Orbital update: better evidence for brand decisions

In brief

Orbital Brand Science uses consumer research and behavioural measurements to help Caribbean businesses decide what to change in a product, package or message. Our work combines EEG, which records electrical activity at the scalp, with eye tracking, sensory and product testing, surveys, interviews and focus groups.

We are developing a research assistant agent alongside that work. It will help researchers find relevant material, check source details and organise evidence for review. We will assess its value through the quality of the resulting work and the time needed to check it.

Connecting the measurement to the decision

Our service portfolio brings several methods to the same commercial question because each records something different. A person can look at a label without understanding it, or understand it without wanting the product. Their account of the experience adds information that a gaze trace alone cannot supply.

For a brand team, the study begins with a change the business can make. That might mean choosing between package designs, testing whether instructions are understood or assessing how customers describe a product after trying it. The Orbital science overview introduces the methods; our research programme examines how the measurements support those decisions.

The assistant's role fits around the researcher: retrieve the source, preserve its sample and outcome, and flag a gap that needs further reading. It must keep evidence from customers separate from examples generated by a model. Our current development work concerns that support role, with people responsible for the research design and the final recommendation.

What the evidence says about AI assistance

Research gives us reasons to test an assistant carefully. In an experiment involving 758 consultants, the Harvard Business School team reported that AI assistance on tasks within the model's capabilities increased speed by over 25%, human-rated performance by over 40% and task completion by over 12%. These are results from that experiment, not measured gains at Orbital. HBS research summary, 2023 study.

Noy and Zhang's published writing experiment involved 453 college-educated professionals. Time spent on the assigned tasks fell by 40% and evaluated quality rose by 18%. It concerned bounded writing work; it did not establish that a model can replace customer research or discover an unmet need in a Caribbean market. Noy and Zhang, Science, 2023.

The April 2023 working paper Generative AI at Work studied 5,179 customer-support agents. It reported a 14% average increase in issues resolved per hour, with a 34% improvement among novice and lower-skilled workers and little effect on experienced, highly skilled workers. Stanford working-paper record.

A short-story experiment adds another concern. In Doshi and Hauser's study of 293 writers, access to up to five AI ideas raised evaluated novelty by 8.1% and usefulness by 9%, compared with the no-AI condition. Yet AI-assisted stories were more similar to one another. Those outcomes concern the study's writing task, rather than advertising effectiveness or sales. Doshi and Hauser, Science Advances, 2024.

A Caribbean brand team needs to test whether an assistant helps staff produce useful options while keeping those options distinct. Average speed alone cannot answer that question. Our evaluation will examine the evidence behind each proposal and the errors that survive review.

What this means for a Jamaican brand

Imagine a Jamaican food business choosing between two packages. This is an illustrative test, not a report of a client engagement. One version makes the product name easier to notice; another explains preparation more clearly. A gaze trace can help show what participants looked at. A comprehension question can reveal whether they understood the instructions. A purchase test asks a different question again.

The research design should follow the decision the business can actually change. We apply Adrian Dunkley's People First, Problem Second, AI Last approach by naming the customer need and the commercial decision before choosing an instrument. If the problem is that people cannot find the price, a clear comparison of label designs can be more useful than a large collection of signals with no agreed interpretation.

Local context enters through the sample and the task. A Kingston supermarket purchase, a parish shop visit and a product sent to relatives overseas involve different constraints. The design should state where participants shop, how they encounter the product and which customers were not represented. A national label on a small convenience sample would hide a limitation that the buyer of the research needs to see.

AI can help prepare a research plan or organise existing material, but a generated response is not a customer interview. We will keep synthetic examples separate from responses collected from people. A business using generated personas to rehearse questions still needs evidence about whether those questions make sense to its intended customers.

Measuring a signal without inventing its meaning

Our research programme needs clear boundaries around each measurement. Eye tracking records estimates of gaze. Facial coding describes patterns of visible movement. These measurements become useful through a defined task and an evidence-based interpretation. Calling all of them direct readings of purchase intent would erase important differences.

The caution is especially strong for facial movements. Barrett and colleagues reviewed the evidence behind inferring emotional states from a face and challenged the assumption that a particular facial configuration reliably identifies a particular emotion across contexts and populations. Their review gives us a reason to preserve context and validate an inference, rather than treat a software label as the participant's emotional truth. Barrett and colleagues, 2019 review.

For our work, the practical consequence is to write the inference beside the observation. If a report says that a package drew attention, show how attention was measured. If it says that people preferred the product, identify the preference question or behaviour. If the client wants a sales claim, the evaluation must measure sales or explain why the current study cannot answer that question.

The research assistant agent's job

We are developing the assistant around tasks a researcher can inspect: finding a relevant source, extracting the described sample, comparing how studies define an outcome and preparing a summary that links to the underlying material. The researcher must be able to inspect each of those steps.

A useful answer should identify which sentence comes from a source and which part is our interpretation. It should retain contradictory findings and indicate when only an abstract was available. The researcher then decides whether the source fits the question. Saving reading time is valuable only if the assistant does not hide the step where an unsuitable study would have been rejected.

That is also a practical response to Cognitive Debt, Adrian's term for the risk that builds when people automate decisions without learning from them. In this setting, a source-linked summary should help a team inspect the argument. A polished conclusion without the evidence would make that work harder.

A test your brand team can run this week

Here is a test we suggest for a Jamaican brand manager evaluating an AI research assistant. It is an evaluation design, not a completed Orbital result.

1. Write one decision the team needs to make, the customer group affected and the evidence that would change the choice. Keep a copy before using the assistant.
2. Compare 10 source summaries over the next 14 days. Give a researcher the originals and record every incorrect number, missing limitation and unsupported conclusion in the assistant's output.
3. Measure review time as well as drafting time. Include the time spent correcting errors, and record whether a corrected summary changed the proposed decision.
4. Check whether the proposed options remain distinct. Ask a reviewer who did not create them to describe what would make each option useful to a different customer need.

The test can begin with public research and non-sensitive product information. Customer interviews and biometric recordings should enter the workflow only after the team has agreed the purpose, access and handling process with the relevant people.

Risks for customers and research teams

A false interpretation can affect both the customer and the business. A participant's facial movement might be labelled as dislike; a small sample might be reported as a market; an AI summary might omit a result that changes the recommendation. The business then spends money on an answer that the evidence did not support. We will guard against those failures by separating observations from inferences, stating sample limits and requiring source checks before a recommendation is released.

Privacy also shapes the design. A recording collected for one agreed study should not become training material for an unrelated tool through an unnoticed upload. The team must identify who can access the data, which system receives it and what the participant was told. The main unresolved question for our assistant work is how much useful time it saves after that verification and handling work is counted. We need a local evaluation before claiming a return.

What to expect from the series

Our next work connects the assistant to the evidence a brand team needs, then tests whether its summaries preserve the findings and their limits. The accompanying research series will examine attention, comprehension, consumer sampling and sensory tests. We will also contribute to the Caribbean AI Association's think-tank discussions on responsible AI.

Follow the Orbital research archive for the series and StarApple AI for the wider applied AI programme. The next useful milestone is an evidence trail a brand manager can inspect and a decision test the team can repeat.