AI Consulting & Strategy

Almost every company has more AI ideas than capacity to carry them out. At the same time the market is hard to read. Ready-made assistants, features in existing software and custom builds compete for the same tasks, and vendor promises are almost impossible to compare from the outside. Many initiatives fail less because of the technology than because of unclear value, missing data or missing ownership.

We help decide where AI pays off, in which form and in what order. The result is a rated list of use cases, an assessment of data readiness and a roadmap that can be turned into projects. Nothing is built at this stage. Whether an approach reaches the required quality on your own data is for a proof of concept to find out, and how a company works with data as a whole is for the data strategy. We also say openly where AI is not the right answer.

Finding use cases with the business teams

Good ideas come from where the work happens. We talk to the business teams about workflows, not technology. Where do waiting times arise? Where do people keep gathering information by hand? Where are large amounts of text, documents or images processed, and where do decisions depend on forecasts? We usually end up with requests and tickets, documents, knowledge search, reports and demand forecasts.

Each idea is described briefly: task, people involved, current effort, data needed, expected impact. Then we rate it on five criteria. Value asks how much time, money or quality is at stake and how the effect can be measured. Feasibility asks whether the task can be solved at the required quality with today’s models, and data readiness whether the data is available, accessible, current and usable with the right permissions. Risk describes what happens when things go wrong and how easily errors are noticed, whether in business, legal or organizational terms. Running cost, finally, covers licenses, usage, infrastructure, maintenance and support.

We weight the criteria together, because a small time saving at high volume can be worth more than a big effect on a rare task. The obligations that arise from the EU AI Act are covered in the topic of the same name, and questions on hosting and data flows in the topic on GDPR-compliant AI.

Where AI helps and where it does not

AI suits tasks that involve language, documents or many similar cases, where a suggestion that a person reviews is enough: drafts of texts and replies, summaries, sorting requests, extracting data from documents, searching knowledge bases or forecasts based on existing time series. Less suitable are tasks with fixed rules that can be automated more reliably and cheaply without AI, exact calculations that must always be reproducible, rare tasks with small volume, and decisions with high potential damage where errors go unnoticed.

A common finding is also that the process itself is unclear. AI would then automate a procedure nobody can describe precisely. In that case it makes more sense to clarify the workflow first or to start with simple automation. It goes similarly when someone starts with a tool and looks for applications afterwards. In the end there are licenses that hardly anyone uses.

Buy, configure or build

For the same task there are usually three routes. A ready-made assistant such as Microsoft Copilot covers general tasks like writing, summarizing and research. Rollout is quick and needs little adaptation, but the benefit depends on whether the teams really use it day to day. Many applications also bring their own AI features, such as CRM systems, BI tools or ticketing systems, and low-code tools let you put together custom assistants without programming. The advantage is proximity to data and permissions that already live in the software.

A custom application calls a language model or a machine learning model through a cloud provider’s platform. That pays off for special workflows, deep integration or requirements no product meets. In return you have development, testing, operation and maintenance.

We first check whether an existing product covers the task. What gets compared is license cost per person, usage-based cost, vendor dependency and the effort of running it. How to roll out assistants and choose models is described in the topic on generative AI.

Data readiness, roadmap and ownership

Whether a use case works is often decided by the data. For the prioritized initiatives we check whether the data exists, is of usable quality, is technically accessible and current enough, and whether permissions and data protection fit. This covers only the data the particular use case needs. It does not replace a data strategy, but it shows where foundations are missing.

The roadmap puts the initiatives in order and accounts for dependencies, such as a data foundation several initiatives use. For each initiative it names a responsible person in the business and sets decision points. Without that person there is no reason to review and develop results further. A few initiatives that get finished are better than many parallel pilots. Organization also covers who is accountable for AI initiatives, who reviews results and which skills must exist in-house. We name what should be built up, bought in or trained.

Measuring value before the project starts

Value is defined before the start, not after. That includes the baseline in today’s process, a target value, a quality measure and the expected running cost. After rollout we record the same figures again, plus usage and quality in daily work, for example through spot checks. Time saved only counts as value if the time is used differently. We also decide in advance at which values an initiative is continued, adjusted or ended.

Inflated expectations mostly come from demos. Language models phrase things convincingly even when a statement is wrong, and a flawless answer in a demonstration says little about the rate in daily use. Comparing against today’s process helps, because manual workflows make mistakes too. We also check vendor claims. Gartner uses the term “agent washing” for existing chatbots and automations being marketed as agents.

Surveys show how far expectation and reality are apart. In McKinsey’s survey “The state of AI in 2025”, nearly two thirds of respondents said their organization had not yet started scaling AI across the enterprise. In a Gartner survey of CIOs published in October 2026, only 13 percent said their enterprise had realized significant or transformational value from AI so far. Gartner also predicted in 2025 that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing costs, unclear business value or inadequate risk controls. The cost per amount of text processed has fallen considerably in recent years for models of a given performance class. Total cost does not fall automatically as a result, because consumption grows with usage, and models that reason longer before answering need more.