Eight steps for picking an AI consultant or agency that ships working systems, builds your team's capability and leaves you in control.
In short
Choose an AI consultant by testing how they work, not what they promise. Define the business problem first, ask for production systems similar to yours, check data engineering and security skills, run a short paid pilot with agreed success measures, and make sure your company owns the code, data, prompts and evaluation sets when the work ends.
Searching for an AI consultant (AI-konsult) or an AI agency (AI-byrå) returns hundreds of firms that all claim to deliver value with AI. Very few of them have taken a system past the demo stage and kept it running. This guide from Devies, a Gothenburg-based product and engineering consultancy, is for CEOs, CTOs and business leads who are about to buy AI help for the first time, or who bought it once and ended up with a prototype nobody uses.
What should an AI consultant actually deliver?
An AI consultant should deliver a working change in a business process, measured against a baseline, that your organisation can operate after the consultant leaves. A slide deck, a model in a notebook or a chatbot demo is not that outcome, although each can be a step towards it.
The work usually spans four disciplines, and a credible partner covers all four or tells you openly which ones you must bring:
- Problem framing: choosing the use case and the measure of success.
- Data engineering: getting the right data, with the right permissions, into a usable shape.
- Software engineering: integrating the model into real systems with logging, testing and error handling.
- Change and operation: training users, monitoring quality and handling drift and incidents.
The table below shows how the main types of supplier differ.
Type of partner
Typical strength
Typical gap
Strategy firm
Typical strength
Prioritisation, business case, executive alignment
Typical gap
Rarely builds or runs the system
AI agency or AI studio
Typical strength
Fast prototypes, model and prompt expertise
Typical gap
Integration, security and long-term operation
Product and engineering consultancy
Typical strength
Building production software and data platforms
Typical gap
May need your domain experts closely involved
Platform vendor's services team
Typical strength
Deep knowledge of one product
Typical gap
Recommends its own product by default
Freelance AI specialist
Typical strength
Narrow expertise at short notice
Typical gap
Capacity, continuity and breadth
Eight steps to choose an AI consultant
The safest way to choose is to make the selection itself a small piece of real work. Follow these steps in order.
- Write down the problem, not the technology. Describe the process, who does it today, how long it takes and what a better result looks like. If you have several ideas, rank them first; our guide on how to prioritise AI use cases gives a scoring method.
- Check your data before you check suppliers. Know where the data lives, who owns it and whether you are allowed to use it for this purpose. An AI readiness check of your data prevents paying a consultant to discover that the data is missing.
- Ask for production references, not demos. Ask which systems they have in production, how long those have run, who operates them now and what went wrong. A partner with real experience will talk about failures without being prompted.
- Test data engineering and integration skills. Most effort in a business AI project goes into data access, permissions and integration with ERP, CRM or document systems. Ask who on the proposed team does that work.
- Test security and governance knowledge. Ask how they handle prompt injection, per-user access to retrieved documents, logging and data processing agreements with model providers. Frameworks such as the NIST AI Risk Management Framework or the ISO/IEC 42001 AI management system standard are useful vocabulary; a partner does not need a certificate, but should recognise the ideas.
- Agree how quality is measured. Before any build, agree a test set of real questions or cases, the target accuracy and how a person reviews the output. Without this, every demo looks good.
- Run a short paid pilot with an exit. Four to eight weeks on one use case, with a decision point at the end: scale, change or stop. Pay for it; free pilots tend to be sales material.
- Secure ownership and knowledge transfer. The contract should state that your company owns the code, infrastructure definitions, prompts, evaluation sets and documentation, and that your staff work alongside the consultants.
Which questions reveal a weak partner?
A few questions separate partners who have shipped AI from those who have only presented it. Ask them in the first meeting and listen for specifics.
- "What happens when the model is wrong?" A good answer covers review steps, confidence thresholds and how errors are logged and fed back. A weak answer is "the model is very accurate".
- "How will our team run this without you?" Look for a plan with pairing, documentation and a handover date.
- "Which model would you use, and why?" A good partner explains trade-offs between cost, latency, data location and quality, and expects to change models later. A partner tied to one vendor will not.
- "What will this cost to run each month?" Model calls, hosting, monitoring and support are ongoing costs. Someone who has run AI in production can estimate them.
- "What would make you advise us not to do this?" The best partners say no to some projects.
Devies saw the value of building capability rather than dependency at Hedin Mobility Group, where the team coached Hedin IT's BI team in code review, logging and reusable code while working on real changes, and moved infrastructure into code so that AI agents can read and change the platform with traceable changes.
What should the contract cover?
The contract should protect your ownership, your data and your ability to leave. Read it with the same care as any software contract, and add the points specific to AI.
- Intellectual property: your company owns deliverables, including prompts, configuration, evaluation sets and fine-tuned weights where applicable.
- Data processing: the consultant and every model provider they use are covered by data processing agreements, with stated data location and retention.
- Use of your data: your data is not used to train the consultant's or a provider's general models without explicit consent.
- Regulation: roles under the EU AI Act are clear. As of October 2026, the AI Act's AI literacy obligations have applied since 2 February 2025, and high-risk obligations for Annex III use cases apply from 2 December 2027 after the AI Omnibus amendment. A partner should help you determine whether you are a provider or deployer.
- Exit: a handover plan, documentation standard and notice period, so you can move the work in-house or to another supplier.
When should you build in-house instead?
Build in-house when AI is core to your product, when you will run many use cases over several years, or when the domain knowledge is so specific that outsiders slow you down. Most mid-sized companies end up with a mix: a small internal team that owns priorities, data and operation, supported by a partner for architecture, peak capacity and skills the team does not yet have.
A useful rule is to buy the first one or two projects as joint work and use them to hire and train. If after a year you still cannot change a prompt or retrain a model without the consultant, the partnership has failed at knowledge transfer, whatever the system delivers.
Keep reading
From our work
Client Case: Stena LineStena Line modernised its Azure data platform - improving Power BI reliability, reducing pipeline runtimes and scaling with medallion architecture.Read the caseDevies
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 does an AI consultant do?
An AI consultant helps a company choose, build and run AI solutions for specific business problems. In practice that covers selecting use cases, preparing data, choosing models, integrating them into existing systems, measuring output quality and training staff. The best consultants also leave the company able to operate and improve the solution without them, through documentation, pairing and handover.
What is the difference between an AI consultant and an AI agency?
The terms overlap. An AI agency (AI-byrå) is often a small studio focused on prototypes, chatbots and generative AI content, sometimes with a marketing angle. An AI consultant or consultancy more often works on business systems, data and integration over a longer period. What matters is not the label but whether the team has taken systems into production and can show how they run today.
How long should an AI pilot with a consultant take?
A focused pilot on one use case usually takes four to eight weeks, including data access, a working version used by real people and an evaluation against agreed measures. Longer pilots often indicate unclear scope. Shorter ones rarely test integration and permissions, which is where most problems appear. End every pilot with a decision to scale, change or stop.
Should we choose an AI partner that is certified?
Certification is a weak signal for AI delivery. Standards such as ISO/IEC 42001 describe an AI management system and are useful for governance, but few delivery teams hold it, and holding it does not prove they can build. Ask instead for production references, a named team, a test-based approach to quality and clear answers on security, data processing and ownership.
