Method for the “Decisions with AI” practical tasks

This appendix sets out the procedure for the “Decisions with AI” practical tasks. The purpose of these tasks is to develop three essential competences of a present-day analyst:

• the ability to formulate an evaluation problem and to integrate artificial intelligence (AI) tools into the business valuation process;
• the skill of critically verifying AI-generated results;
• readiness to take personal responsibility for the final decision.

(A broader rationale for including these tasks in the textbook is given in section 5 of the Introduction.)

The analytical “Decisions with AI” tasks conclude the review questions and exercises section of every chapter of the textbook. By their nature they differ fundamentally from traditional tasks: standard exercises test what the student can do alone, whereas this format tests the ability to manage technological tools while keeping a critical distance (avoiding blind reliance on the AI tool). Because the mechanics of the tasks are the same in all chapters, they are described in detail only in this appendix. Sections A.1–A.7 are intended for a general audience (both students and lecturers), while section A.8 is designed specifically for the needs of lecturers (as moderators and assessors of the learning process).

Illustrative example. The whole appendix draws on a single hypothetical case. Kavos stotelė, a private limited company (UAB) that runs one café in Kaunas, is considering expansion: opening a second outlet. The owner asks an AI assistant to assess whether the investment is worthwhile. Within a few minutes the AI assistant generates the following summary:

The second café will require an investment of EUR 120 thousand (equipment and fit-out). It is likely that 200 customers will visit per day, the average bill will be EUR 3.50, and the café will operate 300 days a year. Annual revenue will be EUR 210 thousand. Variable costs (coffee, milk, cups) will amount to 35 % of revenue, and fixed costs (rent, wages, utilities) to EUR 96 thousand a year. Annual cash flow will be about EUR 40 thousand, so the investment will pay back in 3 years. Recommendation: open.

There are no arithmetic errors in this AI summary: the calculations are accurate. However, as the verification steps below will show, in this case absolute mathematical accuracy is no guarantee at all that the conclusion is economically sound.

A.1 Two tasks and levels of openness#

Each chapter contains two tasks.

Basic task. The student is given an assessment that has already been prepared — for example, an AI-generated summary, a calculation or a table. Everything looks convincing and the calculations are correct, but an error or a weak assumption has been deliberately built into the result. The student checks this assessment against the protocol (A.2) and states whether its conclusion is correct. Because the lecturer knows the error in advance and it is the same for all students, this task shows whether the check actually finds anything.

Example: the lecturer presents the café summary from the beginning of this appendix and asks whether the AI’s recommendation can be accepted. In this appendix the café is only an illustration: its weak points are revealed in A.2–A.6, so for an assessed basic task the lecturer prepares a separate variant (A.8).

Extended task. Here the student formulates the problem themselves: decides which question needs to be answered, breaks it down into smaller subtasks, assigns some of them to AI, checks the results and draws conclusions.

Example: the owner of Kavos stotelė asks whether expansion is worthwhile. The student decides on their own that the real question is not “will the café pay back” but “will a second café increase the cash flow of the whole company”, because some customers may simply move from the first café.

Qualitative and quantitative assessment. First, the factors that determine the decision and their causal links are identified — this is the qualitative assessment. Quantitative assessment usually counts in the student’s favour, but in the tasks of some chapters it is required (the task itself states this clearly). The requirements and examples of a task do not go beyond the chapter material. Anything that cannot be calculated with the methods set out in that chapter or in earlier chapters of the book is not required. In that case a qualitative assessment is enough, because what matters most is whether the student has understood. Methods from other topics are not required, but applying them on the student’s own initiative is encouraged and counts in the student’s favour.

Example: first it is established that the café’s result depends most on the number of customers and on the fact that some customers may move from the first café (qualitative assessment). Then, if break-even analysis has already been covered in that chapter or an earlier one, the student calculates the minimum number of customers per day at which operating cash flow is not negative (about 141 customers).

What is not assigned to AI. In the extended task, two things must not be handed over to AI: formulating the problem itself and taking the final decision (the verdict). In the basic task the problem is already given, because what is tested there is precisely the ability to check results; but the final verdict there, too, remains solely the student’s responsibility.

Extended task workflow (E1–E7). The extended task is carried out in seven stages. They are labelled E1–E7 so that they are not confused with the protocol steps (0a, 0b and 1–6, A.2). The chapter task gives examples for its own topic.

