Building an In-House Insights Team in 2026

Last updated: 1 September 2026

An insights function is accountable for the quality of decisions made under uncertainty. It is not accountable for the number of studies it runs, and confusing the two is the single most common reason a new function is judged a failure while producing a great deal of work. The teams that survive their second budget cycle are the ones whose output can be traced to decisions that were made differently.

That framing determines the first hire, and it is not the hire most companies make. The instinct is to recruit someone who can run studies. The scarcer skill is different. It is deciding which questions are worth answering at all, and having the standing to tell a stakeholder that the study they asked for will not change what they do.

Almost everything else is a question of what stays inside and what does not, and the honest answer is that a small function should own the questions and the interpretation, and own very little of the fieldwork.

What Is an Insights Function Actually Accountable For?

Four things, in descending order of how often they are neglected.

• Deciding what is worth knowing. Triaging incoming requests against the decisions they would inform, and declining the ones that would not change anything. This is the function's core product.
• Choosing a method that fits the question. Most bad research is a good method applied to a question it cannot answer.
• Owning interpretation. Translating findings into implications a decision-maker can act on, including saying when evidence is too thin to support the conclusion someone wants.
• Maintaining institutional memory. Preventing the organization from paying twice for the same finding, which is the cheapest value a function delivers and the first thing lost when it is busy.

Running fieldwork is not on that list, which is deliberate. Fieldwork is a capability that can be bought at consistent quality. Knowing which question to ask cannot. How that looks for a smaller team is covered in consumer insights platforms for mid-sized teams.

Who Should the First Hire Be?

Someone senior enough to say no, with enough method breadth to know what each approach can and cannot deliver.

The common mistake is hiring an executor as the first researcher: someone who runs surveys competently and lacks the standing to challenge a request from a marketing director. That person will be fully occupied within a month, delivering studies that answer the questions they were handed rather than the questions the business has.

Three specific capabilities matter more than years:

• Method breadth over method depth. A first hire needs to know when a qualitative approach is required, when a trade-off design is required, and when the answer is already in the CRM. Deep expertise in one method is what vendors are for.
• Sufficient standing to decline. Reporting line matters more than title. A researcher reporting into the team that commissions most of the research cannot decline it.
• Translation. The ability to move between a stakeholder's business question and a research design, in both directions, is the rate-limiting skill.

How Much Research Can a Small Team Actually Run?

Less than expected, and the constraint is analysis rather than fieldwork.

Fieldwork scales by buying it. Analysis and interpretation do not, because they require the person who understands the business context, and that person is the bottleneck by design. A single researcher can meaningfully own a handful of substantial studies a year alongside continuous stakeholder work, and teams that promise more than that usually deliver the fieldwork and skip the interpretation, which converts research into a reporting service.

How Do You Know When a Study Is Finished?

Sample sizing helps here more than most people expect, because knowing when a study is finished prevents a great deal of unnecessary work. A published method for assessing thematic saturation applied bootstrapping to three interview datasets. Saturation came at six interviews in two of those datasets and eight to nine in the third, using a threshold of no more than five percent new information across a run of two.

Those figures describe relatively homogeneous groups with narrow aims. The federal review of participant numbers in evaluation research names the three factors that push them up: number of distinct groups, heterogeneity within them, and complexity of what is being studied.

The practical use of that evidence inside a small function is defensive. The saturation paper also reports earlier work in which 70 percent of 114 identified themes appeared in the first six interviews and 92 percent within the first twelve, which is the clearest published answer to a stakeholder asking for more. Faulkner's 2003 usability test of 60 participants is the other half of the argument: random groups of five found 85 percent of problems on average but as little as 55 percent at worst, which is the case against cutting below that floor.

Together they give a researcher a published basis for stopping a study at a defensible point, rather than running a larger sample because a stakeholder finds a bigger number more convincing. The distinction between a tool and a service is drawn in platforms that run the study for you.

What Should Stay Outside, Even With a Team?

