The eight steps between a promising demo and a system people rely on, and where most AI projects stall on the way.
In short
AI development is the work of turning a business problem into an AI solution that runs reliably in daily operations. It moves through eight steps: frame the problem, check the data, choose the simplest approach, build an evaluation set, prototype, harden for production, classify risk and govern, then deploy gradually and monitor. Most projects stall between prototype and production.
A working AI demo can be built in days. A system that a team trusts with real customers, real money or real decisions takes considerably more, and the extra work is mostly not about the model. This guide from Devies, a Gothenburg-based product and engineering consultancy, is for CTOs, product owners and business leaders who are planning AI development (ai utveckling) and want to know what the path to production really involves, and which decisions to make early.
What is AI development?
AI development is the process of designing, building, testing and operating software in which a learned model, rather than fixed rules alone, produces predictions, content, recommendations or decisions. The model is usually a small part of the work; data, integration, evaluation, security and operations are the rest.
Today most business AI solutions (ai lösningar) fall into one of three families: classic machine learning models trained on your own data, such as demand forecasts or churn scores; applications built on large language models through an API, such as assistants, document extraction or answers grounded in company documents; and agents that take actions in other systems on a user's behalf. The AI lifecycle is similar for all three, but the risks and the testing differ.
AI development is not the same as buying an AI feature. When the capability already exists in a product you use, configuring and governing it is usually cheaper than building your own. Build when the problem is specific to your business and the value justifies owning the solution.
Which approach fits your problem?
The right approach is the simplest one that meets the quality bar, so compare options before choosing a model. A rule or a report sometimes solves the problem without AI at all.
Approach
Fits when
What you build
Main risk
Rules and automation
Fits when
The logic is known and stable
What you build
Workflow or code
Main risk
Brittle when cases vary
Classic machine learning
Fits when
You have labelled history and a clear target, such as a forecast or a score
What you build
Training pipeline, model, monitoring
Main risk
Data drift over time
Language model via API
Fits when
The input is text or documents and the output can be checked
What you build
Prompts, integration, evaluation set
Main risk
Confident wrong answers
Retrieval over company data (RAG)
Fits when
Answers must come from your own documents
What you build
Search index, access control, evaluation
Main risk
Outdated or unauthorised sources
Agent that takes actions
Fits when
A task spans several systems and steps
What you build
Tools, permissions, human approval points
Main risk
Actions with real consequences
Buy a product with AI built in
Fits when
The need is common across companies
What you build
Configuration and governance
Main risk
Lock-in and limited control
Choose before you build, and write down why. That record is also useful later for risk classification.
How does an AI solution go from idea to production?
An AI solution goes from idea to production through eight steps, and each step produces something the next one needs. The order matters more than the speed: teams that skip step 4 usually discover in step 6 that they cannot tell whether the system works.
- Frame the problem and the value. Name the process, the user, the decision the AI supports and how success will be measured in business terms. Our guide on how to prioritise AI use cases helps choose which idea to take forward.
- Check the data and the rights to use it. Confirm that the data exists, is good enough, can legally be used for this purpose and can be reached by the system in production, not just exported once for a demo.
- Choose the simplest approach that could work. Use the table above. Start with an existing model or service where possible; train your own only when needed.
- Build an evaluation set before the solution. Collect realistic examples with the expected answer or a clear judgement rule, including difficult and adversarial cases. This is your definition of "good enough", agreed with the people who will use the system.
- Prototype against the evaluation set. Build quickly and measure. Our comparison of a prototype, an MVP and a pilot explains what each first version should prove.
- Harden for production. Add access control, logging, error handling, cost limits, fallbacks and a way for people to review or correct outputs. Integrate with the systems where the work actually happens.
- Classify the risk and set up governance. Decide whether you are a provider or a deployer under the EU AI Act, which risk tier applies, who owns the system and what human oversight looks like.
- Deploy gradually, monitor and improve. Start with a limited group, track quality, cost and usage, and feed real errors back into the evaluation set.
