A spreadsheet with 300 enthusiastic customer responses can make a new product look ready for investment. If all 300 responses were generated by AI, the company has not spoken to 300 customers. It has produced 300 model outputs. That distinction belongs in the investment proposal, before a forecast becomes a purchasing commitment.
For Norwegian leaders, the decision is where simulated customer research is useful and what evidence is needed before committing money to a launch. My recommendation is to use AI to sharpen the questions, explore objections and prioritise what to investigate. Require separate evidence from the relevant market for claims about demand. More generated answers do not, by themselves, provide more independent observations of customers.
Find out who actually answered
An AI tool can be asked to act as a purchasing manager, a busy parent or a small-business owner. The resulting answer may sound plausible and specific. A service may also build simulations from interviews or survey responses collected from real people. Those are materially different starting points.
Before using either result, ask what was observed and what was predicted. Was a customer directly asked this question? Did a model infer the answer from older interviews? Or was the entire customer profile invented for the exercise? A realistic quotation is still generated text unless someone actually said it. Do not turn it into a testimonial or present it as an interview finding.
Keep three categories separate in the report: what people told us, what people actually did, and what the model predicted. None is automatically a complete picture of the market. A real interview about willingness to pay is also different from an accepted offer or a repeat purchase.
The research shows both limits and promise
Pew Research Center reported on 30 September 2026 that AI-generated responses differed from human responses by an average of about 12 percentage points across three US survey waves. This measured average absolute differences in answer shares, not a 12% rate of wrong individual answers. The study concerns public opinion, not Norwegian purchasing behaviour.
Pew's methodology used profiles of actual participants, including prior political-attitude responses, rather than invented personas alone. The principal results used Claude Opus 4.6. Most questions were political. The findings do not establish the performance of every model or every simulation method.
There is also evidence that richer human data improves simulation. A study of 1,052 Americans, revised on 28 June 2026, found that agents built from interviews and surveys reached 86% of participants' own two-week response consistency on General Social Survey questions whose answers had been withheld from the agents. That benchmark accounts for people giving different answers when asked again; it does not measure purchase prediction. The English-language US study does not establish performance for Norwegian customers.
Gallup is testing simulations with Simile, a commercial partner, using real interview and survey data. Its methodology article, first published on 11 May 2026 and subsequently updated, describes encouraging early results near the original interview topics. It also says validation is continuing and simulated answers will not replace direct measurement in its tracking research. This is not independent confirmation of a particular product's suitability for your market.
My conclusion is that the label “synthetic research” settles very little. The underlying human data, the question being predicted and validation against relevant people determine whether a result deserves weight. These studies give reasons to investigate the method, not permission to label generated respondents as customers interviewed.
Test the product, not just the description
Consider a hypothetical Norwegian manufacturer evaluating returnable transport boxes for small food producers. Management is considering NOK 400,000 in tooling. A team asks AI to simulate 300 bakery and catering buyers. The generated responses favour the proposed box and praise its return scheme.
The exercise may reveal useful questions. Will the box fit existing shelves? Who stores the empties? Who pays for collection? How much extra handling will staff accept? But a positive response from an invented buyer does not show that a bakery has room for the boxes or will pay the proposed price.
The next step should investigate the condition that could make the investment fail. Let relevant buyers handle a prototype in their own workflow, explain their current alternative and assess an actual offer with price, delivery and return terms. Include prospective customers beyond the company's most enthusiastic existing contacts. Record refusals and the reasons for them.
A small number of interviews is useful for discovering objections; it is not a representative estimate of Norwegian demand. A paid trial provides stronger evidence of willingness to try at that price, but does not establish repeat demand or profitable operation at scale. Describe each finding at the level the evidence supports.
For this manufacturer, storage and returns might matter more than the wording of the product description. Asking AI to generate another thousand favourable answers will not resolve those questions. A limited physical trial may do so before the full tooling commitment.
Put the evidence beside the commercial claim
Ask the commercial owner to keep a short evidence table with the launch proposal. The following is a proposed format for the hypothetical manufacturer, not a validated research standard:
| Claim being assessed | Evidence needed | What a simulation contributes |
|---|---|---|
| Buyers have a problem worth solving | Relevant buyers describe current costs and workarounds | Candidate problems and interview questions |
| The box fits their work | Buyers try the prototype in realistic conditions | Possible obstacles to investigate |
| Buyers accept the offer | Responses to a concrete price and terms, with refusals recorded | Hypotheses about price objections |
| Demand justifies tooling | A defensible demand estimate and economics, with uncertainty | Scenarios to test, not observed orders |
Decide in advance which finding would change the product, limit the launch or stop the investment. Management may still choose to proceed with uncertainty. It should state the assumption it is taking a risk on, rather than letting a large synthetic sample make that uncertainty disappear.
In a small AS, the commercial owner and product lead may be the same person. They still need to distinguish the generated material from customer records. A larger business commissioning research should require the supplier to disclose the origin, age and coverage of the human data, the model and method, the validation and which reported results are simulated. Ask whether performance was checked on questions and people that were not used to tune the method. A demonstration selected after seeing the results is weak assurance.
Make customer research faster without removing the customer
AI is worth testing in preparation: suggesting alternative explanations, finding gaps in an interview guide and helping organise permitted interview material. Have a researcher check summaries against the original responses, including contradictory and minority views. Measure whether the team reaches a better-informed decision sooner, including review time and research cost.
For a Norwegian market, the research must cover the actual buyers and buying process. Fluent Bokmål and a profile labelled “Norwegian purchasing manager” do not establish that the simulation reflects local contracts, distribution, geography or purchasing authority. If personal or confidential customer information is involved, use approved tools and resolve the permitted processing before uploading it.
Do not make perfect evidence a prerequisite for every small experiment. A cheap, reversible concept test can be proportionate even when uncertainty is high. A major tooling commitment or a demand forecast used to raise finance needs a more defensible basis. The AI investment guide explains how to weigh investment against the cost of waiting. The pricing article addresses the separate decision about the value the customer pays for.
Management owns the research brief, validation and commercial decision. Where the investment is material, the board should ask how much of the demand case comes from actual customers, what behaviour has been observed and which assumption remains untested. Directors do not need to run interviews or select the model.
Before approving the next market study, ask one plain question: How many relevant people actually answered, and which answers were simulated? That answer helps the company buy useful learning instead of an impressive volume of artificial certainty.
Sources checked on 25 October 2026. The manufacturer, 300 responses, NOK 400,000 investment and evidence table are hypothetical. The proposed working methods are the author's recommendations. The illustration is AI-generated.