Below the
surface

A deep dive into full-trust AI for SMEs.

In this guide

Real AI adoption needs sensitive data, and commercial LLMs can't be fully trusted with it.

The path to a full-trust AI stack is easier and more affordable than most companies think.

We compare 6 models, explain 5 layers in the stack, and outline 4 steps to get you started.

Introduction

Business AI doesn't go deep enough. It should.

Most SMEs have adopted AI at a superficial level. Drafting and summarising are useful, but the real value lies in the sensitive records AI is not allowed to touch. For a good reason: commercial AI was not built for trust.

This paper sets out what full-trust AI requires, which models deserve attention, what the overall stack looks like and how to approach it. To explore what this means for your organisation, book a free consultation on our website.

Giles and Theo
Your Beside Partners

Giles Thurston and Theo Paraskevopoulos

What we observed

Most AI projects stop where the data gets interesting.

Trust is everything

Customers and regulators are asking harder questions.

Generated text has made competence cheap to imitate. A polished proposal takes a minute to produce, so polish no longer tells a buyer anything. The questions have moved to how a firm knows what it claims: who checked the work, what it drew on, and where the data went.

Regulation is moving the same way, and not on the timetable most assume. Transparency and literacy duties already apply; only the high-risk rules were deferred. Forward-thinking firms could start treating accreditation as a commercial asset rather than a compliance cost.

Are you ready for the EU AI Act?

ObligationApplies from
Prohibited practicesFeb 2025
AI literacy dutiesFeb 2025
General-purpose model rulesAug 2025
Transparency, Article 50Aug 2026
Marking synthetic contentDec 2026
High-risk uses, Annex IIIDec 2027
High-risk, inside productsAug 2028

Regulation (EU) 2026/1744 (Digital Omnibus on AI), amending Regulation (EU) 2024/1689. In force 27 July 2026.

Why AI remains shallow

Policies exclude the data that makes AI valuable.

Most AI in an SME runs on the green zone alone: material that was already public, where nothing is at risk and little is at stake. The amber zone is rarely ruled out explicitly. It is simply left undefined, so people avoid it. The red tier is excluded outright.

Value climbs as the lights change. Almost nothing in the green zone moves a number; nearly everything that would sits in amber and red. None of it is blocked by the technology. All of it is blocked by a classification made, usually quickly, when the policy was written.

Public data

  • Product brochures
  • Public research
  • News and blog posts
  • Team bios
  • Help content

Internal data

  • Meeting notes
  • Support triage
  • Compliance
  • Complaints
  • Project updates

Restricted data

  • Client data
  • Contracts
  • Financial data
  • IP documentation
  • HR files

The limits of commercial AI

Commercial AI is built for exploring, not for trusting.

Commercial AI comes with serious limitations. None of these limits are accidental, and none are villainous. They follow from what the commercial frontier stack was designed to be: an open, general tool sold at enormous scale.

That design serves a great many purposes well. It serves poorly the narrow case of a mid-sized firm that needs a component it can govern, forecast and evidence.

Designed for business?

Business softwareCommercial AI
Solve real challengesExplore AI frontier
Minimise riskMove fast, break things
Control over dataUS residency (largely)
Business focusConsumer focus
Reduce switching costsVendor lock-in
Predictable cost structureUncertain token pricing

What gets missed

Would you run your website on LinkedIn?

Running a business on a commercial LLM alone is a bit like publishing a company on social media. The reach is real, and the cost of entry is near zero. The lack of ownership can be costly.

The comparison is useful because the resolution is already known. Firms did not abandon social media when they understood the trade. They kept the reach and built a site alongside it.

The right tool for the job?

Social mediaOwned content platform
Noisy discourseControlled environment
Standard user experienceTailored experience
Tactical reachStrategic projects
Short-term valueLong-term value
Limited IT governanceFull IT ownership
Simple publishingContent workflow

How we approach it

Owning the whole AI stack is easier than you might think.

Benefits of an owned stack

AI adds value when you can trust it.

An owned stack is not an end in itself. It removes the one condition that caps every business case: the data the work depends on can finally be used. Customer records come into scope, and AI moves to where the work already happens.

What follows is control. The experience is designed rather than inherited, spending becomes a line a firm can forecast, and residency, audit trails and access control turn into things that can be shown to an assessor rather than described.

  • Support customers

    Customer records are in scope

  • Team workflows

    AI sits where the work happens

  • Unique experience

    Where AI appears and how it is labelled

  • Control spending

    Predictable token and operations costs

  • Governance and compliance

    Residency, audits and access control

Beyond the frontier

Six models, many reasons to switch

NameHosting and licenceBest forCapability(frontier = 100%)
Mistral FranceEU cloud · Licence not yet publishedMaximum flexibility, European throughout66%
Lumo Switzerland / EUHosted only, EU data centres · ProprietaryConfidentiality built-in, no maintenanceVaries
Qwen ChinaA single GPU · Apache 2.0Self-hosting at a lower cost59%
DeepSeek ChinaMulti-GPU server or private cloud · Open weightsWhen cost per token is the binding constraint67%
GLM ChinaMulti-GPU server or private cloud · MITStrongest self-hosted model. Few licence constraints.78%
Gemma United StatesLocal or Cloud · Google’s Gemma termsStart with zero infrastructure at a lower cost29%

Derived from Artificial Analysis Intelligence Index, retrieved 8 October 2026 · Epoch Capabilities Index, May 2026. Versions compared: Mistral Large 4, Qwen3.8 27B, DeepSeek V4.1 Flash, GLM-5.3 and Gemma 4 26B.

Stack architecture

The model is only one part of the stack.

Most conversations about AI adoption are about which model to pick. That is the one decision a firm can reverse cheaply. The layers around it decide what matters: where the keys live, who sees the logs, and which jurisdiction does the processing.

The gateway is worth taking first. It holds the keys, caps the budgets and records every request, and it sits in front of whatever a firm already uses. Each layer can come in-house as the work justifies it.

  1. Interface Open WebUI, LibreChat, custom app

    Branded user experience

  2. Gateway Otari, LiteLLM, OpenRouter

    Holds keys, budgets, routing and usage logs

  3. Inference Ollama, vLLM

    Direct queries and runs models

  4. Model Claude, Mistral, Lumo, Qwen

    Reasoning, generation and agentic AI

  5. Hosting Azure, AWS, own hardware

    Physical data storage and processing

Cost of ownership

Trust and control require some upfront investment.

Renting is cheap to start and expensive to leave. Owning reverses both. The investment goes in at the front — people, review, a test harness — and buys the freedom to change model, host or vendor without rebuilding the work on top.

That freedom compounds. The second use case costs less than the first, the third less again, and the capability stays in the business rather than on someone else's roadmap.

Where the resources go

ResourceInvestment nowBenefit tomorrow
ComputeCloud or serverLower token cost
PeopleHire or trainBuild expertise
Model selectionEvaluation processNo lock-in
SecurityExternal testingVerified compliance
SupportSupport deskFaster resolution
ImplementationDevelopment costRoadmap control

Next steps

Start your journey to full-trust AI.
We can help.

  1. 1. Discovery

    You need to ask the right questions from the right people.

    We can help uncover valuable workloads that the current AI policy rules out.

  2. 2. Select vendors

    You need independent and impartial system comparison.

    We only look after your interests, and receive no referral fees from anyone assessed.

  3. 3. Roadmap

    You need a step-by-step approach to solve the challenge.

    We will design an architecture that builds on what is already in place.

  4. 4. Handover

    Your team will need training and change management.

    We put in place an operating model that keeps running once the project is complete.

Keep a copy of the full 14-page guide.

Download PDF