AI practice
Grapefruit is an AI-first digital and software company from Romania that helps enterprises move AI from pilot to production. We find the use cases worth funding, prove them on your data in weeks, then build, secure and run AI agents, knowledge assistants and custom models, with GDPR and EU AI Act controls from day one.
Proof, not adjectives
What we do not do: train foundation models from scratch, sell AI licences, or give legal opinions. We build on proven models, pick them per use case, and work with your counsel on compliance.
Stuck in pilots?
Pilots start from a tool demo, so nobody can say what success is worth or compare it with other options.
How we fix it: The Opportunity Scan ranks use cases and models the ROI before anything is built.
The demo looked impressive, but there was no test set, so legal and the business do not trust it in front of customers.
How we fix it: Every prototype has an evaluation set and a hallucination report agreed in week 1.
Data flows, hosting and EU AI Act duties were raised after the pilot, and the project stalled in review.
How we fix it: We classify risk and document data flows at the start, in your cloud tenant.
The pilot team moved on, models changed, and quality quietly dropped until people stopped using it.
How we fix it: We operate what we build: monitoring, re-evaluation and monthly improvements.
How to start
See where your company stands on AI in 10 minutes, and what to fix first.
Find the 3 AI use cases worth funding in your business, with the numbers to defend them.
A working AI prototype on your own data in 4 weeks, so you decide with evidence, not slides.
From validated prototype to a production AI system your team trusts and runs.
After launch: AI Operations Partner. Your AI keeps working, stays compliant and keeps improving.
What AI can realistically do in your industry this year, with examples from companies like yours.
Find out how ChatGPT, Gemini, Perplexity and Google AI Overviews describe your brand, and what to fix so they get it right.
Talk to someone who has built this before, not a salesperson.
Capabilities
AI agent development for enterprise workflows: agents that triage, draft, update systems and report, with permissions, evaluation and human approval built in. Start with a prototype sprint.
Starts with: AI Prototype Sprint
AI consulting for enterprises: use-case discovery, data readiness, EU AI Act risk classification, ROI model and a 90-day roadmap. Start with a 2-week AI Opportunity Scan.
Starts with: AI Opportunity Scan
Custom AI development from proof of concept to production: generative AI features, fine-tuned models and prediction tools, evaluated on your data and deployed in your cloud.
Starts with: AI Prototype Sprint
Enterprise knowledge assistants built with retrieval-augmented generation (RAG): answers from your documents with citations, access control, evaluation and human hand-off.
Starts with: AI Prototype Sprint
AI for marketing and sales teams: brand-voice content models, AI-assisted creative and localisation, CRM workflows and AI search visibility (GEO). UNTOLD cut newsletter creation time by 80%.
Starts with: AI Visibility Audit
Use cases by industry
Problem: Content volume grows faster than the team, and generic AI drafts get rewritten from scratch.
Approach: Models fine-tuned on your approved content draft newsletters and campaign copy; editors approve every piece.
Problem: Customers call about bills, contracts and policies that are already explained somewhere on your portal.
Approach: An assistant in your portal or app answers from your documents with citations and hands over to an agent when unsure.
Problem: Contact-centre and sales staff search several systems for one answer while the customer waits.
Approach: Retrieval over procedures, product sheets and tickets, respecting existing access rights, inside the tools staff already use.
Problem: Loyalty programmes hold years of member data but every member still gets the same message.
Approach: Segmentation and next-best-offer suggestions on your existing loyalty and CRM data, with marketers approving what goes out.
Problem: Buyers ask ChatGPT, Gemini and AI Overviews, and your brand is missing or described incorrectly.
Approach: Measure share of answer on your buyers' real questions, then fix content, entities and technical signals.
Problem: Energy producers, suppliers and distributors plan on forecasts that live in spreadsheets few people can read.
Approach: We designed and built the product experience (UX, UI and front-end) that makes AI forecasts usable for daily decisions.
Problem: Teams re-key requests between email, CRM and ticketing, and backlogs depend on who is on shift.
