AI architecture consulting and engineering for systems that must hold up in production

It depends on who you hire. Published US federal rate data puts AI consulting at about $131 to $195 per hour, and UK government rate cards show large consultancies charging £400 to £2,855 per day. Vrisic prices by outcome instead: an AI architecture review costs $8,000 to $20,000, evaluation and hardening $20,000 to $60,000, and an ongoing engineering retainer starts at $4,000 per month, fixed in writing after a free call.

An AI consultant helps a company decide where AI fits and how to build it safely. Strategy consultants focus on use cases, business cases and operating models. Engineering consultants, which is what Vrisic is, review system architecture, choose models, design evaluation and monitoring, find security gaps and often write the code that fixes them. Ask any consultant which of the two they actually deliver.

AI architecture is the design of how an AI system is put together: how requests flow, which models handle which tasks, how context is retrieved, how agents call tools, how outputs are checked, and how the whole thing is measured, secured and paid for. Good AI architecture keeps the model replaceable and makes quality, cost and risk visible to the people who own the system.

A Vrisic architecture review includes a code and infrastructure walkthrough, a baseline measurement of quality, latency and cost on a starter golden set, a threat model mapped to the OWASP Top 10 for LLM Applications, a governance gap check, and a written target architecture with a prioritised roadmap. It takes 2 to 4 weeks and ends with a readout for your engineering and leadership teams.

LLM evaluation is the practice of scoring an AI system’s outputs against defined expectations so you know whether a change made it better or worse. It combines automated checks such as schema validity, retrieval recall and exact answers, model graded scoring with an LLM as a judge, and human review of samples. Evaluation runs before every release and continuously on production traffic.

Only after you calibrate it. A judge model should score against a clear written rubric, be tested against a few hundred human labelled examples, and reach agreement with people that you have measured, not assumed. Judges have known biases, such as favouring longer answers or the first option shown, so we randomise order, use a different model family where possible and keep humans on disagreements.

AI observability means recording what every AI request did: the prompt version, retrieved documents, model and parameters, tool calls, tokens, cost, latency and final output, linked into one trace. It lets engineers explain any single answer, spot rising cost or error rates, and feed real failures back into the test set. Tools include Langfuse, LangSmith, Arize Phoenix and OpenTelemetry.

The biggest savings usually come from routing easy requests to smaller models, caching repeated prompts and retrieval results, trimming oversized context, capping agent loops, and batching work that does not need an instant answer. We measure cost per successful task before and after each change against the golden set, so savings never come from quietly lowering quality.

Multi agent AI makes sense when a process has genuinely separate stages, different permission boundaries or specialist knowledge that would overload a single prompt. It is not worth it when one agent with good tools can finish the job. Multiple agents add coordination bugs, higher token spend and harder testing, so we usually start with one agent and split only when traces show a clear reason.

It can. The EU AI Act applies to providers and deployers outside the EU when their AI system is placed on the EU market or its output is used in the EU. Transparency duties for certain AI systems apply from August 2026, while most high risk obligations were moved to December 2027 by the 2026 AI Omnibus amendment. Take legal advice on your specific case.

Large consultancies are strong on enterprise strategy, change management and programmes across many countries. Vrisic is a small senior engineering team: the people who review your architecture also write the evaluation suites, fix the code and set up the dashboards. That usually means a faster start and lower cost for technical work, and less capacity for organisation wide transformation programmes.

Usually not, unless customers or regulators ask for it. ISO/IEC 42001 is a management system standard for governing AI across an organisation, and certification is done by accredited auditors, not by builders like us. Many teams use it as a checklist to organise policies, risk assessments and monitoring. We help produce the technical evidence an auditor will ask for.