Frequently asked questions

A free consultation call. We ask about your goals, current systems and constraints, and tell you honestly whether AI is the right lever. If it is, we propose a fixed-price discovery sprint of two to three weeks that ends with a ranked list of opportunities, an architecture sketch, a delivery plan and a costed proposal you can take to your leadership team.

We score candidate use cases on business value, data readiness, technical risk and time to production, then recommend one that can prove value within a quarter. Typical first projects are support automation, document processing, a knowledge assistant or a forecasting model, because they use data you already have and their impact is easy to measure.

No, but it helps. During discovery we audit what data exists, where it lives and how clean it is. Many projects begin with unstructured data such as documents, tickets or call recordings, which need less preparation than people expect. Where data is missing we design a way to capture it as part of the first release.

A description of the process or product area you want to improve, rough volumes (tickets per month, documents per week, calls per day), the systems involved and any constraints on data or hosting. You do not need a specification; we help shape the problem together.

Yes. Our consulting and strategy work covers opportunity mapping across departments, feasibility and ROI analysis, architecture and governance recommendations, and a phased roadmap. Many clients start with a roadmap and then move into delivery of the first one or two priorities with us.

Discovery sprints are fixed price. Delivery projects are scoped after discovery and priced either as a fixed-price milestone plan or as a monthly team engagement, depending on how well-defined the work is. We share indicative ranges early in the conversation so you can decide quickly whether it fits your budget.

Three: fixed-scope projects with milestones, a dedicated team or specialists embedded with your engineers on a monthly basis, and advisory or fractional AI leadership. Many clients combine them, for example a fixed-price first release followed by an ongoing team for iteration and support.

A discovery sprint is a fixed, modest fee. A focused first production use case, such as a support assistant or document-processing pipeline, is usually scoped in weeks rather than months. We would rather give you a realistic range on the first call than a number here that does not fit your situation.

Yes, in two forms: infrastructure and model usage billed directly to your cloud and API accounts, and optional support or improvement retainers with us. We estimate running costs during discovery and design for cost control, for example by routing simple requests to smaller models.

We guarantee the things we can control: fixed prices for discovery, weekly working demos, code and documentation delivered in your repositories, and no vendor lock-in. We do not promise specific business results before we have seen your data, but we agree measurable success criteria with you before delivery starts.

In your own cloud accounts and repositories. We work inside environments you control, with least-privilege access that you can revoke at any time. Any temporary copies used for development are documented and deleted at hand-over. A data-flow diagram is part of every project’s documentation.

No. When we use hosted models we use enterprise or API tiers whose terms exclude training on your data, and we can add a zero-retention configuration where offered. Where that is not acceptable, we deploy open-source models inside your environment so nothing leaves it.

Yes. We regularly deploy open-source LLMs, vision and speech models on client GPUs, private cloud tenancies or air-gapped networks. This is common for healthcare, legal, finance and utility clients, and we help size the hardware and estimate the running cost against hosted alternatives.

We sign data-processing agreements, minimise personal data, apply role-based access and encryption, log access and model calls, and design human oversight for sensitive decisions. We are not a law firm, so we work with your compliance team to map their requirements onto the architecture and provide the technical evidence they need.

Not yet. We follow the underlying practices (access control, encryption, secure development, incident response) and can supply our security overview and complete customer security questionnaires. If certification is a hard requirement we will tell you upfront and can discuss deploying entirely within your certified environment.

Discovery takes two to three weeks. A first production release usually follows in six to twelve weeks depending on integrations and validation needs. We work in weekly iterations with a demo at the end of each, so you see progress continuously and can adjust priorities rather than waiting for a big reveal.

Asynchronously by default (a shared channel in Slack or Teams, written weekly updates and a project board you can see at any time), plus scheduled calls in overlapping hours. Most of our clients are in Europe, North America, the Middle East and Australia, so this rhythm is well practised.

We build an evaluation set from real examples, define acceptance metrics with you, and run automated and human evaluation on every change. Guardrails, edge cases and failure modes are tested explicitly. For customer-facing systems we run a limited pilot with real users before full rollout.

We find out early, because feasibility is tested in discovery and quality is measured weekly. If a use case is not viable we say so and recommend an alternative rather than pushing on. During delivery, evaluation results guide whether we adjust prompts, retrieval, models or data before scaling up.

Yes, and we encourage it. We can lead the AI work while your engineers own the surrounding product, embed a specialist in your team, or advise a team that does the building. Pairing sessions and documentation are part of hand-over so your team can maintain and extend the system.

We are model-agnostic. We use OpenAI, Anthropic Claude and Google Gemini through APIs, and open-source models from Hugging Face such as Llama, Mistral, Qwen and DeepSeek when cost, privacy or customisation favour them. The choice is made per use case on quality, cost, latency and data-handling requirements, and can change as models improve.

Start with prompting and retrieval; they cover most needs and are cheap to change. Fine-tune when you need very consistent formats or style, a specialised domain the base model handles poorly, or a smaller and cheaper model to match a larger one on a narrow task. Often the answer is a combination.

Hosted frontier models are the fastest way to prove value and are usually the right start. Open-source models make sense when data must stay in your environment, when volumes make API costs significant, or when you want full control over model behaviour. We design systems so the model layer can be swapped later.

Python with PyTorch, scikit-learn and Hugging Face for modelling; LangChain, LangGraph and custom orchestration for agents; Pinecone, Weaviate, Qdrant, Milvus or pgvector for retrieval; FastAPI and Docker for services; AWS, Azure and Google Cloud for hosting; and n8n, Make or Zapier for automation glue. We fit into your stack rather than imposing ours.

Yes. Most of our work is integration: connecting models to CRMs, helpdesks, ERPs, databases and internal tools through APIs, middleware and event pipelines. We build with your authentication, logging and deployment practices so AI features behave like any other part of your system.

A hand-over period with documentation, runbooks and pairing sessions is included in every project. After that you can choose a monthly support and improvement retainer covering monitoring, model updates, evaluation runs and new features, or manage the system in-house with us on call.

We instrument quality, latency, cost and usage, track model drift and unanswered or escalated queries, and set alerts for anomalies. Dashboards are in your monitoring stack. Regular evaluation runs on fresh samples show whether quality is holding as data and models change.

It depends on how fast your data changes. Forecasting and fraud models are typically retrained on a schedule or when drift is detected. LLM-based systems change less through retraining and more through updated content, prompts and occasional model upgrades, which we test against your evaluation set first.

Yes. Everything is built in your repositories and cloud accounts using mainstream tools, with documentation written for your team. Clients who take over in-house often keep a light advisory arrangement for model changes, but there is no technical or contractual dependency on us.