Research Agent
Does the background reading and shows you where every claim came from.
What is research agent?
A research agent gathers published information on a company, market or supplier, writes a short summary of what it found, links every claim to the source it came from, and states explicitly what it could not establish.
Research is slow mostly because the majority of what you read turns out to be irrelevant. The useful part is small and its location is unpredictable.
An agent reads the volume and writes the summary. That much is unremarkable, and plenty of tools do it badly by producing confident prose with no way to check any of it.
The design decision that matters is sourcing. Every point links back to where it came from, and anything the agent could not establish is stated as unknown rather than smoothed over. A research summary you cannot verify is worse than no summary, because it gets quoted.
Why the agent says when it does not know
Duara builds research agents from Nairobi for businesses that check out suppliers, partners and markets regularly enough that the reading has become a real cost.
We build them to declare ignorance loudly. An agent that produces a complete looking summary while quietly omitting what it could not find is the most dangerous kind, because the gaps are invisible to whoever reads it.
Who this is for
- You check out suppliers or partners regularly
- You need market background before decisions
- Staff spend hours reading to find little
- Reports get quoted without anybody checking them
Industries we serve
- Investment and private equity
- Procurement functions
- Trade and export
- Professional services
- Banking credit teams
- NGO programme design
Problems this solves
These are the conditions where research volume has outgrown the time available to do it properly.
- Most reading is wasted. The useful material is a small fraction of what must be read to find it, and nobody can tell in advance which fraction.
- Summaries nobody can check. A confident paragraph gets quoted in a decision paper with no way to trace whether it was ever true.
- Gaps that are invisible. What the research failed to establish is omitted rather than stated, so readers assume the picture is complete.
- Research skipped under time pressure. The check gets abbreviated when a decision is urgent, which is exactly when it mattered most.
What we deliver
Structured gathering across published sources, a short written summary, a link behind every claim, and an explicit statement of what could not be established.
Written summaries on companies, markets or suppliers
A source link behind every individual claim
An explicit list of what could not be established
Recurring research on a schedule where useful
Consistent structure so summaries are comparable
Raw sources retained alongside the summary
Features
- Every claim sourced. Each point links to where it came from, so a decision paper quoting it can be checked rather than trusted.
- Unknowns stated explicitly. What could not be found is listed. An agent that omits its gaps produces summaries whose weaknesses are invisible.
- Consistent structure. The same headings every time, so two supplier summaries can be compared without rereading both.
- Scheduled monitoring. Recurring research on a company or market, so changes surface rather than requiring somebody to check.
- Sources retained. The underlying material is kept alongside the summary, because a link that dies later leaves the claim unverifiable.
Benefits
- Decisions get made on evidence. Research that would otherwise be skipped under time pressure actually happens.
- Claims can be checked. A sourced summary survives challenge in a way a confident unsourced one does not.
- Gaps are visible. Knowing what you could not establish changes how much weight the summary should carry.
- Comparison becomes possible. Consistent structure means three supplier summaries can be read against each other quickly.
Technologies we use
- Structured source gathering. Published material collected systematically rather than through a single search.
- Claim to source linking. Each statement in the summary tied to the specific passage supporting it.
- Source archiving. Underlying material retained, because links break and the claim then becomes unverifiable.
- Gap detection. Identifying what the research set out to establish and failed to, so it can be reported.
- Scheduled re-running. Recurring research on named companies or markets, with changes highlighted against the previous run.
Source retention matters more than search sophistication. Links rot, and a claim whose evidence has disappeared is indistinguishable from an invented one.
How the process works, step by step
A research agent is usually producing usable summaries within three to four weeks, once the standard structure is agreed.
-
Step 1
Free scope call
We establish what you research, how often, and what a decision actually needs to know.
-
Step 2
Structure agreed
The summary headings are fixed so outputs are comparable rather than shaped by whatever was found.
-
Step 3
Source selection
We agree which sources are acceptable, including any subscriptions you already hold.
-
Step 4
Calibration
Early summaries are checked against research your team has already done, which tests both coverage and sourcing.
-
Step 5
Live and scheduled
On demand research plus recurring monitoring where a market or supplier justifies it.
Quality assurance and how we work
The commitments attached to research that will inform decisions.
- Every claim carries a source. Nothing appears in a summary without a link to what supports it, because unsourced claims get quoted and cannot be checked.
- Gaps are reported prominently. What could not be established is stated rather than omitted. Invisible gaps are the most dangerous failure in research.
- Sources are archived. The underlying material is retained, since a dead link leaves a claim indistinguishable from an invented one.
- We do not infer beyond the source. The summary reports what the material says rather than what it suggests, and speculation is labelled if included at all.
- Only agreed sources are used. Which sources are acceptable is your decision, and paid databases are used only where you hold the subscription.
Frequently asked questions about research agent
Can we trust what it finds?
Check the sources, which is why every claim links to one. The agent reports what was published, and published information can itself be wrong.
What if it cannot find something?
It says so explicitly. An agent that omits its gaps produces summaries whose weaknesses are invisible to whoever reads them.
Why archive the sources?
Because links break. A claim whose evidence has disappeared is indistinguishable from an invented one, so we retain the underlying material.
Does it search paid databases?
Only ones you already subscribe to and connect. Otherwise it uses publicly available sources, and the summary says which.
Will it speculate or infer?
No. It reports what the source says rather than what it implies, and any speculation is labelled clearly if included at all.
How long does a summary take?
Minutes once configured, which is the point. The setup takes three to four weeks to agree structure and sources.
Why fix the structure?
So summaries are comparable. Three supplier reports with the same headings can be read against each other; three shaped by what was found cannot.
Can it monitor a company over time?
Yes, on a schedule, with changes highlighted against the previous run so you see what moved rather than rereading everything.
Is this suitable for credit or investment decisions?
It supports them. It gathers and sources published material, and the judgement about what that means stays with your team.
Does it work on Kenyan companies?
Yes, though public information here is thinner than in some markets, and the gap reporting makes that visible rather than hiding it.
Can it research individuals?
We are cautious about that and scope it carefully, since it raises data protection questions that company research does not.
What sources does it use?
Whichever you approve. Source selection is a deliberate step because it determines what the summary is worth.
How is this different from asking a chatbot?
A chatbot answers from training data with no sources. This gathers current material and links every claim to it.
Can it summarise documents we provide?
Yes, and combining your own documents with published sources is a common arrangement.
What if two sources disagree?
Both are reported with their sources rather than one being silently chosen, because the disagreement is usually the interesting part.
Does it produce a recommendation?
No. It gathers and summarises. Recommending stays with the people who understand your commercial context.
How do we know coverage is adequate?
Calibration against research your team has already done tests exactly that before you rely on it.
Can several people request research?
Yes, and requests can be routed by team with different structures for different purposes.
Is our research history confidential?
Yes. What you research stays yours and is never used to train public AI models.
What does each research summary cost?
Request volume and source access determine it, quoted against your expected usage before anything is agreed.
Glossary of terms
Research terms as used here.
- Source
- The specific published material supporting a claim, retained so the claim remains checkable.
- Gap
- Something the research set out to establish and could not, reported rather than omitted.
- Link rot
- The tendency for web sources to disappear over time, leaving claims unverifiable.
- Desk research
- Investigation using published material rather than direct enquiry or interviews.
- Monitoring
- Recurring research on the same subject so changes are noticed without somebody checking manually.
- Provenance
- The chain showing where a piece of information came from and how it reached the summary.
Related services in AI Agents
Talk to us about research agent
Give us a supplier you have already researched properly. Comparing our summary against what you found is the fastest test of whether this is useful.
Start your project