Practice
AI Consulting and Integration
Murarum's AI consulting and integration service designs and ships AI agents, copilots and LLM workflows into the production systems of companies with 10 to 250 people, with evaluation, guardrails and a named owner. It starts with a diagnosis that sets the scope and timeline, and builds only where a process is worth it.
When this fits
You recognise this
Pilots that never reach production
A demo works in a meeting, then stalls. MIT NANDA's State of AI in Business 2025 report found that 95% of generative AI pilots deliver no measurable profit and loss impact. The cause is rarely the model. It is brittle workflows, no evaluation and no owner.
AI demanded before a process exists
The board wants AI on the roadmap. Nobody has named the repetitive process it should take over, the data it needs or the decision it is allowed to make. We start there, and we say when the answer is no.
Data and security questions left open
Which data leaves the company, where the model runs, who sees the outputs, what GDPR requires. Without written answers the project stops at legal or at IT. We answer them in the design phase, in writing.
Specific solutions
When the problem is this specific.
How it runs
Diagnose, design, build, scale
- 01
Diagnose
We map the processes where AI could earn its place and rank them by volume, cost of error and data readiness. We check what data exists, where it lives and whether it can be used. Some candidates are dropped here; that is the point.
A short list of use cases with a written case for each, or a written case for none.
- 02
Design
For the chosen use case we design the workflow, the points where a person approves or overrides, the guardrails on inputs and outputs, and the evaluation criteria. Data flows, retention and access are documented for your DPO.
Workflow specification, data flow document and an agreed evaluation set.
- 03
Build
We integrate the model with your existing systems, ERP, CRM, email, document stores, and build the human approval points into the tools your team already uses. Every version is run against the evaluation set before it is released.
A production system, logged and tested, used by the people who own the process.
- 04
Scale
We add monitoring for accuracy drift, latency and cost, set budget limits with alerts, and extend the workflow to the next process only when the first one holds. Then we hand off: documentation, runbooks and a named owner on your side.
Monitoring in place, cost under control, your team running the system without us.
What we deliver
Scope, not slides
- AI readiness assessment: use cases ranked by value and data readiness
- Workflow design with human approval points
- Evaluation set and accuracy criteria agreed before build
- Guardrails: input and output checks, permissions, logging
- Integration with existing systems: ERP, CRM, email, documents
- Production deployment with monitoring and cost limits
- Data flow and retention documentation for GDPR review
- Documentation, runbooks and hand-off to your team
- Claude
- OpenAI
- Gemini
- LangChain
- Python
- Node.js
- TypeScript
- PostgreSQL
- Supabase
- n8n
- Vercel
Stack
Proven models,
tested on your own examples
Claude, OpenAI or Gemini, chosen per task and tested against your own examples, with evals around everything and your data kept in your own database.
Proof
No named AI case yet. We will not invent one.
We have not published a named AI integration case study yet, and we will not invent one or turn an unnamed pilot into a percentage. Trust here rests on the engineering record: WHO, Sunday, ExproSov and Georgescu & Gonț were shipped to production by the same people who build our AI work. The first AI case will appear on this page when a client clears it, with its numbers and dates.
Until then, the data we publish is our annual resource, AI adoption in Romania: 2026 statistics, compiled from Eurostat and the European Commission.
Questions about AI consulting and integration
Which processes are a good fit for AI?
Repetitive processes with steady volume, written inputs and a clear definition of a correct result: sorting inbound email, extracting data from documents, drafting replies from a knowledge base, classifying tickets, preparing routine reports. Processes that run rarely, need a person's judgment every time or have no usable data are not a fit, and we say so during Diagnose.
How do you evaluate accuracy?
Before we build, we agree an evaluation set: real examples from your process with the correct answer for each. Every change to the prompt, model or workflow is run against that set, and the results are reported as numbers, not impressions. The threshold for going live is set with you, per use case, and it stays in the monitoring after launch.
How do you handle data privacy and GDPR?
In the Design phase we decide which data the model may see, where it is processed (EU region or your own infrastructure), how long it is retained and who can read the outputs. Personal data is minimised or masked before it reaches a model. The decisions are written into a data flow document your DPO can review. We do not train models on your data, and we choose providers whose terms do not either.
What does it cost to run models in production?
Two costs: model usage, billed by the provider per token or per request, and the surrounding system: hosting, monitoring, logging. During Design we estimate the cost per task and per month from your real volumes, set a budget limit with alerts, and choose the smallest model that passes the evaluation set. Project fees are scoped after Diagnose; we do not publish prices.
How we priceWhat happens when the model is wrong?
It will be, some of the time, so the workflow is designed for it. Low-confidence or high-impact outputs route to a person for approval. Every output is logged with its inputs, so an error can be traced. Guardrails block outputs outside the allowed format or scope. When a pattern of errors appears, it becomes a new evaluation case and the workflow is corrected.
You have no named AI case. Why trust you with AI?
Because the risk in AI work is rarely the model. It is the integration, the data and what happens in production, and that is the same engineering our case studies show, done by the same people. For AI, we agree an evaluation set before we build, ship only what passes it, and set a budget limit with alerts in production. If the diagnosis shows AI does not belong in a process, we say so and do not build it.
Further reading
Read more on this.
Related services
Where AI work usually starts
By industry
See what it looks like in your industry.
Tell us the process. We will tell you if AI belongs in it.
Describe the process, the volume and the systems it touches. We will say plainly whether AI earns its place there, and what it would take to run it in production.