AI consulting for companies: where to start with AI

Part 01

Why most AI pilots stall

The pattern repeats across European companies. Someone runs a pilot with a chatbot on a few documents, the demo impresses, and six months later nothing runs in production. The model was rarely the problem.

In its 2024 survey of 1,000 executives, BCG found that 74% of companies had yet to show tangible value from AI. The same study traced around 70% of the challenges to people and processes, 20% to technology and only 10% to the algorithms (BCG, October 2024). Most meeting time goes to the model. The outcome depends on the process around it.

Eurostat sees the same gap from the other side. In 2025, 19.95% of EU enterprises with 10 or more employees used AI. Size matters: 55.03% of large enterprises used it, against 17% of small ones (Eurostat, December 2025).

Ask the companies that considered AI and decided against it, and the top reason is skills, not cost or doubt about the technology.

Only 20.68% said AI was not useful for them. The other answers point to missing expertise, legal uncertainty and data protection. None of them is solved by a better model. A clear choice of where to start, a fixed scope and a plan for the rules solve most of it.

Part 02

Where to start with AI: five tests for the first use case

Think of an accounting team matching supplier invoices to purchase orders, a property manager sorting tenant requests, or a fund screening the decks that arrive every week. Each is a candidate. Run each through five tests.

1. VolumeHow often does it happen? Forty invoices a day beats one board report a month. Volume is where the hours are, and it gives you enough cases to test against.
2. Rules clarityCould a new hire do it with a written checklist and a few examples? If the answer depends on one senior person's judgment, the AI will struggle in the same places the new hire would.
3. Data accessIs the input digital and reachable? Emails, PDFs and an ERP with an API are fine. Paper files in a cabinet or a system nobody can export from are a project before the project.
4. Cost of an errorWhat happens when it gets one wrong? A misrouted tenant email costs minutes. A wrong clause in a signed contract costs much more. Start where a person can catch mistakes cheaply.
5. Measurable baselineDo you know today how long it takes, how often it goes wrong and what it costs? Without a baseline you cannot prove the result, and the project loses its budget at the next review.

A pre-seed VC fund in Europe started with deal-flow scoring: high volume, criteria the partners could write down, and data already in their inbox and CRM. The system was live in production 9 days after kickoff in 2026. The speed came from the choice of use case as much as from the build.

Part 03

Score your use cases with a weighted matrix

Copy this worksheet into a spreadsheet. Score each candidate from 1 to 5 per criterion, multiply by the weight and add up. The maximum is 50. Volume gets the highest weight because it drives both the value and the number of test cases.

Here is the worksheet filled in for six typical candidates in a mid-sized European company.

To see the trade-off, split the score in two. Value is volume and baseline. Feasibility is rules, data and cost of an error. Plot both and the order shows at a glance.

The invoice case wins on both axes. The board report is easy but rare, so a template or a simple automation covers it. Contract review is valuable in theory, but an error is expensive and the rules live in a lawyer's head. It belongs later, once a first agent has earned trust.

Part 04

Build or buy, and what data and access you need

Before anyone writes code, check whether a product already does the job. Your accounting software may already read invoices. Your helpdesk tool may already sort tickets. Buying is often the right answer, and a consultant who never says so is selling hours.

Whichever way you go, the same inputs decide whether it works:

• Real past cases with the correct outcome attached: the invoices and how they were booked, the tenant emails and who handled them. They become the evaluation set.
• Read access first, through a service account with the narrowest permissions that work. Write access comes later, step by step.
• One person on your side who knows the process and can answer questions within a day.
• A data processing agreement and a clear answer on where the data is processed. For a European company, the sensible default is the EU.

Part 05

What to measure, and the governance basics

Measure the process before you touch it. Then measure the same things after go-live, on the same definitions.

Governance for a first use case does not need a committee. It needs two things the EU AI Act already points to.

The first is AI literacy. Article 4 has applied since 2 February 2025. In the current consolidated text, providers and deployers of AI systems must take measures to support the AI literacy of their staff and others who operate AI on their behalf, taking into account their knowledge, experience and the context of use. A company that uses an AI system at work is a deployer. In practice, train the people who will work next to the agent, on their own cases, and write down that you did.

The second is human oversight. Article 14 requires that high-risk AI systems can be effectively overseen by people while in use. Most first back-office use cases are not high-risk, but the same design is good practice anyway: a confidence threshold, a person who approves what falls below it, and a log of every decision.

Part 06

What an AI Audit should deliver

An AI Audit is the first step of AI consulting. Whoever runs it, you should walk away with four things you can use without them:

1. A process mapHow the work runs today, step by step, with the time each step takes and the tools it touches.
2. A ranked opportunity registerEvery candidate scored on value and feasibility, like the worksheet above, with a build-or-buy view for each.
3. A 12-month roadmapWhich use case comes first, which follows, and what each one depends on.
4. A fixed price for the first use caseAgreed before you commit to anything. An open estimate moves the risk of overruns to you.

If you cannot say how long a process takes today, you cannot prove that AI made it faster.

Alpgency

This is what our AI consulting work produces. The Audit is free and starts with a 30-minute call. You keep the roadmap and the findings whether you build with us or not. If you go ahead, the path has three steps.

Before the build, you see a working demo on your own data. A demo on sample data proves little about yours. After go-live you choose how it runs: your team takes it over with the code, prompts and evaluation set, you pay a fixed monthly support fee, or we stay on as tech partner on a fixed schedule. We never bill by the hour.

FAQ

Frequently asked questions