Skip to content
Interface language: English
Imperium / Signal

Business experiments: a hypothesis card that keeps the result from being fitted to the plan

Prepare a business experiment with an explicit hypothesis, comparison and decision condition. Assess the result without changing the rules afterwards or adding false precision.

Article language: English

A seated woman beside a glass sphere with dark reflections.
In this article

“Let's see if it works” sounds reasonable. The trouble begins when you wait until after the trial to decide what was supposed to work. A hypothesis card records an expectation, a way to check it and a condition for making a decision before the result is known. You do not need a lab coat. The spreadsheet with seventeen tabs can wait too. You need a question the trial will actually answer, and a willingness to accept an answer that is less convenient.

Here, a business experiment means a limited trial of a specific assumption. The procedure below is a suggested way for a service company to work. It does not turn a few conversations into representative research or promise a causal link where one cannot be demonstrated. It can help distinguish an observation from a wish and decide whether the next step is worthwhile.

Start with a decision that actually awaits you

Write down the choice facing the team. Should you change the service description? Prepare a larger pilot? Start conversations with a particular group of people? “We want to learn something” is too broad. Specify what you will do differently once you have the result. If no answer would change the plan, you may be organising a presentation rather than a trial.

Separate the decision from a solution someone already likes. A team may need to handle enquiries faster without yet knowing why delays occur. In that situation, buying an AI tool should not be written into the objective as the only possible answer. First, trace the process using the article on auditing automation (in Polish). Otherwise, you are testing a purchase against a problem nobody has identified.

Write a hypothesis without the magic word “better”

A useful hypothesis identifies a change, the situation, the expected observation and what it would mean. In an invented example: if we show the service scope before the form, people carrying out the task will correctly identify which materials to prepare for a conversation. You still need to define the group, how you will ask and the condition for assessment. “The new layout will be better” gives you no basis for comparison.

Also record why you expect that result. The reason might come from customer questions, an earlier observation or a known process constraint. Mark what is fact and what is an assumption. Printing a hypothesis on company letterhead does not turn it into evidence. You still need to check its basis and allow for the possibility that the mechanism works differently.

Choose the assumption with the most at stake

Not every unknown deserves the first experiment. Choose one whose rejection would change the direction of the work or prevent an expensive continuation. Whether a customer understands the service outcome may matter more than a button colour. Instead of comparing language models, you may first need to establish whether the company has a current source for its answers. If that assumption is hard to identify, a premortem conversation about possible project failure (in Polish) can help: the causes you name may suggest what to test first.

GOV.UK describes the alpha phase as a way to try different solutions and test the riskiest assumptions (in English). A prototype may cover only the part needed for the trial. This is guidance on how the British public sector develops its services, not a mandatory timetable for a Polish company. We use one narrow principle: first test the unknown that could undermine the whole plan.

Make a card that can be read before the meeting

On one card, record the decision question, hypothesis, basis for the expectation, scope of the trial and person responsible. Add a comparison, how observations will be collected and conditions for the next decision. You do not need a form longer than the process being tested. Each field should help resolve a specific uncertainty during execution or interpretation.

  • Decision: which next step depends on the result?
  • Hypothesis: what change is expected to produce what observation?
  • Scope: which people and cases does the trial cover?
  • Comparison: what will the result be assessed against?
  • Data: what will we record, and what will we deliberately leave out?
  • Boundary: when will we stop the trial or consider it inconclusive?
  • Responsibility: who will conduct the test and who will make the decision?

Atlassian's Project Poster (in English) separates the problem description, checking assumptions and preparing a solution. The document develops as the team learns. Our card is a separate, narrower tool for one trial; it does not reproduce the provider's entire workshop. It should preserve the reason for the test even when, a week later, everyone remembers a slightly different version of the idea.

Set the assessment condition before seeing the answers

Describe what you will consider a result that supports the hypothesis, what contradicts it and what leaves it unresolved. The condition should match the decision. In a handful of conversations, you might assess understanding of a description and recurring questions. Do not announce sales growth across the entire market on that basis. The conclusion must fit what was actually tested.

If you use a numerical threshold, record its rationale and limitations. Participant numbers, observation time and the comparability of cases all affect interpretation. Do not add a convenient threshold afterwards because the result fell just below your expectation. A changed condition may belong in a new trial, but it should not quietly rewrite the previous result.

Choose a comparison the company can manage

Not every trial is an A/B test with random assignment and enough traffic for statistical analysis. A small company can begin by observing people perform a task with two versions. Take care with order, familiarity with the earlier version and how the conversation is conducted. Describe these limitations rather than using a method's name as a stamp of scientific authority.

When checking a first impression, use the five-second test (in Polish). When studying interest in a product, see the article on validating demand in an Asian market (in Polish). Both trials can produce useful observations, but they answer different questions. Better understanding of a headline is not yet a confirmed order.

Protect the trial from too much help

The facilitator often knows the project so well that they supply missing information. The participant chooses the right answer and the team celebrates an effective screen that remains unclear without commentary. Prepare a neutral instruction and specify when a hint may be given. In the record, distinguish independent completion from a result achieved with help.