Best kept outside even at large scale: multi-market fieldwork. Very large organizations that have built internal fieldwork capability in one market almost always buy it in the others. The fixed cost of local recruitment, incentive rails, language coverage and regulatory compliance does not amortize against a single company's volume. This is a case where the in-house instinct is simply wrong, and it stays wrong as the team grows.

Whichever side of the line a capability sits, the professional standards governing respondent treatment, consent and disclosure apply to the work rather than to the supplier. The ESOMAR code and guidelines and AAPOR standards and ethics materials are what an internal function should hold both itself and its vendors to.

Which Skills Do Not Transfer From Marketing?

Insights functions are frequently staffed from marketing, which is reasonable and carries three specific hazards.

Advocacy versus assessment. Marketing rewards making a case. Research requires being genuinely open to the answer being unwelcome, and the habit of building a case is hard to switch off.

Sample intuition. A marketer reading "eight of ten people said" hears a proportion. Reading it correctly as qualitative signal rather than measurement is a trained skill, not an obvious one.

Question construction. Writing a question that does not lead the respondent is harder than it appears, and confidence in writing copy transfers poorly, since good copy is persuasive and a good research question must not be.

Who Is Missing From the Evidence Your Function Runs On?

A function's standing sample frame quietly becomes the organization's picture of its market, and skew inside it compounds across every study.

Pew Research Center's mobile technology fact sheet reports 16 percent of US adults as smartphone-only internet users. Pew puts that at 34 percent among adults in households under $30,000 a year against 4 percent above $100,000, and the ITU's connectivity statistics show a considerably steeper gradient across markets. Pew also puts overall smartphone ownership at 91 percent of US adults, so the gap is not device access but the broadband and privacy a research session assumes.

A function running everything through one browser-based panel will accumulate a consistent, invisible bias toward more connected and more affluent respondents, and consistency makes it harder to spot rather than easier.

Two cheap defensive habits: weight or at least compare recruitment profiles against a public frame such as US Census population data, and vary the fieldwork channel deliberately rather than defaulting to whichever one is already contracted.

Where a study needs to reach past the standing panel, the channel is the variable that moves it. Alchemic runs interviews natively inside WhatsApp with no link and no app to install, and by outbound phone call for respondents reachable by voice rather than browser. Recruitment runs as managed fieldwork or bring your own, so an internal team can use its own customer list where it has one and a managed recruit where it does not. The same selection effect is examined in sample validity and who you miss.

Where In-House Functions Stall

Four failure modes, all recoverable if named early.

Becoming a reporting service. The function ends up producing dashboards and wave summaries nobody acts on. The symptom is that requests arrive specifying a method rather than a question.

No authority to decline. If every request becomes a study, capacity is consumed by low-value work and the function is judged on throughput.

Memory that lives in one person. Institutional memory that is not written down leaves with the researcher, and the organization pays again for findings it already owns.

Measuring the wrong thing. Studies delivered, respondents interviewed and reports published are all easy to count and none of them is the objective. A function measured on volume will produce volume.

A fifth stall is budgetary and arrives in the second year. A function justified on cost savings against agency fees is measured on those savings, and the saving is easy to reverse: cutting the function restores the agency line and looks neutral on a spreadsheet. A function justified on decision quality is harder to cut because the counterfactual is a decision made badly. It is also harder to evidence.

That is why the framing has to be established from the start rather than argued during a budget round.

Where the organization needs an external reference point for what it is buying either way, the US Census Statistics of U.S. Businesses gives a public sense of how fragmented the professional services supply base is. That matters when the alternative to building is a market of many small suppliers rather than a few large ones. Which model to buy from is worked through in research agency against platform and managed service.

The honest counterweight is that not every organization needs a function. A company making a handful of significant decisions a year is usually better served by buying research well than by hiring someone to run it, and recognizing that is a legitimate answer rather than a failure to invest.

Teams working out which methods to commission will get more from the concept testing overview and the qualitative research primer than from another team-structure model.