AI consulting for companies: how to start and what you get

AI consulting is a collaboration in which you and outside experts review your processes and data, choose the uses of artificial intelligence that pay off and come away with a workable plan: a pilot, a cost estimate and rules for how AI is used. It makes sense when your company has repetitive work and data in digital form, and on your side it needs one accountable person, access to your systems and a willingness to try the first working version in day-to-day work.

When does AI consulting make sense, and when does it not?

Bringing artificial intelligence into a business makes most sense where people do the same work with data every day: answering similar enquiries, retyping orders from emails into the ERP, putting reports together or looking for information across several systems. If you recognise this kind of work in several departments but it is not clear where to start, AI consulting helps you decide what comes first before you spend money on development.

It makes less sense when there is no concrete problem, only a wish to “bring in AI”, when the key data exists only on paper, or when nobody in the company has time to run the project. In that case the better first step is to name a concrete problem, put the data in order or decide who will lead the project.

What exactly do you get?

A good engagement ends with concrete outputs you can base decisions on:

• A process review. A record of how the work is actually done today and where delays and errors arise, gathered in conversations with the people who do it.
• A prioritised list of use cases. Each possible use case is rated for benefit, effort and risk, with a clear proposal of what comes first and what is not worth doing.
• A data readiness assessment. Data readiness means that the data needed for a particular task is available and accurate enough, and that you are allowed to use it for that purpose.
• Build or buy. For each use case you get a reasoned recommendation: an off-the-shelf tool, a custom solution or a combination of the two.
• A pilot plan. A pilot is a first working version that shows, on real data and with a small group of users, whether the idea works in practice; the plan sets out its scope, users and success criteria.
• Governance rules. Governance rules set out who is responsible for the system, which tools are allowed, what a person must check and how errors are logged and corrected.

How can you tell a useful engagement from a slide deck?

A simple test: could someone in your company start work tomorrow on the basis of what you received? You can tell a useful engagement by these signs:

• the consultants talk to the people who do the work, not just to management;
• every recommendation has an accountable person, the data it needs and a measurable success criterion;
• you also get a list of what not to do, with the reasons;
• the cost estimate includes maintenance and the cost of using AI models, not just development.

You can spot a slide deck by its general trends instead of your processes, its recommendations made without checking the data and integrations, and a solution that happens to be exactly what the adviser sells.

How do you choose a partner?

Whether you are looking for an AI agency, an independent consultant or a software company, the following questions reveal more than a list of logos on their website:

• Who owns the code and the data? The contract should state who owns the source code, the data and the system settings, and how you get them back if you end the relationship.
• Where does the system run? Ask whose servers it runs on and in which country, and which outside providers, including AI model providers, see your data.
• References you can call. Ask for a contact at one of their clients and find out how the system performed after go-live, not only during development.
• How are changes charged? Agree in advance how change requests are estimated, approved and billed.
• Who maintains the system after go-live? Someone has to monitor it, respond to faults and adapt the system when the AI models or your processes change.
• Does the solution work well in Croatian, or in the language your customers use? Ask for a demonstration on your own examples, with the technical terms and product names your customers actually use.

No approach is best for everyone; the difference lies in what you get and what you risk:

What drives the cost?

The cost of an AI project is not set by the technology itself but by these five items:

• Scope: how many processes, users and channels (web, email, phone) the project covers.
• Integrations: whether the system has to exchange data with your ERP, CRM or telephony, and whether there is a documented application programming interface (API) for it.
• Data quality: how much work it takes to make the data accurate, complete and accessible.
• Languages: each language has to be tested separately, and a voice channel needs speech recognition on top of text understanding.
• Maintenance: monitoring, updates, support, servers and the cost of using AI models, which grows with the number of queries.

Data protection requirements, such as an impact assessment or specific retention periods, can add work to any of these items. When you compare quotes, ask for each item to be listed separately: a quote without maintenance is not cheaper, just incomplete. Our AI implementation cost page describes how we link the price to the scope and which engagement models we offer.

How long it takes: from the first conversation to go-live

1. First conversation. Goals, processes, systems and constraints; at the end you know whether it makes sense to continue.
2. Scope and estimate. A written description of what will and will not be done, a cost estimate and a phased plan.
3. First working version. In our process it arrives 4–8 weeks after the scope is agreed, and it runs on real data with a small group of users.
4. Go-live. All intended users work with the system, supported by training, monitoring and maintenance.

Factors outside the technology can also stretch the timeline: waiting for access to systems and data, decisions nobody has the authority to make, and legal reviews that only start once the system is finished.

What should you check in the EU AI Act?

The EU Artificial Intelligence Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024 and applies in stages. First establish your role under Article 3: a “provider” is whoever 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” is whoever uses an AI system in the course of its business. A company that commissions a custom solution and uses it under its own name can therefore also have a provider’s obligations.

