Last updated: 4 September 2026
A customer insight reaches a decision when it arrives attached to three things: a decision that is already scheduled, a named person who can make that call, and a statement of what evidence would change the answer. Findings that arrive without those three get read, complimented, and filed.
In 2020 the US Government Accountability Office surveyed roughly 4,000 federal managers across 24 agencies. An estimated 95 percent had at least one type of evidence for their programs. Only about half to two-thirds reported using it in decisions such as allocating resources.
Supply was close to universal. Use was not.
That gap is not a communication problem, and better slides do not close it. Grimshaw and colleagues concluded, in a widely cited review, that passive dissemination of research findings is largely ineffective. Translation works when the strategy is chosen against the specific barriers between the evidence and the people who act on it.
Circulating a deck is not a strategy. Naming the barrier is.
Where Research Stalls Between Fieldwork and Decision
Research rarely fails at the fieldwork stage. It stalls in the handoff, in three recognizable places. Either the study was commissioned without a decision attached, or the decision window closed before delivery, or the finding is true but not falsifiable, so nobody can tell what acting on it would look like.
Commissioning is the first. A study gets approved because a question is interesting, not because a decision is pending, and nobody notices until delivery. The federal evidence machinery has the same failure and has documented it: GAO's 2019 review found agencies had reasonably strong processes for assessing evidence they already held, but weaker ones for prioritizing which new evidence to generate, when, and how. Knowing what you know is easier than deciding what is worth finding out.
Timing is the second. Decisions have windows, and the window usually closes before the report is polished. A directional read inside the window beats a definitive one after it, and teams that miss this produce excellent evidence for decisions already made.
Falsifiability is the third, and the one researchers control. A finding that cannot be wrong cannot be acted on. "Customers value trust" survives any outcome and therefore changes none. "Mid-tier buyers abandon at the plan comparison step because they cannot tell what the middle plan adds, and eleven of nineteen said so unprompted" can be wrong, which is what makes it useful.
What Separates a Finding From a Customer Insight?
A finding describes what respondents did or said. A customer insight names the decision it changes and the direction it pushes that decision, and it carries enough method with it that someone can judge how much weight it holds. The difference is not depth of analysis. It is whether the sentence points at an action.
Three tests separate them, and a finding has to pass all three.
- The decision test: name the decision, and name it specifically enough that someone could put it on a calendar. Not "improve onboarding" but "whether to rebuild step two before the Q1 launch."
- The direction test: say which way the evidence pushes, in one sentence. If a recommendation needs three qualifiers to survive, the evidence is not strong enough to carry one, and saying so is the honest output.
- The provenance test: the finding travels with its own method. The American Association for Public Opinion Research publishes disclosure standards covering what must be released alongside results, including the population studied, how the sample was recruited, the modes of collection, the field dates, and the exact question wording. Those items let a reader discount a result appropriately. Stripped out for readability, they leave a number that looks more certain than it is.
The third test is the one that gets skipped, and skipping it produces the confident wrong decision below.
Who Owns an Insight After the Study Closes?
By default, nobody, which is the whole problem. Ownership splits across three roles that have to be assigned explicitly, and written down before fieldwork rather than after: the researcher owns the evidence and its limits, the decision owner owns the call, and a third person owns the follow-through.
The first is a professional obligation rather than a preference. The ICC/ESOMAR International Code makes researchers responsible for how findings are presented and for not allowing results to be reported misleadingly, and the Insights Association Code of Standards sets comparable duties in the US. In practice the researcher is accountable for the caveat surviving summarization, which is hard when a finding is retold third-hand.
The second, the decision owner, is whoever holds budget or roadmap authority over the thing being decided. If that person did not know the study existed, the study was commissioned wrong. It is a routing problem before it is an org design problem.
The third role is the one almost nobody assigns. Not the decision, the aftermath: whether the change shipped, whether the metric moved, whether the finding held. In a small insights function all three collapse into one or two people, which works until the team grows and the roles stay merged. The questions a mid-sized insights team should ask a platform look different once the constraint is translation capacity rather than fieldwork capacity.
How Should a Finding Be Structured for a Decision Maker?
Structure it around the format the decision actually runs on, not the format the research team enjoys producing. Every delivery format buys something and costs something, and the failure mode is picking one by habit. The table sets out that tradeoff.
| Delivery format | What a decision maker can act on | Where it fails | Best used for |
|---|---|---|---|
| Full report or deck | The complete argument, method and limits attached | Length is a filter. The people who most need it read the summary slide | Long-horizon calls, and evidence re-read by people not in the room |
| One-page decision memo | One recommendation, the evidence for it, the case against it | Compression hides the disagreement inside the data | A decision already scheduled, with a named owner |
| Live dashboard access | Their own follow-up questions, when they think of them | Access is not interpretation. Unattended dashboards get opened once | Teams who know the study and keep querying it |
| Verbatim or video evidence clips | The texture behind a number, in a form that survives retelling | Vivid quotes travel further than representative ones | Moving a room that saw the chart and did not move |
| Searchable research repository | Whatever a previous study already answered | Retrieval only helps if someone thinks to look | Organizations re-commissioning studies they have already run |
| Working session with the decision team | The decision itself, in the room, with the evidence open | Costs the most senior calendar time available | Contested calls where the argument is about interpretation, not facts |
When a Specialist Tool Beats a Research Vendor
When the real problem is retrieval rather than evidence. If an organization cannot find what it already learned, a dedicated repository is the better purchase: a standalone tool such as Dovetail for research collected elsewhere, or an insights layer that already holds the studies, which is what the next section describes. Better fieldwork does not fix a retrieval problem.