• E1. Formulating the research problem. A single main question is formulated: its answer will determine the final decision, and the available data do not answer it directly. It must be a specific question, not an abstract research topic.
• E2. Analytical objective and format of the result. The student defines which intermediate data are needed to construct the answer and in what form they will be presented. If the assessment is quantitative, the result is expressed as a range of values (an interval) with clearly stated assumptions, not as a single point estimate; if it is qualitative, the possible interpretations and the conditions under which the conclusion would change are stated.
• E3. Decomposition into subtasks and assignment to AI. The objective is broken down into 3–5 separate subtasks. For each subtask assigned to AI, the student completes a handover record (A.3). Formulating the research problem and reaching the final verdict must not be assigned to AI; at least one important subtask is done by the student alone, with an explanation of why it was not assigned to AI. Before E4, the student writes down protocol steps 0a and 0b (see E5).
• E4. Execution and prompt log. The planned subtasks are carried out. All AI prompts and the answers received are copied into the prompt log (A.3) without editing: the lecturer must see the authentic AI-generated text, not a stylistically or logically corrected version.
• E5. Verifying the results (critical check). The result is checked with all eight protocol steps (A.2). The expectation and the falsification criterion (protocol steps 0a and 0b) are recorded before any prompts are submitted to AI.
• E6. Research conclusions and review condition. The student answers the research question and assesses whether the result can be relied on in making a decision. A review condition is set (A.4): an indicator whose change would require the conclusion to be reviewed. Concluding that the result can be relied on is as acceptable as concluding that it cannot; in both cases the protocol step 5 log clearly states what could not be verified and what risk this creates.
• E7. Defence, critique and reflection. In the seminar the student, as defender, presents and defends the conclusion reached from their analytical perspective, and then acts as opponent to another student’s work (A.5). The opponent’s fundamental task is to identify the load-bearing assumption — the assumption without which the defender’s whole conclusion would collapse.

Submission. Unless the task states otherwise, up to 4 pages are submitted (E1–E6 records; shorter for remote work, see A.5), with the handover record and the unedited prompt log in an annex. After the seminar the defender submits a one-sentence reflection (what they would change in their work in light of the opponent’s criticism, and what they would not change). The opponent submits a three-sentence summary (which load-bearing assumption was found, why it is load-bearing, and what data were missing).

Levels of openness. A task can start in different ways — from very specific to very open. The higher the level, the less initial information the student is given and the more decisions they have to take themselves. Levels 1 and 2 apply to the basic task, levels 3–5 to the extended task.

The level is chosen by the lecturer (see section A.8). The basic task usually remains at level 2, which makes it possible to compare how the student’s ability to check has changed from the first level-2 task to the last chapter; level 1 suits the first, introductory task.

A.2 Verification protocol#

The protocol is the sequence of steps used to check a result produced by AI (or by any other source). The steps are the same in both tasks and in all chapters. The first two steps (0a and 0b) are carried out before the AI answer is received — afterwards they can no longer be done honestly, because the answer is already known. In the basic task, where the AI result is given together with the task, steps 0a and 0b are written down before the result is analysed, on the basis of the situation description alone; in class the lecturer may first present the situation and only later the AI summary.

Example of step 5 (residual uncertainty log).

Why are steps 2 and 3 kept separate? Step 2 checks a fact on which the result relies: the fact is traced and verified in a primary source, whether or not the AI cited a source. Step 3 looks for a claim made without any source or support. These are entirely different analytical skills, and the latter (step 3) is harder: the student has to notice what is missing from the text, not just check what is written. In a hypothetical case, where the company’s data cannot be verified, step 2 checks a general fact mentioned in the summary (for example, an inflation target).

What is “independent reconstruction” (step 1)? It is the reconstruction of the most important part of the result by the student, using another data source or an alternative method of calculation. The following do not count: asking the same AI the same question again, asking another AI model, or asking a friend to agree. Different AI models can make identical mistakes, so agreement between their answers creates an illusion of confirmation. In reality it is just one more opinion, not an independent check.

Why is the answer “I checked everything” not acceptable for step 5? Analytically, such an answer says nothing. The log must state clearly what was checked, what could not be checked and why. If no gaps were found within the limits of the check, this may be stated, but the limits of the check must still be specified precisely.

Poor: “I checked everything; there are no errors.”

Good: “I checked the logic of the calculation and the rent; I could not check the number of customers, and it is precisely on this number that the decision directly depends.”

A.3 Handover record and prompt log#

The phrase “I did it with AI” says nothing — it is unclear what the AI did and what the student did. Therefore, for each subtask assigned to AI, the student completes four fields:

• subtask — what had to be done;
• assigned to — which tool and in what role. The same AI can be used as a calculator and as a source finder — these are two different assignments with different risks;
• what was received — briefly, what the answer was;
• what was done with it — accepted, checked, corrected or rejected. This field is essential: without it, it is unclear whether the AI answer was checked.

Subtask not assigned to AI. At least one subtask that was deliberately not assigned to AI is named, with an explanation of why. Acceptable reasons: confidential data; the need to trace where a result comes from; the student’s personal responsibility for the claim; or the absence of a similar case that could be relied on. Example: “I did not upload the existing café’s cash register data to AI, because they are confidential company data — I calculated the number of customers myself.” The reason “I wanted to do it myself” is not considered sufficient.