• Since 2 February 2025 the prohibitions in Article 5 have applied, for example the ban on manipulative techniques that materially distort a person’s behaviour and cause significant harm, and the ban on inferring people’s emotions in the workplace, except for medical or safety reasons. The AI literacy obligation in Article 4 has applied since the same date.
• Since 27 July 2026 an amendment to the Act, Regulation (EU) 2026/1744, has been in force. Under it, providers and deployers take measures to support the development of their staff’s AI literacy, without having to guarantee a particular level of knowledge for each individual. The same amendment adds new prohibited practices, which apply from 2 December 2026.
• Since 2 August 2026 the transparency obligations in Article 50 have applied: people must be informed, at the latest at the first interaction, that they are communicating with an AI system, unless this is obvious, and deep fakes must be labelled as artificially generated.
• From 2 December 2027 the obligations for high-risk systems listed in Annex III apply, for example systems used to select job candidates, to make decisions about employment or to assess creditworthiness; for AI in regulated products covered by Annex I, the date is 2 August 2028.

The European Commission publishes an overview of these dates on its page about the regulatory framework for AI, and its Compliance Checker, currently in beta and not available in Croatian, can help with a first orientation. At this stage, write down the role and purpose of every AI system you use or plan to use, and check whether it falls into any of the categories above.

GDPR: the basics for AI projects

The General Data Protection Regulation, Regulation (EU) 2016/679, applies as soon as an AI system processes personal data, whether during development, testing or everyday use. The Croatian data protection authority, AZOP, advises on its page on AI and personal data protection (in Croatian) that data protection should be considered from the planning and procurement stage, not only after the system goes live. For an AI project this means:

• Purpose and legal basis (Articles 5 and 6): for each use case, write down why and on what basis you process the data, and use only the data you need.
• Informing people (Articles 13 and 14): customers and employees must know which of their data is processed and for what purpose.
• Contract with the processor (Article 28): if a partner or tool provider processes data on your behalf, the relationship must be governed by a contract.
• Impact assessment (Article 35): it is mandatory where processing, in particular using new technologies, is likely to result in a high risk to the rights and freedoms of individuals.
• Automated decisions (Article 22): special rules apply to decisions with legal or similarly significant effects that are based solely on automated processing, and in its recommendation on applying Article 22 (in Croatian) AZOP stresses that human involvement must be real, not merely formal.
• Transfers to third countries (Chapter V): check whether data is processed outside the European Economic Area, including by sub-processors.

For publicly available chatbots and similar tools, AZOP’s recommendation on the safe use of large language models (in Croatian) advises an internal usage policy, a list of permitted tools, a ban on entering personal data without prior approval, business rather than private user accounts, and involving the data protection officer before the tools are introduced.

The first thirty days: what to do

1. Appoint a project owner: one person with the authority and the time to make decisions and bring people together.
2. List the AI tools your employees already use and, until you have rules in place, ban entering personal and confidential data into publicly available tools.
3. Pick two or three tasks that take the most time and write down how they are done today, how long they take and where errors occur. This is your baseline for later comparison.
4. Check the data for those tasks: where it is, who has access to it, how accurate it is and whether it contains personal data.
5. Establish your role under the AI Act and check whether any use case falls into the high-risk areas.
6. Talk to at least two potential partners, call their references and ask each of them for a written scope proposal with success criteria for the pilot.
7. Decide for each use case: an off-the-shelf tool, custom development or a combination of the two.
8. Agree the pilot: scope, users, success criteria, a review date and the data processing agreement.
9. Run a short training session for employees on what AI tools do well, where they make mistakes and what must not be entered into them.

From our own work

Divine Solutions develops and maintains its own AI products, which are in everyday use: CallVista, a platform for AI voice agents that handle inbound and outbound calls in Croatian and other languages for call centres; AdReady, for brand management and automatic banner variants; and Business Booster, a business platform. For Instar informatika, a Croatian IT equipment retailer, we connected their ERP system, Luceed by Tomsoft, to an AI layer that answers business questions asked in plain language, and set up automatic sales and stock reports by brand. A layer like this is only as good as the data in the ERP, and the reports have to respect the same access rights as the ERP itself.

What does a first conversation with Divine Solutions cover?

The first conversation is free. In it we go through the tasks that take up most of your time, the systems you use (ERP, CRM, email, telephony), the state of your data and constraints such as data protection or deadlines. It ends with one of three decisions: start with a pilot, prepare the data first, or hold off on the project for now. If we go ahead, you get a written scope, a cost estimate and a phased plan, you follow progress in our client portal, and the solutions we build and maintain are hosted on servers in the EU. Our AI strategy and consulting for companies page describes what the whole engagement looks like.