← All posts
AIinsight agenciesmarket research

Will AI replace research agencies? The work that changes and the value that remains

Sajin Kabeer · November 9, 2026

AI will replace some tasks that research agencies used to charge for. It may also squeeze agencies whose offer is an undifferentiated survey, a stack of transcripts and a deck. But it does not follow that every agency disappears: clients still need someone to ask the right question, check who and what the evidence represents, and make a recommendation they can defend.

We hear the replacement question from inside the agencies we work with. The uncomfortable version is not “Can AI draft a report?” It is “Why would a client pay us for a report they can generate themselves?” The answer cannot be that researchers type more slowly. It has to be a better research decision.

Disclosure: Meaningful builds tools for insight agencies. This is our view of how their offer may change, informed by the linked research and industry perspectives—not a forecast of agency employment or a measured claim that AI has already replaced a fixed share of work.

Which part of an agency is under pressure?

The most exposed work is repeatable production sold as if the labour itself were the product. A client who needs a quick standard survey, a first pass at open ends or a descriptive slide deck now has more ways to do parts of that work in-house. That does not make every self-serve answer decision-ready; it changes what an agency must demonstrate beyond delivery.

The MRS Delphi Group's Who owns understanding? report (2026) brings together practitioner perspectives, not a representative survey of agency outcomes. Its commercial-structure contributors argue that as process becomes cheaper, agencies will need to make expertise, interpretation and decision impact more explicit in their contracts. That is a sharper claim than “AI makes research faster”: the invoice has to reflect what the client can no longer get by pressing a button.

“80% of the insight for 20% of the cost” is a way of describing one buyer's tolerance for a good-enough answer, not a measured market ratio. Some decisions genuinely need only a fast directional check. Others carry enough brand, product or policy risk that a plausible summary without sound sampling and contrary evidence is a false economy. Agencies should be honest about which job they are bidding for.

What work can AI make faster?

AI can help prepare, organise and draft across a research project. The gain is real only when the agency counts the time spent correcting errors and securing approval, not just the time to the first output.

Work that may accelerateWhat a researcher must still verify
Transcribing and summarising interviewsSpeakers, meaning, missing passages and any quotation used with a client
Suggesting codes for open-ended responsesCodebook fit, minority and negative cases, respondent-level counts
Cleaning and tabulating survey dataQuestion wording, bases, weights, exclusions and changes between waves
Bringing surveys, interviews and prior studies togetherWhether sources describe the same population, period and decision
Drafting a readout or deckThe strength of each claim, its caveats and what the client should do next

“Days of work, now hours” can be a useful ambition, not a service-level promise. A fast transcription pass can create a slower review if speakers are confused. An attractive deck can be expensive if the agency has to reconstruct every base and source link. Our first-30-days guide sets out how to measure net time and quality on a finished, permission-cleared project.

What can agencies do with the time they get back?

They can ask a wider question, investigate disagreement and test a conclusion from more than one direction. More data is valuable only when each source has a reason to be there; adding social posts, old decks and AI-generated summaries to a folder is not the same as understanding them.

Imagine a retailer sees that consideration has fallen after a packaging change. A survey shows where the shift is strongest. Short, AI-moderated follow-ups or researcher-led interviews explore why. Social conversation suggests language people use unprompted, but cannot stand in for a representative sample. An earlier study reveals whether the confusion predates the redesign. An agency then checks whether these sources agree, whether the timing fits and which alternative explanation would overturn its recommendation. This is an illustrative study design, not a reported client result.

AI-moderated interviews, social conversations, open-web material, AI-assistant descriptions of a brand and older studies are possibilities, not a compulsory source list. A concept test might need depth interviews and a robust survey; an AI-perception audit might be relevant to a brand visibility question and irrelevant to packaging. Spend the saved time on the missing evidence, not on making every project bigger.

Why is human judgement still part of the product?

Because an AI system can process material without knowing whether the right people were asked or whether the resulting pattern matters to this client now. Researchers have to challenge the brief, choose a defensible method, recognise a consequential exception and take responsibility for the recommendation. “Human in the loop” is not a signature at the bottom of an unchecked AI report.

There are technical reasons to resist equating speed with understanding. In a 2023 preprint studying 2,407 open-ended interviews, 789 expert-annotated, Ashwin, Chhabra and Rao found that errors from the language models they tested were not random with respect to interviewee characteristics. That is one setting with older models, not a verdict on every current tool. It shows why a fluent aggregate can distort who is being heard. Long-input systems can also miss relevant evidence even when it fits in their context window; our context-rot explainer explains that separate coverage risk.

An experienced researcher might notice that three unusual complaints matter more than 30 routine compliments because the client is about to launch a product into the very market those complaints describe. Or that a rise in a tracker is due to a changed sample, not a campaign. Neither call comes from counting quotations alone. The agency earns its fee by showing the decision, the evidence and the uncertainty that travels with both.

This does not mean simply declaring every researcher irreplaceable. The MRS Delphi Group report also raises a harder skills question: if junior researchers never work close to transcripts, tables and fieldwork, how will they learn to challenge automated analysis? Agencies need deliberate apprenticeship—reviewing source-level evidence, diagnosing mistakes and explaining why a plausible conclusion fails—not just a promise that humans will remain involved.