Prompt and answer log. During the class or the work, all prompts submitted to AI, the answers received and what was done with them are recorded in a table (for example, in Excel or in a form provided by the lecturer). Answers are copied without editing; a long answer may be shortened to an excerpt, with the omitted part marked […]. The log is submitted together with the work (for example, attached in Moodle). It allows the lecturer to assess not only the final result but also the route to it — which makes assessment more objective and feedback more precise.

A.4 Verdict#

The verdict is the student’s final decision: whether the conclusion is correct and whether the recommendation is acceptable. It must be accompanied by a sentence stating when the decision would be reviewed (the parts “until then I will monitor” and “responsible for the review” are included when there is someone who will carry out the decision):

“I will review this decision if …; until then I will monitor …; responsible for the review: …”

Example: “Do not open the second café now. I will review this decision if counting passers-by outside the new premises, together with data from a similar café (what share of passers-by become customers), shows that at least 200 customers per day can be expected; until then I will monitor offers of vacant premises and the customer flow of the first café; responsible for the review: the owner.”

If the decision is irreversible, this is noted, together with how this risk can be reduced in advance. Example: “The lease is signed for five years, so terminating it would be expensive. Before signing, the right to terminate the lease after the first year should be negotiated.”

Defending a decision does not mean never changing it: if the opponent (A.5) points out an important error, changing the verdict is the right thing to do.

A.5 Defence and critique#

The student has to defend their decision — explain it and answer questions. How this happens depends on the circumstances.

Why the written work is shorter for remote work. When there is a conversation, what is assessed is what the student says and explains, not what is written. Anything written that the student cannot explain does not count. A long text therefore adds nothing.

Remote work without contact takes place in two stages. First, the work is submitted and locked. Then a random changed circumstance from a list prepared by the lecturer is revealed, and within a set time the student answers in an audio or video recording: how this circumstance changes the assessment and whether the verdict changes. A defence text written in advance is not acceptable, because it could be prepared together with the work — including with AI.

Roles in class. Four roles operate in the seminar.

• The defender presents their work and answers questions. After the seminar they write one sentence: what they would change and what they would not change after the opponent’s questions. Example: “I would change the assumption about the number of customers to a range of 150–200; I would not change the verdict ‘do not open now’, because the number of customers has still not been verified.”
• The opponent is another student who has received the defender’s work in advance. Their task is to find the assumption without which the conclusion would collapse (the load-bearing assumption). After the seminar they write three sentences. Example (opposing a work whose verdict is “open”): “Assumption found: 200 customers will come per day. It is load-bearing because, with 150 customers, the investment pays back only after almost 19 years even under the AI assumptions, and with the actual rent the cash flow is close to zero. Data on customer flow at that specific location are missing.”
• The lecturer moderates and assesses (see A.6, A.8).
• The audience — the other students — also learns to assess. Before the critique, each of them writes down which assumption, in their view, supports the conclusion and how confident they are in that view (from 0 to 100). After the critique, they write down whether the opponent’s argument showed that the conclusion does not hold, why, and how their confidence has changed. Example: “Before: load-bearing assumption — the rent; confidence 60. After: the opponent showed that the number of customers matters more; my confidence in my initial view is 20.”

The audience must make the first record before the critique — later the assumption has already been said aloud, and the record would no longer show independent judgment. The lecturer takes the audience records into account (for example, as participation): what matters is not whether a student’s view matches the majority view, but what the change in it is based on. The audience records do not affect the defender’s grade.

A.6 What is assessed#

Assessed:

• how the problem is formulated (whether it is a question rather than a topic);
• whether the factors that determine the decision and their causal links have been identified; quantitative assessment and the use of methods from other topics count in the student’s favour, and quantitative assessment is required only when the task says so;
• how the subtasks are divided between the student and AI, whether the handover record is completed and at least one subtask not assigned to AI is named;
• whether the AI prompts are precise (log, A.3);
• whether all protocol steps have been carried out;
• whether the limits stated in the step 5 log are genuine rather than a mere formality. A mere formality: “The result may be inaccurate.” Genuine: “We do not know how many customers will come; with fewer than ~150 per day (taking the actual rent into account), operating cash flow becomes negative.”
• whether the verdict is justified and a review condition is given (A.4);
• how the conclusion is defended and how the critique is conducted (A.5).

In the basic task the problem is given, so the first criterion does not apply; the handover record, the log and the subtask not assigned to AI are assessed if AI was used, and the defence and critique are assessed if the task provides for them.