Approach: Agents triage, draft and update systems within scoped permissions, with a person approving sensitive steps.
See it work
Client work · Entertainment & Events
80% faster newsletter creation
Custom GPT-4.1 models fine-tuned on UNTOLD's own newsletters, in Romanian and English, draft each issue and copywriters edit. The first version was ready in 1 week with one AI engineer and one copywriter.
Read the UNTOLD case study →Built by Grapefruit · Own product
Our own generative AI product: personalised bedtime stories in which the child is the protagonist, orchestrating several AI models behind a simple interface. We practise on our own product what we build for clients.
Visit Povelia →Client work · Energy & Utilities
A B2B platform that uses AI and machine learning to forecast energy demand and prices. From fall 2020, Grapefruit owned the UX, interface design and front-end code, taking it from MVP to a stable platform.
Read the Ogre.ai case study →Our method
Find and rank the use cases worth funding: interviews, data readiness, EU AI Act risk and an ROI model.
Prove the top use case on your data in your tenant, measured against an evaluation set agreed up front.
Production architecture, integrations, guardrails, human review, documentation and training.
Monitoring, re-evaluation on model updates, cost optimisation and a quarterly use-case review.
Responsible AI & data
Systems run in your cloud tenant or an EU region we agree on. Your data is not used to train public models, and a data processing agreement covers our work.
Data minimisation, documented data flows and access limited to named people, reviewed with your DPO before anything touches personal data.
Each use case is classified by risk at the start. Article 50 transparency duties have applied since 2 August 2026; high-risk obligations are expected from 2 December 2027 under the Digital Omnibus.
People approve the outputs and actions that matter. At UNTOLD, AI drafts and copywriters edit and publish.
Accuracy and hallucinations are measured on your own test cases before launch and after every model update. No absolute guarantees.
We are model-agnostic and hand over code, prompts, evaluation sets and documentation, so you can switch models or partners.
We do not give legal advice; we work alongside your counsel. Security documentation, our AI policy and the certifications we actually hold are on Trust & compliance.
Who builds it
Delivery Manager
AI · AI solutions delivery · Digital product delivery
FAQ
Most start small and fixed-scope. Teams still exploring take the free AI Readiness Assessment or a leadership briefing. Teams that need to pick use cases book a 2-week AI Opportunity Scan. Teams with a chosen use case start with a 3 to 4-week AI Prototype Sprint. Each step ends with a clear recommendation.
No. We build in your cloud tenant or an EU region we agree on, use enterprise model terms under which your data is not used for training, and sign a data processing agreement. Data flows are documented for your security team and DPO before any personal data is involved.
We measure them instead of promising they will not happen. Every system gets an evaluation set of your real cases, run before launch and after every model update, plus guardrails, citations where facts matter and human review where errors are costly.
Article 50 transparency duties, such as telling people they are talking to AI, have applied since 2 August 2026. High-risk obligations are expected from 2 December 2027 under the Digital Omnibus. We classify each use case by risk at the start and document the controls; your legal counsel confirms formal positions.
No. We are model-agnostic, keep models swappable, and hand over code, prompts, evaluation sets and documentation. Every step on the ladder ends with no obligation to continue, and our deliverables are written so your team or another partner can pick them up.
You own what we build for you: code, prompts, fine-tuned models in your account, evaluation sets and documentation. Ownership and any licensing of pre-existing components are written into the contract before work starts, so there are no surprises at handover.
Integrations and data preparation usually matter more than the model. Other drivers are security requirements, the number of languages and users, and how much human review is needed. That is why we start with fixed-scope steps and estimate the build from measured prototype results.
Because production AI is mostly software, data and user experience, which we have shipped for enterprises since 1999. Our AI work is built in-house, UNTOLD's fine-tuned models cut newsletter creation time by 80%, and we run our own generative AI product, Povelia.
Start with a 2-week AI Opportunity Scan, or talk to a senior practitioner first. No obligation to continue.
We reply within 1 business day, usually the same day. Not ready to talk? Take the free AI assessment.
