Data Analysis Agent
Ask a question in plain words and see exactly how the answer was worked out.
What is data analysis agent?
A data analysis agent turns a question asked in ordinary language into a query against your own data, returns the answer, and shows the query and figures it used so anybody can check how the number was reached.
Most businesses have the data and cannot reach it. The question is simple, the answer sits in a database, and getting it requires somebody who writes SQL and has three other things to do first.
So the question goes unasked. Decisions get made on the monthly report, which answers last month questions rather than this morning ones.
An agent lets somebody ask directly. The important part is that it shows its working: the query it ran, the rows it counted, the filters it applied. A number nobody can check is not analysis, it is a rumour with a decimal point.
Why the query stays visible
Duara builds data analysis agents from Nairobi for businesses whose questions currently queue behind whoever knows the database.
We refuse to hide the query. Every tool in this category can produce a confident number, and the ones that do not show their working are the dangerous ones, because a wrong answer presented cleanly is indistinguishable from a right one until somebody acts on it.
Who this is for
- Questions queue behind one technical person
- Decisions wait for the monthly report
- Different reports give different numbers
- Nobody can check where a figure came from
Industries we serve
- Retail and distribution
- Financial services
- Logistics
- Manufacturing
- Healthcare groups
- Telecommunications
Problems this solves
These are access problems rather than data problems. The information usually exists already.
- Questions queue behind one person. Anything requiring SQL waits for the person who writes it, and most questions are simply never asked.
- Decisions made on stale reporting. The monthly report answers last month questions, which is not when the decision is being made.
- Numbers that disagree. Two reports give different figures for the same thing and nobody can establish which is right.
- Figures nobody can trace. A number appears in a board pack and no one present can say how it was calculated.
What we deliver
Plain language questions answered directly against your data, with the query and underlying figures shown, and an explicit refusal to answer where the data cannot support it.
Plain language questions answered from your own data
The query and figures shown behind every answer
Charts where a chart communicates better than a number
Scheduled answers to questions you ask repeatedly
Agreed definitions for your common measures
An explicit refusal where the data cannot answer
Features
- The working is always shown. Query, filters and row counts are visible, because a number nobody can check is a rumour with a decimal point.
- It declines rather than approximates. Where your data cannot answer the question asked, it says so instead of producing a plausible figure.
- Definitions agreed once. What counts as an active customer is settled with your team, which is usually why two reports disagree.
- Reads only. The agent queries. It does not write to your systems, so a bad question cannot damage anything.
- Repeat questions scheduled. Anything you ask weekly can arrive automatically instead of being asked.
Benefits
- Questions get asked. The ones nobody bothered to raise because of the queue are usually where the useful findings are.
- Answers arrive in time to matter. A question answered this morning affects today decision. The same answer next month does not.
- Reports stop disagreeing. Agreeing definitions once resolves most contradictions between existing reports.
- Numbers become checkable. A figure with its query attached survives challenge in a board meeting.
Technologies we use
- Natural language to query. Turning an ordinary question into a query against your actual schema.
- Read only access. Query permissions without write access, so nothing can be altered by a question.
- Working disclosure. Returning the query, filters and row counts alongside every answer.
- Measure definitions. Encoding agreed definitions so the same term means the same thing in every answer.
- Scheduled queries. Running recurring questions automatically and delivering the result.
Agreeing what your measures mean is the unglamorous prerequisite. Most disagreement between reports is definitional rather than technical.
How the process works, step by step
A data analysis agent usually takes four to six weeks, and defining your measures normally takes longer than connecting the data.
-
Step 1
Free data review
We look at what data you hold, its condition, and which questions currently go unanswered.
-
Step 2
Definitions agreed
Your common measures are defined with the team, which is where most report disagreements are actually resolved.
-
Step 3
Read only connection
Access is configured without write permissions, so no question can alter anything.
-
Step 4
Checked against known answers
It answers questions your team already knows the answer to, which tests it honestly.
-
Step 5
Opened to the team
Access widens once accuracy is proven, usually starting with one department.
Quality assurance and how we work
The commitments attached to a system producing numbers people will act on.
- Every answer shows its working. The query, filters and row counts are returned with the number, because an answer nobody can check is a rumour with a decimal point.
- It declines rather than approximating. Where your data cannot answer the question, it says so rather than producing a plausible figure.
- Access is read only. The agent queries and never writes, so a badly phrased question cannot alter or delete anything.
- We test against answers you know. Validation uses questions your team can already answer, which is the only honest way to check it.
- We report data quality problems. Where your data is too inconsistent to answer reliably, we say so rather than automating on top of it.
Frequently asked questions about data analysis agent
Can anyone use it or only analysts?
Anyone. Questions are asked in ordinary language, which is the point, though the working shown alongside is there for whoever wants to check it.
How do we know the answer is right?
Check the working. Every answer returns the query, filters and row counts, so the calculation is verifiable rather than trusted.
What if it gets something wrong?
The visible query is how you find out. This is exactly why we refuse to hide it, since a wrong answer presented cleanly looks identical to a right one.
Will it make up numbers?
It declines where your data cannot answer the question rather than producing a plausible figure. That refusal is a deliberate design choice.
Can it change our data?
No. Access is read only, so a badly phrased question cannot alter or delete anything.
Why do our reports disagree today?
Usually definitions rather than technology. Two reports count active customers differently, which is why defining measures comes before connecting data.
What data can it use?
Databases, spreadsheets and business systems you connect. It answers from your data, not from general knowledge.
Does it work if our data is messy?
Up to a point, and we report where it is too inconsistent to answer reliably rather than automating on top of it.
How soon can we start asking questions?
Four to six weeks, with measure definitions normally taking longer than the data connection.
Can it produce charts?
Yes, where a chart communicates better than a number, though the underlying figures stay available.
Can we schedule regular questions?
Yes. Anything you ask weekly can arrive automatically instead of being asked each time.
Does it replace our BI tool?
No. It answers ad hoc questions that would otherwise queue. Dashboards still do the job of showing the same measures repeatedly.
Who can see which data?
Permissions follow your existing access rules, so it cannot show somebody data they could not already reach.
Can it explain why a number moved?
It can show what changed in the underlying data. Why it changed commercially is a judgement it does not attempt.
Does it work across several systems?
Yes, though joining systems reliably depends on shared identifiers, and we test that before promising it.
Is our business data confidential?
Yes. It stays in your systems, is queried in place, and no public AI model is trained on it.
Can it forecast?
Basic trend projection, yes, clearly labelled as projection. Anything more serious is a separate piece of work.
What if two people ask the same question differently?
Agreed measure definitions are what make the answers match, which is the main reason that step exists.
Do we still need an analyst?
For anything requiring judgement, yes. This removes the queue for routine questions, which frees analysts for the harder work.
What is the ongoing cost of query access?
Query volume and data size determine it, quoted against your own usage before anything is agreed.
Glossary of terms
Data terms in plain language.
- Query
- The instruction sent to a database to retrieve specific data.
- Measure definition
- The agreed meaning of a business term, such as what makes a customer active.
- Read only access
- Permission to look at data without the ability to change it.
- Data quality
- How complete, consistent and accurate your records are, which sets the ceiling on any analysis.
- Schema
- The structure of your database, including what tables exist and how they relate.
- Working
- The query and figures behind an answer, shown so the result can be verified.
Related services in AI Agents
Talk to us about data analysis agent
Think of a question about your business you have wanted answered for months but never asked because of the queue. That is the test.
Start your project