Can the Claim Be Interrogated?
That is what the delivery format decides. Alchemic's insights layer drills any claim back to the respondent, verbatim or voice clip it came from, and answers plain-English questions from Slack, Microsoft Teams or WhatsApp. That is a structural point, not a quality one: a finding a stakeholder can pull on is harder to dismiss and harder to overstate.
Confident Decisions Built on Partial Evidence
The most expensive failure is not the insight that never reached a decision. It is the insight that reached the decision cleanly, was acted on with conviction, and rested on a sample missing the people the decision was actually about. Coverage is what decides whether an insight generalizes.
Who Never Enters an Online Sample?
More people than most US teams assume, and the exclusions are measurable. Pew Research Center notes that about 7 percent of US adults do not use the internet, 16 percent are not digitally literate, and half cannot read above an eighth-grade level. Those adults skew older, less formally educated and more rural.
The federal skills assessment agrees. In the 2023 PIAAC results, 28 percent of US adults scored at Level 1 or below in literacy, up from 19 percent in 2017, and 34 percent scored there in numeracy. A self-administered online questionnaire is a reading test before it is a research instrument.
How Much of an Opt-In Panel Is Real?
Less than the panel reports, and no screen fixes it cleanly. In August 2026 Pew published an analysis of 11,114 US adults surveyed through an opt-in panel in November 2024, screened three ways. Before screening, 88 percent of open-ended answers were unproblematic, leaving roughly 12 percent that were not: non sequiturs, generic positive ratings, gibberish and probable AI text.
Every remedy cost something. Trap questions flagged 18 percent of cases, automated prescreening removed nearly half the sample, and voter-file matching increased error by removing mostly good respondents. Three fixes, three new problems.
None of that argues against online fieldwork. It argues for knowing which mode produced your sample before a decision rests on it, the same discipline set out in who gets missed in AI-moderated samples and where good survey respondents come from.
Which Channels Reach the Rest?
Channel choice is what moves that boundary. Interviews run natively inside WhatsApp with no link and no app, and outbound AI phone interviews reach people a browser session never gets to. That includes those who will answer a voice note but not a form, and those whose written fluency sits below their spoken fluency.
Alchemic publishes 57+ languages including Hindi, Tamil, Telugu, Arabic, Spanish and Portuguese, recruits through its own panel network or a client's own list or a hybrid top-up, and has fielded studies across 14 markets including the USA and the UK. More channels is not the point. The channel sets the frame, and the frame decides who the insight is true of.
What Should Happen in the Two Weeks After Delivery?
Three things, and only the first is common practice: write the decision record, reread the disconfirming evidence you registered, and close the loop on a fixed date. The date gets set when the decision is made.
The decision record holds five fields. Decision, owner, action taken, the metric that will move, and the revisit date. Most teams get the first four right and treat the fifth as optional, which is where the practice quietly dies. The revisit date is what turns a decision log from a filing system into a learning one.
Registering the disconfirming evidence has to happen before fieldwork, not after delivery. Writing down what result would have changed the recommendation costs ten minutes and removes the option of reinterpreting data once the outcome is known. It also produces the next study's question for free.
Closing the loop is easier when a follow-up is cheap. Alchemic's standard benchmark is a 200-interview qualitative study from brief to live dashboard in three days, which changes what a revisit can mean: another wave becomes a reasonable answer to an ambiguous result rather than a next-quarter proposal. That matters most for brand tracking with qualitative follow-ups, because a tracker that reports movement without explaining it raises a new question every wave.
Where This Approach Falls Short
Plenty of decisions are not made on evidence at all, and no amount of routing changes that. A commitment already announced, a founder's conviction, a closed budget cycle: research arriving into any of those is documentation, not a decision input. Recognizing which one you are in is worth more than a better memo.
Attribution is weaker than it looks. Claiming a study caused a decision is rarely falsifiable after the fact, and the best-known figure in this area shows the trap.
Morris, Wooding and Grant reviewed 23 papers quantifying research translation lags and found estimates from zero to 306 years, depending on which two points were measured. The 17-year average everyone quotes hides exactly that variation, which is the authors' point. Insight-to-decision lag behaves the same way: one number for it is a story, not a measurement.
The practice can also be over-instrumented. A decision record for every finding produces compliance rather than use, and the log ends up unread. Reserve it for decisions large enough that being wrong is expensive.
Finally, how much translation work lands on the buyer varies with the operating model. Self-serve tools generally hand back transcripts, themes and a dashboard, and turning that into a recommendation stays in-house, which is the right trade for a team that has the analysts and wants control. Where that capacity does not exist, the alternative is a model where the vendor carries it.
Once it has the client's brief, Alchemic designs and tailors the study to the decision at stake before fielding, as part of end-to-end consumer research at scale, rather than leaving that work to the buyer. Neither is better in the abstract. They fail differently, and the question is which failure your team can absorb.