How should an agency change what it sells?

Sell a defined decision and the standard of evidence needed to make it, rather than hiding all expertise inside hours of production. That does not mean charging the same price for less work by default. It means agreeing what has become cheaper, what new questions are worth answering and what level of researcher accountability the fee includes.

Consider three different offers:

  1. A directional check: a bounded question, fast turnaround, explicit limits and no claim that it substitutes for a full study. The lower effort should be reflected in scope and price.
  2. A decision-grade study: suitable recruitment, a stated method, mixed evidence where useful, a review of disagreement and a researcher-owned recommendation. Automation makes parts of the work faster; the client pays for the quality of the decision support.
  3. An ongoing learning programme: new waves tested against prior work, a shared evidence trail and regular revision of what the team believes. The value is continuity and challenge, not an endlessly growing dashboard.

On a time-and-materials contract, do not bill hours that were not worked. On a fixed-fee engagement, discuss deliverables, review and outcome before assuming that a faster process means the client wants a discount or will accept a thinner method. Specialist knowledge—of an audience, category or market—becomes easier to value when the agency can show where it changed the question or the recommendation. The MRS report treats this as a commercial proposition from contributors, not a guaranteed winning formula for every firm.

Which agencies have an opening, and which are at risk?

Agencies that can only sell the production process face a harder comparison with self-service tools. Agencies that combine faster operations with distinctive expertise, evidence discipline and client trust have a stronger reason to be commissioned. Neither position guarantees survival: clients differ, budgets move and some work genuinely will be brought in-house.

The risk is to automate the visible work while leaving the offer unchanged. If the proposal still promises “20 interviews, thematic coding and a 60-page report”, a buyer may ask why the price and schedule look like the old process. A stronger proposal might say: “We will identify which barrier is holding back first-time buyers, test the competing explanations with survey and interview evidence, and recommend a change with the cases that could prove us wrong.” The method remains concrete; the value is now legible.

For an agency, “more research from more directions” should mean a better route to a decision, not a larger pile of output. It can also mean refusing a source that adds noise, saying the sample cannot support a claim or telling the client the evidence is not yet sufficient.

How do you test whether the new model works?

Run one finished, permission-cleared project through the old and AI-assisted routes at the same review standard. Compare the approved result, then ask a client or an uninvolved director which version better supports the decision—not which draft arrived first.

Record four things: total researcher hours including correction; material errors in bases, quotations, codes and coverage; whether the additional sources changed or merely decorated the conclusion; and whether the recommendation names its limits and next action. Cost should include tools, fieldwork and reviewer time. Set the pass criteria before looking at the AI output.

If the new route saves time but weakens the evidence, fix the workflow before selling it. If it saves time at the same standard, decide whether to improve margin, lower the price, add a useful line of enquiry or deliver sooner. Different clients will value those gains differently. Our six questions beyond AI seats cover the underlying team, data and evidence controls.

Where does Meaningful fit?

Meaningful is built to help agencies collect and connect surveys, AI-moderated responses, uploaded interviews, prior documents and other supported sources in one project. Its current product page describes batched reading of included responses and transcripts, source-linked findings and draft client outputs. Those are workflow features to test, not proof that every insight is correct or every source is representative.

The agency still chooses the question, decides whether a source is appropriate, challenges a contradiction and signs off on the advice. On a live project, check permissions and processing terms before uploading client or participant material. Then ask to see which records were processed, which quotations support a claim and what happens when a researcher rejects the first interpretation.

What do agencies ask about AI replacement?

The practical question is not whether a tool can produce familiar-looking deliverables. It is whether the agency has a role that a client can see and trust after those deliverables get cheaper.

Will clients bring all market research in-house with AI?

Some routine work may move in-house, but “all” is too broad. Teams still need recruitment, methodological choices, validation, context and independent challenge for consequential decisions. The boundary depends on the client's capability and the risk of the question, not the presence of an AI tool.

Can AI replace a qualitative researcher?

AI can speed transcription, retrieval, coding suggestions and first drafts. It cannot be held accountable for whether the sample or interpretation answers a client's real question. Researchers must check quotations, minority cases and bias, then own the final meaning. Roles may change even when that responsibility remains.

Should an agency charge less if AI saves time?

Price should match the agreed scope, contract and value delivered. Do not charge unworked hours on an hourly agreement. For fixed-fee work, show what the client receives: a quicker directional answer, more evidence at the same fee or a decision-grade study with explicit review. Do not disguise a thinner method as unchanged research.

Does more data automatically mean better insight?

No. More sources can reveal disagreement, but they can also bring different populations, periods and biases into one confident-looking answer. State what each source adds, what it cannot represent and whether it changes the recommendation. A useful agency sometimes removes a source or asks for fresh fieldwork instead.

What should you do next?

Take one study your team already knows well and compare the old route with an AI-assisted one. If the goal is to connect more evidence without handing the judgement to a tool, try Meaningful's current three-day trial on a permission-cleared project and test the source trail as closely as the speed. The question is not whether all agencies survive unchanged. It is what yours will be hired to understand.