The research production layer, compressed

Four production tasks

These are the tasks that sit between a completed field and a delivered finding. Each one is real work today, and each one is where an agent earns its place.

Open-end coding

Thousands of verbatims are themed and counted overnight. The human role inverts: instead of building the codeframe from a blank page, the coder reviews a proposed codeframe against the data and corrects it. Quality assurance comes from disagreement sampling, where a human codes a sample blind and the gap between human and agent is measured, not assumed.

Transcript synthesis

Eight groups in, one findings document out. The discipline that makes this trustworthy is traceability: every claim in the synthesis links back to the transcript line it came from, so a moderator can audit any sentence in seconds rather than re-reading the night's tapes.

Questionnaire QA

Routing logic, translations and quota maths are checked without fatigue. An agent does not tire deep into a long script, and it applies the same scrutiny to the last translation as the first. Humans still own the questionnaire; the agent owns the tedium of verifying it.

First-draft reporting

The agent drafts the data story overnight: the structure, the charts described, the movements flagged. Seniors then spend their hours where they are irreplaceable, on the "so what" that turns a data story into a recommendation.

The honest limits

Research houses trade on trust, so the limits matter as much as the capability. Four of them are non-negotiable.

• Hallucination is real: agents invent when unchecked, so traceability is not a nice-to-have, it is the design requirement. If a claim cannot point to its source line, it does not ship.
• Codeframes need human judgment: nuance, sarcasm and cultural register still defeat automated coding. A Malaysian respondent's "okay lah" is not a neutral code, and a human has to say so.
• Confidentiality comes first: nothing client-confidential goes to a model without a data agreement in place. The procurement conversation precedes the pilot, not the other way around.
• A human signs every deliverable: accountability does not delegate. The name on the report is a researcher's, and that researcher has reviewed what they are signing.

Where the value moves

When production compresses, it does not disappear; its value migrates. Judgment and design appreciate. The questionnaire architect, the moderator who hears what was not said, the analyst who knows which movement is noise: these roles become more valuable, not less, because their hours are no longer spent on mechanical work. The teams that win will be the ones whose analysts direct fleets of agents rather than compete with them.

The Malaysian context sharpens the timing. AI adoption among Malaysian businesses reached 27 percent in 2025, according to a vendor survey by AWS and Strand Partners of 1,000 business leaders. But depth is shallow: 73 percent of adopters remain at basic use, and only 10 percent have reached advanced integration. That gap between broad adoption and shallow depth is the opportunity. Research teams that build real agentic depth early will be operating in a market where most of their competitors, and most of their clients, have not.

Sources