Getting to 80 percent: how AI changes research work

AI is especially useful at getting research work to the first 80 percent.

That phrase is deliberately rough. It does not mean the system is 80 percent correct in every case. It means the system can often compress the messy first pass: finding options, gathering sources, summarising long material, removing duplicates, comparing specifications, translating across languages, surfacing constraints, and preparing questions for a human reviewer.

The final 20 percent still matters. In many professional settings, it matters most. That is where verification, judgment, risk, context, and responsibility live.

The Research Problem

Research work often begins with a wide search.

A team may need to compare suppliers, materials, policies, precedents, vendors, market examples, regulatory positions, product references, or project benchmarks. The early work is slow because the researcher has to discover the landscape before they can make sense of it.

They search manually. They open tabs. They compare PDFs. They copy specifications into notes. They check whether a source is current. They translate supplier pages or standards. They notice that three listings describe the same product with slightly different names. They try to work out whether guidance from one geography applies to another. They try to remember why one option was rejected. They turn scattered material into a first view.

AI can help with that first view.

What AI Can Prepare

For research-heavy work, an agent can prepare:

• candidate options;
• source links and document references;
• short summaries of long pages, PDFs, catalogues, or notes;
• deduplicated records where the same supplier, product, policy, or precedent appears in several places;
• normalised specifications so similar fields can be compared side by side;
• translations or bilingual summaries where relevant sources are in another language;
• geography and industry notes that show where a source may or may not apply;
• obvious constraints and open questions;
• comparison tables;
• missing information;
• questions for suppliers or experts;
• risks that require verification;
• next-step recommendations for review.

This is useful because it changes where the professional starts. Instead of beginning with an empty page or a broad search, the reviewer begins with a structured map.

The important work is not only retrieval. It is clean-up and sense-making. A useful system should reduce repeated reading, make source differences visible, and help the reviewer see when two sources are saying the same thing, when they are saying different things, and when the difference is caused by region, terminology, industry context, or translation.

What The System Is Really Doing

The visible output may be a table or memo, but the useful system work happens underneath.

That work is time consuming when done manually. It is also easy to underestimate because the final deliverable may look simple: a table, a short memo, a shortlist, or a set of questions. The labour is in getting the research clean enough that a person can use it.

Example: Supplier Or Material Research

Consider a team looking at material options for a physical product, fit-out, installation, or prototype.

The old workflow might involve searching marketplaces, supplier pages, spec sheets, forums, catalogues, and prior notes. The work is not just finding names. The team needs to compare properties, minimum order quantities, certifications, lead times, tolerances, finishes, costs, and open questions.

It also needs quieter research hygiene. A supplier may list the same material under multiple product names. A marketplace page may copy a manufacturer's claim without preserving the original context. A certification may apply in one geography but not another. A Chinese-language supplier page may contain useful technical detail that has to be translated before the team can review it. A product may be relevant in furniture but inappropriate for medical, food-contact, exterior, or fire-rated use.

Those distinctions take time. They are also exactly the kind of preparation that makes research review better.

An AI-assisted research packet could show:

The system helps the team get oriented. It does not decide what to buy, test, specify, or recommend.

Example: Policy Or Precedent Scan

In advisory, legal, compliance, or operations work, the research object may be less physical but just as scattered.

An AI research packet might gather:

• relevant policies;
• prior memos;
• comparable decisions;
• clauses or standards to inspect;
• jurisdiction or geography notes;
• translated summaries where source material is not in the team's working language;
• conflicts between sources;
• areas where the source base is stale;
• questions for the responsible professional.

Again, the point is preparation. The system gets the reviewer to a better starting point. It does not make the accountable recommendation.

Here too, the value is often in the clean-up. A policy scan may involve several versions of the same guidance, commentary that refers to an older rule, examples from another jurisdiction, and internal notes that use a different vocabulary from the formal source. The system can help cluster, translate, date-check, and mark relevance before the professional decides what matters.

Why The Final 20 Percent Still Matters

The final part of research is not clerical.

It includes:

• checking whether sources are authoritative;
• verifying current prices, availability, and terms;
• deciding which tradeoffs matter;
• applying professional standards;
• weighing risk and client context;
• knowing when the evidence is not enough;
• deciding what to test or recommend.

Generative AI can produce confident outputs even when information is incomplete or wrong. NIST's Generative AI Profile discusses this risk in terms of confabulation and unsupported claims. That is why research agents should show sources, uncertainty, and missing information clearly.

The Source Map Is The Product

For research work, the most valuable artifact is often the source map.

A source map tells the reviewer:

• what the system found;
• where each claim came from;
• which records were grouped as duplicates;
• which terms, units, or categories were normalised;
• which sources were translated;
• which geographies, industries, or use cases the source appears to apply to;
• which sources disagree;
• which sources are stale or unclear;
• what the system could not verify;
• what should be checked manually.

This is more useful than a polished answer. A polished answer can hide weak evidence. A source map makes the evidence inspectable.

That is why research workflows fit naturally with source grounding. The team needs to see the basis for the comparison, not only the conclusion.

How To Use AI Research Safely

The practical workflow is:

1. Define the research question and decision boundary. 2. Let the agent gather and organise the first pass. 3. Let the system deduplicate, normalise, translate, and mark likely relevance. 4. Review the source map, not just the summary. 5. Mark unsupported claims, missing evidence, and questionable relevance. 6. Verify high-consequence facts directly. 7. Decide what to test, shortlist, purchase, advise, or reject. 8. Feed corrections back into the next research run.

This keeps the agent useful without turning it into the decision maker.

What This Is Not

Getting to 80 percent is not a promise that AI has done most of the thinking.

It means AI can reduce the cost of getting oriented. It can make the first pass faster, wider, and easier to review. The remaining work is not a formality. It is the professional work.