Most companies no longer need convincing that AI is useful. They need help getting a system past the pilot stage and into production, where it has to be correct, affordable, permissioned and auditable. That gap is the work we do.
Gartner projects that up to 40% of enterprise applications will include task-specific AI agents during 2026, up from under 5% in 2025. The same firm expects more than 40% of agentic AI projects to be cancelled by 2027. Both are true at once, and together they describe the market accurately: nearly everyone is experimenting, and only a minority ship.
What we build
AI agents — systems that plan a task, call your real tools and complete multi-step work, with scoped permissions and human approval where the consequences warrant it.
Retrieval and knowledge assistants — grounded answers drawn from your own documents, with citations and access rules applied at retrieval time.
Voice and multimodal — real-time speech agents, and document or vision models that read forms as images rather than discarding their layout.
Product copilots — AI features embedded in software you already run, built around its data model, permissions and latency budget.
Fine-tuned and self-hosted models — smaller, cheaper, more consistent models for narrow tasks, including deployments that never send data to a third party.
Governance and compliance engineering — the inventory, logging, oversight and documentation that regulated AI now has to evidence.
Why projects stall, and what we do about it
Failed AI projects rarely fail because the model was not clever enough. They fail on the surrounding engineering: no way to measure whether output is actually good, no permission model, no cost ceiling, no plan for the day the provider changes a price or deprecates an endpoint, and no clear owner when it gets something wrong.
We treat those as the primary work rather than an afterthought. Every system we ship carries an evaluation set built from your real tasks, tracing on every call, spend instrumentation per user and per feature, and a provider abstraction so one vendor's outage or price change does not become your incident.
Computer vision and AR
This practice also covers our longer-standing computer vision and augmented reality work, including ARKit and ARCore. That field has consolidated sharply — several consumer AR authoring platforms have been retired by their owners, and the individual pages in this section say plainly where a platform is no longer a safe place to invest. Face and biometric work now also sits inside the EU AI Act's high-risk category, which changes what a responsible deployment looks like.
When AI is the wrong answer
A deterministic rule, a database query or an ordinary integration is often faster, cheaper and more reliable than a model, and we will say so before taking the work. Language models are a poor fit where the task has one provably correct answer that simpler code can compute, where no one can define what a good output looks like, or where the cost of an occasional confident error exceeds the value of automating the task at all. Establishing that early is cheaper than discovering it after a build.
Regulation is now part of the build
The EU AI Act's obligations for high-risk AI systems became enforceable on 2 August 2026, with penalties reaching €15 million or 3% of global annual turnover, whichever is higher. Its Annex III categories reach a great deal of ordinary business software, including recruitment, credit scoring, insurance pricing and access to essential services. We build the technical evidence layer this requires and work alongside your legal advisers, who remain the ones to determine how the rules apply to you.
We front-load the questions that decide whether a system is worth building, so the expensive work only starts once the answer is known.
01
Scope one workflow
We pick a single, narrowly defined task rather than a broad assistant. Narrow systems are testable, their failure modes are comprehensible, and they reach production.
02
Prove it on your data
We prototype against your real documents, audio or records. Performance on your genuine inputs diverges sharply from performance on clean samples, and that is where most estimates go wrong.
03
Build the harness with the system
Evaluation sets, tracing, permission scoping and spend limits are built alongside the feature, not retrofitted after something goes wrong in front of a customer.
04
Hand over ownership
You get the system, the evaluation harness and the documentation needed to operate and extend it. We would rather your team could change it without us.
Frequently asked questions
How do we know whether AI is the right solution for our problem?
Start by asking whether anyone can define what a good output looks like. If the task has one provably correct answer that ordinary code could compute, conventional software will be cheaper, faster and more reliable. AI earns its place where inputs are messy or unstructured, where judgement is involved, or where the volume makes manual handling impractical. We assess this before proposing a build, and we will tell you when the answer is no.
Do we have to send our data to a third-party AI provider?
No. Where regulation, contract or risk appetite requires it, we deploy open-weight models inside your own cloud account or on-premise environment so no prompt or document leaves your network. This costs more to operate than a hosted API, so we usually recommend it where there is a real constraint rather than by default.
How long before we see something working?
A narrowly scoped prototype on your own data typically takes a couple of weeks and is deliberately cheap, because its purpose is to answer whether the approach works at all. A production deployment with evaluation, monitoring and permissions generally runs one to three months, driven mostly by how many existing systems it must integrate with.
What stops an AI system from giving wrong or made-up answers?
Engineering rather than prompting. Retrieval grounds answers in your actual documents and cites them, deterministic code validates model output before anything is written, evaluation suites catch regressions when a prompt or model changes, and the system is designed to say it does not know rather than produce a confident guess. A system that admits uncertainty is far more useful in a business setting than one that is always fluent.
Does the EU AI Act apply to us?
It may, including for companies outside the European Union, because the Act can reach providers and deployers whose AI system output is used within the EU. Obligations for high-risk systems listed in Annex III became enforceable on 2 August 2026. Whether your specific system is in scope is a legal determination for your counsel; we build the inventory, logging, oversight and documentation that meeting those obligations depends on, and work alongside your legal team rather than in place of them.
