# Assessing AI Use Case Feasibility: Data, Systems, Error Tolerance

> Judge feasibility along separate dimensions: data, system access, error tolerance, model capability and organisation. Each has a quick test on real cases that gives a yes, a no or a conditional yes within days. A single no in any dimension stops the use case or changes its scope.

Canonical URL: https://www.ruebstudio.de/en/guides/evaluating-ai-use-cases/ai-feasibility-assessment  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-10-22

Most AI projects that stall do so for reasons that were visible on day one. The data sits in a system with no API. Nobody is allowed to write back into the ERP. Or a wrong answer would be so costly that every case ends up checked by hand anyway. A feasibility assessment exists to surface these problems before money is spent, and it does not need to be a lengthy report. A handful of short, separate tests on real cases is usually enough.

## What is an AI feasibility assessment, and why do it?

An AI feasibility assessment checks, before you build anything, whether a use case can work with the data, systems and people you actually have, and whether the remaining errors are acceptable. It does not ask whether the use case is worth doing. That is a separate question about value, and mixing the two hides the information you need.

The split matters because each question needs different people. Sales knows whether an agent that drafts quotes would be useful. IT knows whether it can reach the price lists, the bills of materials and the CRM. Management decides whether a mispriced quote is a risk the business can carry. Squeezing all of that into one score from one to ten throws away the reasons behind the number.

Common triggers for an assessment:

- A department asks for something specific, such as pre-sorting service tickets or matching delivery notes against purchase orders.
- A workshop has produced a long list of ideas and you need to decide which to try first.
- A vendor has shown a convincing demo, and nobody knows whether it would behave the same way with your own data.

The underlying question is always the same: what would need to be true for this to run day to day, and is it true?

## Which dimensions of feasibility should you assess?

Five dimensions can be checked independently: data readiness, system access, error tolerance, model capability and organisation. Each gets its own verdict: yes, no, or yes with a condition. The overall decision comes from these verdicts, not from an average, because a strong dimension cannot make up for a failed one.

Consider a use case with clean data and a capable model, but no permitted place to write the result. Averaging would call it promising. In reality it cannot run. So the rule is simple: any single dimension can stop the use case on its own.

| Dimension | Core question | Typical knock-out |
|---|---|---|
| Data readiness | Is the information written down, current and findable? | Knowledge lives only in people's heads |
| System access | Can an agent read what it needs and write where results belong? | No permitted read or write path |
| Error tolerance | What does a wrong output cost, and who catches it? | Errors go unnoticed and cannot be undone |
| Model capability | Does a current model handle real cases well? | Outputs unusable even with good context |
| Organisation | Is there an owner, and are legal and staff questions settled? | Nobody owns the process |

Operations is a sixth consideration: cost per case, response time, monitoring. It rarely decides yes or no, but it often shapes the design, for example whether the model runs with a cloud provider or on your own hardware. If data must not leave your premises, look at [sovereign AI on your own infrastructure](https://www.ruebstudio.de/en/sovereign-ai) at the same time.

## How do you test data readiness?

Data readiness is good enough when someone doing the job today can handle a few real cases using only documents and systems, without asking colleagues or relying on memory. The quick test is a source trace: ask one person to process five current cases and write down where every piece of information came from.

This takes a few hours and tells you more than a data inventory. Each case produces a short list: customer details from the CRM, prices from a spreadsheet on a shared drive, lead time from a phone call to purchasing, special terms "I just know". The last two items are the warning signs. If a person only knows something from experience or a conversation, an agent cannot know it either.

Check four properties for each source:

- **Exists:** Is the information written down anywhere?
- **Reachable:** Can it be found through a shared folder, a search index or an API, or is it buried in someone's inbox?
- **Current:** Who maintains it, and how would you spot an outdated version?
- **Unambiguous:** Are there competing versions, such as three price lists with different dates?

Missing data is not automatically a knock-out. Often the fix is a maintained table or a tidied folder, which turns the verdict into a condition: "Feasible once special terms are kept in one maintained list." It becomes a real no when the knowledge cannot be written down at all, or nobody is willing to keep it up to date. A separate article goes deeper into judging documents and master data.

## How do you check system access and integration?

System access is feasible when every input can be read in a technical and permitted way, and the result has a place it may be written. The quick test is a one-page sketch: for one case, draw the path of each input and output and mark whether there is an API, an export or only a screen.

With AI agents the write side is the one people forget. An agent that writes a good answer which someone then pastes into the ticketing system by hand is an assistant, not an [AI agent](https://www.ruebstudio.de/en/glossary/ai-agent) in the full sense. That can be perfectly fine, but it changes what you should expect from it. Clarify early:

- **Read:** Does the system offer a documented interface (REST, OData, direct database access), a scheduled export, or only a screen?
- **Write:** May a technical user create or change records, such as draft opportunities in the CRM or comments on a ticket? Is there a draft state that forces a human approval?
- **Identity and permissions:** Which account does the agent run under? Can it see more than the person it is working for? For knowledge management this is the deciding question, because folder permissions can be lost if the search index ignores them.
- **Operations:** Who runs the source system, and are there maintenance windows or rate limits that would slow an agent down?