Scoring scale. Each applicable criterion is scored 0–2 points: 0 — not done or done only as a formality; 1 — partly done; 2 — done and justified. The result is the ratio of the points obtained to the maximum possible total for the applicable criteria; in the basic task the diagnostic criterion (see below) is included in this total as a separate criterion. The lecturer decides how the ratio is converted into a grade. When quantitative assessment is required by the task, the criterion on factors and causal links is scored at most 1 point without it. Additional questions in a chapter task are assessed as part of the criterion on factors and causal links and (in the basic task) of the diagnostic criterion.

Not assessed: the quality of the AI answer, which tool was used, and whether the student’s view is popular with the audience.

Direction of the verdict. In the extended task it does not matter whether the decision is “open” or “do not open” — what matters is whether the decision is consistently justified and whether what is unknown is honestly stated. In the basic task, what is assessed is not whether the student agrees with the lecturer, but whether they found the built-in error, understood how much it matters and took their decision accordingly.

Two separate requirements.

• Procedural minimum. Step 2 (source check) and step 5 (uncertainty log) must be clearly carried out. If they are missing, the work does not pass, however good the verdict may be.
• Diagnostic criterion (basic task only) — whether the built-in error was found: 1 point if it is identified, 2 if its effect on the conclusion is also assessed.

These requirements are assessed separately because they reveal different gaps: in one case, what had to be done was not done; in the other, subject knowledge was insufficient to notice the error.

A.7 Two conditions#

1. The use of AI is disclosed — as is also done in this textbook (see section 5 of the Introduction). The prompt log (A.3) fulfils this condition.
2. No confidential data are uploaded to an AI tool — internal company reports, personal data, trade secrets. For this reason, companies listed on a stock exchange or hypothetical cases (created for teaching purposes), such as the café in this appendix, are recommended for the tasks.

A.8 For lecturers#

Choosing the level of openness. For the extended task, level 3 is recommended at the start of the course, level 4 in the middle and level 5 at the end. However, the level is chosen by the lecturer, because students’ preparation and the order of chapters differ. Example: if the group is dealing with investment appraisal for the first time, it is worth starting at level 3 even in the middle of the course.

Limits of the task. The requirements of a task, its examples and the error built into the basic task must not go beyond the material set out in that chapter or in earlier chapters of the book; if the lecturer changes the order of chapters, the requirements are adjusted accordingly. Anything that cannot be calculated from this material is not required — a qualitative assessment is enough, and applying methods from other topics is the student’s own initiative. The task states whether quantitative assessment is required or only desirable. Example: if neither that chapter nor the chapters studied earlier teach how to calculate the required rate of return using the capital asset pricing model (CAPM), the task must not require it to be calculated that way; a qualitative assessment of risk factors or a method set out in the chapter is appropriate.

Perspectives and pairing. In the extended task the same company or case is usually analysed by a group or subgroup of students, but each student is given a different perspective — the angle from which the company or case is viewed (for example, in Chapter 1: core, risk, perception, strategy, intangible capital). In a large group of students several companies can be assigned — one to each subgroup. Defender–opponent pairs are best formed within the same subgroup so that their perspectives clash, i.e. one perspective exposes the weak point of the other. Example (café): a student who assessed the market (how many customers will come) acts as opponent to a student who assessed the finances (payback) — because the financial conclusion rests on the market assumption. Neither of them can answer alone, and this is exactly what makes the seminar valuable. Companies used as illustrations in a chapter are not assigned for an assessed extended task, because their examples already give part of the answers. Conclusions are drawn by the students; the lecturer may help but does not formulate them.

Seminar planning. It is practical to work in subgroups with no more students than there are perspectives in the chapter (five in Chapter 1), so that perspectives are not repeated. When there are many defences, subgroups work in parallel or some defences take place remotely (A.5). The work is passed to the opponent in advance (for example, at least two days before). In the open variant, where students choose different companies, pairs are formed so that the chapter concepts tested in the two works clash. For the first extended task the lecturer may drop some elements (for example, the audience records or the opponent’s summary) so that the workload is not too heavy.

Variant of the basic task. The solution for the café in this appendix is published and the chapters’ basic tasks are public. For an assessed basic task, the lecturer therefore prepares their own variant. They change the company, the figures and the built-in error without going beyond the limits of the task. They define in advance one main built-in error and distinguish it from additional weak points (this is needed for the diagnostic criterion, A.6). They do not publish the answer key.

Back-up question. The lecturer prepares in advance a question with which to test the defender if the opponent has not found the load-bearing assumption. Example: “What would happen to your verdict if the rent were 30 % higher?”

Logs and Moodle. It is practical to collect the prompt log (A.3) and the audience records (A.5) in a single form or Excel sheet, which students attach in Moodle at the end of the class. Assessment and feedback can then also take place after the class.

Staged submission and recordings. In Moodle it is practical to have the 0a and 0b records submitted and locked before the AI summary or the task data are released; the time for answering the counterfactual is short (for example, 10–15 minutes). Without a conversation this safeguard is only partial. Audio and video recordings are handled in line with the institution’s personal data procedures.