finds the right product in a large technical catalogue with sources, price, stock and lead time; 252 people used it in its first 10 weeks. ### What is AI governance? AI governance is the set of rules that decides what each AI tool may see and do. We build one access layer with rules and policies, so every AI tool reaches the data the same way. Our catalogue assistant reads the ERP as the signed-in user, and our KPI connector can only read. ### Do you work with our existing ERP? Yes. We connect to the ERP you run, including SAP, NetSuite, Dynamics 365 Business Central and Odoo, and choose the approach from your workflows and constraints. ### How is custom work priced? We price each project after scoping it with you. A phase with a clear scope and an agreed definition of "working" can have a fixed price. Production cloud costs sit on your own cloud bill. ### What happens after a system is delivered? The team that built it can keep it running, with a monthly package of support hours for questions, fixes, small changes, dependency updates and monitoring. ## Ready to put AI to work in your systems? Book a call and tell us which process you want to fix and where its data lives today.