A no here is rarely final. There is often a workaround: a nightly export, a shared mailbox as the inbound channel, a draft instead of a direct booking. Record workarounds as conditions and judge honestly whether they would hold up in daily use. In [technical service](https://www.ruebstudio.de/en/ai-agents/technical-service), this is often where you decide whether an agent only suggests or actually acts.

## How much error tolerance does the use case have?

Error tolerance is sufficient when wrong outputs are either cheap, reliably caught by a person before they cause harm, or easy to reverse. The quick test is an error cost table: for each kind of mistake, write down what happens, who notices, and what it costs to fix.

Every model makes mistakes. What matters is which mistakes and what they lead to. The [Google PAIR Guidebook](https://pair.withgoogle.com/chapter/user-needs/) points out that it is tempting to weigh false positives and false negatives equally, but that this rarely matches their real consequences. For an agent that triages service tickets, an urgent ticket filed as routine is a very different problem from a routine ticket flagged as urgent.

A simple error cost table for a draft quote:

| Error | Who notices? | Consequence if missed | Reversible? |
|---|---|---|---|
| Wrong item proposed | Sales, at approval | Wrong goods shipped | Yes, with effort |
| Outdated price | Possibly nobody | Lost margin | No, once accepted |
| Typo in cover letter | Sales | Looks careless | Yes |

The table tells you how to shape the use case. Where an error can go unnoticed and cannot be undone, you need either a human approval at exactly that point or an automatic check, such as comparing every price against the current list. PAIR names this situation directly: AI is probably not the right choice when the cost of errors is very high and outweighs the benefit of a small gain in success rate. If an error can be neither caught nor reversed, the verdict is no.

## How do you test whether a model can do the task?

Model capability is established when a current language model, given the same context the finished agent would have, produces outputs that a subject expert rates as usable on a sample of real cases. The quick test: twenty to thirty real cases, run through the model by hand and scored against criteria you wrote down beforehand.

Order matters. Define success first, then test. Anthropic's [guide to success criteria and evaluations](https://platform.claude.com/docs/en/docs/test-and-evaluate/define-success) says good criteria are specific, measurable, achievable and relevant, and that most use cases need several criteria at once. For a service use case that might mean: correct category, complete follow-up questions, no invented spare part numbers.

A practical run looks like this:

1. **Pick cases:** recent, real ones, including messy and borderline cases, not just the clean examples from a demo.
2. **Give realistic context:** exactly the documents and data the agent would get from the systems later. Adding more makes the result look better than it will be.
3. **Write criteria first:** what counts as usable, and what counts as a dangerous error?
4. **Have an expert score:** ideally the person who does the job today, without knowing which output came from which variant if you compare several.
5. **Sort the failures:** was it the model, missing context, or an unclear task?

Step five is the most useful. Often the model could do the job but lacked information, which is a finding about data readiness rather than model capability. Check the other direction too. PAIR notes that a rule-based or heuristic solution often works as well as an AI version, or better. If the task can be fully expressed as fixed rules, conventional software is usually the sturdier option.

## What organisational conditions does an AI use case need?

A use case is organisationally feasible when a named person owns the process, the people affected are involved, and legal duties such as data protection and staff consultation can be met before launch. The quick test is a one-hour meeting with the business owner, IT, data protection and employee representatives, using a fixed list of questions.

The list:

- **Ownership:** Who decides on process changes and who judges quality once it runs? Without that person, any agent stays an experiment.
- **Staff consultation:** Depending on the country, employee representatives may have a say over systems that could monitor staff. In Germany, for example, [section 87 of the Works Constitution Act](https://www.gesetze-im-internet.de/betrvg/__87.html) gives works councils co-determination rights over technical systems intended to monitor employee behaviour or performance. Whether your agent falls under such rules is a case-by-case question, so raise it early.
- **Data protection:** [Article 35 of the GDPR](https://gdpr-info.eu/art-35-gdpr/) requires a data protection impact assessment when processing, particularly with new technologies, is likely to result in a high risk to the rights and freedoms of individuals. Your data protection lead should judge this before effort goes into building.
- **Regulation:** [Annex III of the EU AI Act](https://artificialintelligenceact.eu/annex/3/) lists, among other areas, AI used for recruitment and candidate selection, for decisions on promotion or termination, and for monitoring and evaluating the performance and behaviour of workers as high-risk. An agent that drafts quotes will normally sit well outside this; one that screens job applications sits inside it.
- **Skills:** [Article 4 of the AI Act](https://artificialintelligenceact.eu/article/4/) requires providers and deployers to take measures to support AI literacy among their staff, and has applied since 2 February 2025. In practice, whoever works with the agent must understand what it does well and where it goes wrong; the guide to [AI literacy under Article 4](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/ai-literacy-article-4-ai-act) covers what that means in practice.
- **Acceptance:** Do the people doing the work today want this help? PAIR advises against automating tasks that people have explicitly said they do not want automated.

Treat these points as orientation, not legal advice. Their job is to show where a specialist review is needed, early enough to plan for it.