The same problem can arise with AI. An author may revise a prompt repeatedly until one excellent result appears. If the everyday process does not include that level of attention, account for it separately in the trial. The article on AI pilot criteria (in Polish) helps you consider human work, exceptions and how the result will be assessed. One successful demonstration does not yet describe an ordinary day for the team.

Aeneia conceptual photography

Collect observations that disrupt the comfortable story too

Record errors, missing answers and cases excluded from assessment, with the reason for each. Do not remove an inconvenient task simply because it disturbs the average. If a case was genuinely outside the scope, that should follow from a rule set in advance. Otherwise, it remains important information about a limit of the solution.

Separate observation from interpretation. “The participant asked for clarification of the scope” records an event. “Customers do not want this service” is a much broader conclusion. Show how you move from one to the other and which alternative explanations were checked. The problem may lie in the name, the order of information or recruitment, rather than in the offer itself. For a new version, also examine replacement of an existing service with a new one (in Polish): an existing customer's choice of the new offer does not yet explain whether a new use has appeared or the way an existing need is served has changed.

Recognise a trial that cannot resolve the question

Absence of the expected observation does not always disprove the hypothesis. Perhaps the participants were not in the agreed group, the task did not bring out the problem being studied or the results recording stopped working. Those conditions should be visible in the report. Do not, however, revise the rules after every inconvenient answer. First establish whether a specific execution error occurred, then decide whether to repeat the trial.

In a working example, you want to test whether a new description makes it easier to choose an option, but the facilitator explains the difference before showing the screen. The result then concerns the description with help, not the page alone. Keep the observation and prepare a neutral instruction for the next trial. You do not need to declare the whole project a failure. You do need to state honestly which question the material actually answers.

On the card, distinguish a result that supports the hypothesis, one that contradicts it and one that is inconclusive. For the last, give the reason and a condition for the next trial. This distinction protects against two shortcuts: treating every difficulty as a learning success and treating every failed demonstration as proof that the idea has no value. Either conclusion can run ahead of the available data.

Separate the cost of the trial from the cost of the whole solution

Record the work of preparation, execution, checking and subsequent maintenance. A small trial may use manual help that would be too demanding at full scale. That is a valid way to learn about the problem if the help remains visible. Do not present the tool's processing time as the time required for the whole service. The comparison should also show the work of people correcting results and explaining exceptions.

For AI, assess the total cost of the process, as explained in the article on process costs (in Polish). Break your own trial down into activities instead of adding a single unsupported aggregate amount. If a solution needs constant supervision, check whether the person who would provide it is available. A good presentation with an expert present does not prove the arrangement is feasible when that expert moves to another project.

Set a limit on further investment too. Who can expand the trial, what information must they receive and what justifies stopping? This is not an obligation to stick with the first idea. It protects a decision against the gradual expansion of a test that, after a few weeks, begins to resemble an implementation nobody consciously approved. A small experiment should genuinely be able to remain small.

Work through a mixed result

Imagine a service company testing a new description of an initial consultation. Participants identify the materials they need more accurately, but some expect a complete implementation after one meeting. This is an invented example. The first observation supports part of the hypothesis; the second reveals a significant risk in the promise. There is no reason to turn both into one enthusiastic satisfaction percentage.

The team revises the sentence about the outcome and plans another trial focused on that boundary. It keeps the earlier result, the version of the description and the new question. In its project decision register, discussed in the article on project decision registers (in Polish), it records why it is not expanding the campaign yet. The mixed result leads to a more precise action instead of an argument about whether the test generally “worked”.

Turn the result into the next move

After the trial, choose whether to continue, change the scope, check something further or stop. For each decision, identify the observations, limitations and responsible person. If more information is needed, name the specific gap. “Let's do a little more research” can be a very busy way of postponing the same decision.

Take lessons about how the trial itself was conducted into a project retrospective, as described in the article on project retrospectives (in Polish). If the result justifies commissioning more substantial work, prepare a supplier selection brief, described in the article on supplier selection briefs (in Polish). The card then need not end its life as a neat file in a “Research” folder. It becomes a basis for a next step that can be checked.

Keep a short summary for someone who was not there

The report should let someone reconstruct the question, conditions and decision without attending the entire trial. Record what was done, what was observed and what the result does not explain. Attach the relevant version of the material. A person approving the next stage should also see unsuccessful cases if they affect the choice.

Read the summary after a short break. Can you distinguish a record of fact from interpretation? Does a statement about improvement say what is being compared? Does the next step have an owner? If the report works only when its author adds an oral explanation, fill that gap before handing it over. This is the card's last test: the knowledge should remain in the company when the person conducting the trial moves to their next task.

Which trial should you start with?

Choose one assumption on which the current project depends. Record the decision, a simple scenario and the assessment condition, then ask someone outside the idea to read the card. If they do not understand which result would change the plan, improve the question before starting. That costs less than interpreting a well-documented answer to the wrong question.

For help organising a process or AI trial, explore the scope of Aeneia Audit. A review concerning a website can be combined with Aeneia Design.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear