Menu
Ray Quattromini from Fortuna Data talks to Nipa Soni, who leads AI and HPC enablement at Fsas Technologies, about why Fsas built a fully on-premise Private GPT solution — and what it actually takes to deploy one safely.
Nipa Soni leads AI and HPC enablement across Fsas Technologies in the UK — working directly with customers to understand their business challenges and bring practical AI solutions to the table. The starting point for Fsas's Private GPT was simple: a clear market need for something private, secure, and aligned with government regulation.
Many organisations were already using cloud-based GPT tools accessible to anyone — but as Nipa puts it, an organisation's data is fundamental to what it does. Without proper protection at every layer, particularly with AI, that data is exposed in ways most businesses haven't fully reckoned with.
Private GPT runs fully on-premise. All data, all prompts, and all information provided stays entirely within the customer's own environment — giving full control over security and flexibility, without paying for capability the organisation doesn't need.
One of the clearest differences from cloud AI tools is licensing. Every command sent to a cloud AI service typically consumes tokens you pay for. Fsas's Private GPT doesn't work that way — it's licensed per number of users, with no token-based restriction. Departments that use it heavily aren't penalised with rising costs the way they would be on a token-metered cloud platform.
Although there are no limitations on the number of users, you are limited by your hardware. You wouldn't be able to run an effective agentic system with just a single L40S GPU, for example — the right infrastructure has to be sized to match the workload, not just the user count.
Not sure what hardware your organisation would need to run Private GPT effectively? Our AI workshop covers exactly this.
Book a workshop conversation →Security and data sovereignty sit at the foundation of everything Fsas builds — not as an add-on, but as the starting design principle. Every Private GPT deployment runs entirely on-premise with no outside access to the internet. Document-heavy processes can be automated, and users can have a full conversation with their own data without the risk of hallucination that comes from a model reaching outside into the open internet for unreliable or outdated information.
When new feature updates are released — Nipa gives the example of a customer on Private GPT version 1.5 receiving a 1.6 update — those updates are delivered via a downloadable link the customer applies themselves. No additional remote access is granted to Fsas at any point, even during upgrades.
Every Private GPT deployment starts with a workshop, priced from £2,500. The reasoning is straightforward — AI consultancy costs can spiral very quickly, and Fsas deliberately avoids that. The workshop lets the customer experience their own data inside a working Private GPT model before any larger commitment is made.
Stakeholders from across the business join a session — typically half a day to a full day — alongside Fsas specialists such as AI consultants, mathematicians, or data scientists depending on what the customer needs. In a typical workshop with around 15 stakeholders, somewhere between 30 and 60 potential use cases usually surface. The job of the workshop is to narrow that down to one or two specific, high-value use cases worth pursuing.
Within a week to two weeks of the workshop, the customer receives a full report covering findings, recommendations, and next steps. There's no additional cost between that report and a proof of concept — the customer only moves to a commercial quotation once the proof of concept has proven it meets their objectives.
Fortuna Data runs three workshop formats with Fsas Technologies, designed to match where an organisation is in its AI journey:
A focused session to map your current AI readiness, identify the highest-value use cases across your organisation, and build a clear roadmap before any infrastructure is specified.
For organisations exploring agentic AI specifically — understanding what agentic systems can realistically do for your workflows, and what infrastructure that actually requires.
A hands-on session for technical teams to design the right AI platform architecture — covering hardware sizing, integration points, and deployment planning in detail.
Want to find the right starting point for your organisation? Tell us where you are in your AI journey and we'll recommend the right workshop.
Speak to our team →Private GPT supports both API connections and MCP (Model Context Protocol), meaning customers can connect their own applications into the platform. This matters for organisations wanting to link Private GPT into existing systems — Nipa specifically references asset management platforms like IBM Maximo as an example use case explored during the workshop process.
The platform supports APIs and MCP connections, meaning customers can connect their own applications. However, Fsas does not build the MCP servers or API connectors themselves — that work is done by the customer, or by using existing connectors where they already exist. For organisations wanting deeper custom development beyond this level, Fsas moves into a different enterprise-tier solution using models such as Cohere rather than Mistral.
A common concern from organisations considering Private GPT is whether different departments can be properly siloed from each other — finance shouldn't see what sales is doing, and vice versa. This is handled entirely at the technical configuration level when the project is set up in the customer's environment. Access rules can be defined down to individual department, group, or seniority level — including segregating R&D or intellectual property-sensitive teams from the rest of the organisation entirely. All of this is customisable to however the customer wants to structure access.
For non-technical users, Private GPT works like any familiar chatbot interface — a saved link, a login, and a conversation, except the model is talking to your own data rather than the open internet. An administration portal allows IT teams to pre-load department-specific prompts — for example, different default prompts for HR versus finance versus sales — while still allowing individual users to create their own. Training is provided as part of the Fsas and Fortuna Data partnership, covering both the technical administration side and day-to-day use for non-technical staff.
The EU AI Act regulates the AI market by forcing vendors to classify their systems according to a defined risk matrix. While Fsas is a Japanese company, its UK operation sits within the European business and is consequently more closely aligned with European regulatory frameworks — including the AI Act — than with equivalent US approaches to AI governance. For UK organisations, particularly those in regulated sectors, this alignment matters when evaluating which AI vendor's compliance posture fits their own obligations.
One recent deployment involved a UK manufacturing organisation that supplies parts to multiple third-party manufacturers. They were experiencing a high error rate in sending the correct part for the correct product — a problem with real financial impact through returns and rework. After loading their own data into a Private GPT model and training it against their specific parameters, the organisation saw a significant reduction in errors within around three months, directly reducing returns and improving profitability.
Elsewhere in Europe, Fsas has seen particularly strong adoption in public sector and law firm environments — both sectors with heavily document-driven processes that benefit from being able to query large volumes of internal documentation conversationally, securely, and without the data ever leaving their own environment.
Because Private GPT runs entirely on-premise under the customer's control, ethics and governance are described as being built in "from the ground up" rather than bolted on afterwards. This matters particularly for regulated sectors such as education, where Fsas has gone beyond the technical platform itself — running a programme for first-year college students exploring the ethics of AI, including genuinely difficult questions like whether AI should be used to mark exam papers. The principle running through both the technical platform and this educational work is the same: AI can usually do whatever you ask it to — the harder question is whether it should, and there always needs to be a human in the loop.
If an organisation is interested but doesn't yet know exactly what they want from AI, that's a normal starting point according to Nipa — Fsas expects to help shape that direction, not require it upfront. After an initial conversation to understand basic requirements, Fsas typically needs around two weeks to assemble the right specialists for the workshop itself, meaning the realistic timeline from first phone call to workshop day is roughly a fortnight.
If something in this episode sparked an idea for your organisation, let's have that conversation. No pressure, no obligation — just a straight discussion about whether Private GPT is the right fit.
Get in touch →