ROI Without Risk Is Marketing

180 percent ROI sounds great. Until someone asks how likely it is. A return figure without a risk profile is not half the truth - it is the pleasant half, the one you like to show while the unpleasant half stays in the fog. An AI project with 120 percent ROI and high risk can be the worse decision than one with 80 percent and low risk. But you only see that if you measure risk at all - and in a structured way, not from the gut and not as an afterthought that "feels doable".

We measure it across five dimensions. Here they are, what they mean, and why they carry different weight.

The five dimensions

Complexity. How clear is the input, how many exceptions and special cases exist, how many systems hang off the process? The more chaotic the thing the AI is meant to take over, the more fragile the result. A cleanly structured process with clear rules is a good AI candidate. A process that is 40 percent exceptions carried in an experienced employee's head is a minefield. Complexity is the most frequent silent killer of AI projects, because it never shows up in the pitch.

Volume. How often does the process run, how much does the load fluctuate? A use case that runs twenty times a month forgives errors differently from one with tens of thousands of runs, where every small weakness multiplies. High volumes are good for the ROI, but they raise the bar for stability and monitoring.

Financial leverage. How quickly does it pay off, how long is the break-even horizon? A project that only turns black after two years carries a very different risk from one in the black after four months - simply because more can intervene in the longer window.

Dependency. How easily can you switch provider, how portable is your data, and - the most important point - how many people genuinely understand the system?

Maintenance. How much ongoing care does it need, how often does the underlying basis change, is there monitoring and alerting, or are you flying blind until something surfaces?

Why not everything counts equally

The obvious question: why do we weight complexity and financial leverage at 25 percent each, volume at 20, and dependency and maintenance at 15 each - rather than all five equally? Because not every dimension threatens the ROI to the same degree. Complexity and financial leverage attack profitability most directly: an overly complex process or a distant break-even tips the case immediately and completely. Dependency and maintenance are just as real, but they act more slowly and can be cushioned by measures.

Equal weighting would flatten these differences and produce a false picture - it would give an easily fixable maintenance risk the same weight as a fundamental complexity trap. The weighting is therefore not a technical detail but the actual substantive statement of the model: it tells you what to look at first.

The bus factor: the most expensive risk with the most harmless name

One dimension deserves particular attention because it is almost always underestimated: dependency on individual people. In many AI projects the entire body of knowledge - the prompts, the logic, the workarounds, the instinct for when the system is off - sits in a single head. This is called the bus factor: how many people would have to drop out for the project to stop?

If the answer is "one", you do not have an AI project, you have a concentration risk with a login. That person goes on holiday, changes job or falls ill, and suddenly nobody knows why a given prompt is built the way it is or what happens if you change it. The mitigation is unspectacular and effective: document it, spread the knowledge across several shoulders, build a small centre of competence instead of a hero. It costs some discipline and saves a total outage later.

When risk becomes a board matter

In regulated industries one dimension tips the whole picture: compliance and data protection. What is a gut feeling in a marketing agency becomes a board question in a bank, an insurer or in healthcare. A highly critical data protection finding is not a deduction of a few percentage points on the risk score - it is a stop signal that makes the ROI temporarily irrelevant until the question is settled. The finest business case is worthless if the project breaches data protection law.

This is exactly where the ROI view connects with your compliance strategy, around revised Swiss data protection law and the EU AI Act. At this point risk is no longer a feeling but a documented fact with an owner.

Two projects, one decision

Make it concrete. Project A: an AI chatbot for customer service, 120 percent ROI over twelve months. Impressive - until you lay the risk profile next to it. The process is highly complex, every second enquiry is a special case, the whole setup depends on one external developer, and it involves customer data in a regulated industry. High ROI, high risk across three dimensions at once.

Project B: AI-assisted invoice pre-capture, 80 percent ROI. Less spectacular - but the process is clearly structured, runs at high and stable volume, the knowledge is documented within the team, and it is uncritical in data protection terms. Lower risk across every dimension.

Look only at the ROI and you pick A. Factor in risk and you see that B is the calmer, more predictable investment, the one that tips over less often in operation - and possibly the smarter choice, even though the naked number is smaller. You only get that insight when both profiles sit side by side.

What you do with this

Never look at ROI and risk separately. A return figure belongs next to a risk profile, otherwise it is marketing. Only both together allow the honest question: is this return worth its price in uncertainty? Sometimes the answer is yes, even at high risk - but then it is a deliberate decision instead of a blind one.

Our AI ROI calculator assesses every initiative across these five dimensions - and is honest enough to mark a risk as "not answered" instead of inventing a reassuring "medium" that nobody entered.

→ Assess your risk profile