## What are the steps of a feasibility assessment?

A lean assessment has five steps: describe the use case precisely, check the knock-out criteria, run the quick tests, record a verdict per dimension, and turn conditions into a pilot scope. It usually takes a few days to two weeks, and the main cost is time from people who know the process.

1. **Describe the use case in one sentence.** Trigger, input, output, target system. For example: "When a request arrives by email, the agent creates a draft quote in the CRM for sales to approve." Without this, none of the dimensions can be tested.
2. **Knock-outs first.** In a short meeting, check what would end things immediately: no owner, a high-risk classification nobody wants to take on, no conceivable write path. This spares you testing ideas that were never going to happen.
3. **Run the quick tests in parallel.** Source trace, integration sketch, error cost table, model test and the organisation meeting involve different people, so they can run at the same time.
4. **One verdict per dimension.** A word and two sentences of reasoning each. Make conditions concrete: what exactly, who, by when.
5. **Derive the pilot scope.** Conditions become the design: approval instead of direct booking, one product group only, read-only at first.

The result fits on one or two pages. When you compare several use cases, add the value assessment; a separate article on prioritisation covers how to set value against feasibility without diluting the knock-outs. The same questions apply almost unchanged to [quotation](https://www.ruebstudio.de/en/ai-agents/quotation), [knowledge management](https://www.ruebstudio.de/en/ai-agents/knowledge-management) and [sales](https://www.ruebstudio.de/en/ai-agents/sales) processes.

If you want a broader framework, the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers one. Version 1.0 was released on 26 January 2023 and is organised around four functions: Govern, Map, Measure and Manage. A companion Generative AI Profile, NIST AI 600-1, followed on 26 July 2024. For a first go/no-go decision it is more than you need, but it makes a useful checklist for later stages.

## What are the limits of a quick feasibility check?

Quick tests show whether a use case can work in principle, not how it behaves over months in production. Rare edge cases, data that slowly goes stale, and reviewers who start approving without reading will not show up in twenty test cases. The assessment prepares a measured pilot; it does not replace one.

Common pitfalls:

- **Hand-picked cases.** Testing only tidy examples tests the demo, not the job. Include the messy ones on purpose.
- **Too much context in the test.** If the model test gets information the agent will not have later, the result is too optimistic.
- **Approval as a fig leaf.** Someone approving a hundred drafts a day will stop reading closely. If error tolerance depends on a human check, that check has to be realistic in volume.
- **Feasible is not the same as worthwhile.** A positive verdict says nothing about whether the effort pays off. That needs the value side, for example a [calculation of AI ROI](https://www.ruebstudio.de/en/guides/evaluating-ai-use-cases/how-to-calculate-ai-roi).
- **Stale verdicts.** Models change quickly. A no on model capability from a year ago deserves a retest with current models; a no on data or organisation usually does not.

For use cases near the high-risk categories or involving sensitive personal data, a quick check is never enough on its own. There it only tells you which fuller reviews are required.

## How do I get started?

Start with one use case, perhaps drawn from a list of [AI use cases by department](https://www.ruebstudio.de/en/guides/finding-ai-use-cases/ai-use-cases-by-department), that has a clear owner and happens often enough to collect twenty real cases. Describe it in one sentence, run the source trace on five cases and check the knock-outs. By then you usually know whether the remaining tests are worth running.

A good first candidate meets three conditions: the information it needs is mostly written down, errors are caught before anything reaches a customer, and the person doing the work today would welcome the help. Draft quotes, ticket triage and answers to recurring internal questions often fit. Hiring decisions, payment approvals and legally binding statements rarely do.

Write the outcome down, even when it is a no. A reasoned no that states the condition under which it would become a yes is a solid starting point for the next round. An overview of which kinds of [AI agents](https://www.ruebstudio.de/en/ai-agents) suit which processes can help you phrase that first one-sentence description.

## Frequently asked questions

### How long does an AI feasibility assessment take?

With one quick test per dimension, usually a few days to two weeks, depending on how fast you can collect test cases and get access. Most agent use cases do not need a months-long study.

### Do I need a working integration before testing?

No. The model test uses real cases copied in by hand, and the systems check only confirms that read and write paths exist and are permitted.

### Why do AI use cases usually fail feasibility?

Rarely because of the model. Missing data access and unclear ownership are more common, which is why it pays to check data and organisation first.

### Is a feasibility assessment the same as a data protection impact assessment?

No. It only tells you whether a DPIA, works council consultation or similar review is needed, so you can plan for it early.

## Sources

- [Anthropic: Define success criteria and build evaluations](https://platform.claude.com/docs/en/docs/test-and-evaluate/define-success)
- [Google PAIR Guidebook: User Needs + Defining Success](https://pair.withgoogle.com/chapter/user-needs/)
- [NIST: AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)
- [Art. 35 GDPR – Data protection impact assessment](https://gdpr-info.eu/art-35-gdpr/)
- [German Works Constitution Act, section 87 (in German)](https://www.gesetze-im-internet.de/betrvg/__87.html)
- [EU AI Act, Annex III: High-risk AI systems](https://artificialintelligenceact.eu/annex/3/)
- [EU AI Act, Article 4: AI literacy](https://artificialintelligenceact.eu/article/4/)