What changes when an AI solution goes into production?
In production, an AI solution has to be secure, measurable and accountable, which is where most of the remaining effort goes and where many pilots stop.
Security. AI systems that read user input or documents can be manipulated. The OWASP Top 10 for LLM Applications (2025 edition) lists prompt injection first, followed by sensitive information disclosure, and includes excessive agency, the risk of giving a model more permissions than the task needs. Give the system only the access its users already have, and require human approval for actions that are hard to undo.
Quality over time. Models and data change. A forecast trained on last year's patterns drifts; a language model provider updates its model. Rerun the evaluation set on every change and on a schedule. For machine learning models, our explainer on MLOps covers versioning, retraining and monitoring.
Accountability and regulation. Under the EU AI Act, Regulation (EU) 2024/1689, a provider is an organisation that develops an AI system, or has one developed, and places it on the market or puts it into service under its own name, while a deployer uses an AI system under its own authority. Building an AI solution for your own operations can therefore make you a provider, and if the use falls in a high-risk area the obligations are substantial. The NIST AI Risk Management Framework, a voluntary framework released in January 2023, organises the work into four functions: govern, map, measure and manage. ISO/IEC 42001 sets out requirements for an AI management system for organisations that want a certifiable structure.
Devies used an AI-first development process with SEVR Insights. According to the case, the platform connects to a firm's accounting data and generates budgets, forecasts, liquidity analyses and management reports, and it was built in four months with multiple agents working in parallel across design, development and testing. The case also notes that the hard problem was not making AI generate a budget, but making it generate one a consultant could put their name on, which is exactly what steps 4 and 6 address.
How long does AI development take, and what drives the effort?
AI development effort is driven far more by data access, integration and the quality bar than by the model itself, so two projects with the same model can differ by months.
The main drivers are:
- Data readiness. Clean, accessible data with clear rights shortens everything; scattered or undocumented data lengthens everything.
- The cost of a wrong answer. An internal drafting aid can tolerate errors a person will catch. A system that affects customers, money or safety needs a stricter evaluation, more review and more governance.
- Integration depth. Reading from one system is simple; writing back into several, with permissions and audit trails, is not.
- Risk class. High-risk uses under the AI Act require documentation, logging and oversight that take time to set up.
Plan in stages with a decision point after the prototype: continue, change approach or stop. Stopping a weak idea after a few weeks is a good result, not a failure.
Devies
Want a second opinion?
Devies, a Gothenburg-based product and engineering consultancy, works on questions like this with Nordic product and IT teams. Tell us where you are and we will suggest a sensible first step.
Frequently asked questions
What is the AI development lifecycle?
The AI development lifecycle is the sequence of stages an AI solution passes through: problem framing, data preparation, choice of approach, evaluation design, prototyping, production engineering, governance and deployment, followed by continuous monitoring and improvement. Unlike traditional software, the lifecycle never really ends, because model quality depends on data and conditions that change, so evaluation and monitoring continue for as long as the system runs.
Should we build our own AI model?
Usually not at first. Most business needs can be met with existing models through an API or with AI features in products you already use, combined with your own data and integrations. Training your own model makes sense when you have a specific prediction task, enough labelled history and a clear advantage from owning the model. Start with the simplest approach that passes your evaluation set.
Why do so many AI pilots never reach production?
Pilots often stall because they were built to impress rather than to be measured. Common gaps are no agreed evaluation set, data that was exported by hand for the demo, no plan for security and access control, unclear ownership and no budget for running costs. Defining success, data access and a production owner before the pilot starts avoids most of these problems.
Does the EU AI Act apply to AI we build for internal use?
It can. The AI Act applies to AI systems put into service in the EU, including those an organisation develops for its own use, and the developer may then count as a provider. Most internal tools fall into minimal-risk categories with light duties, but uses such as recruitment, worker management or credit assessment can be high-risk. Classify each use early and check it with legal counsel.










