Case study · Climate data · Australia
The climate agent that will not answer from memory.
OnlyFacts publishes the Australian climate data that newsrooms quote by name, to a free public site and to a paid product used by media houses, research organisations and analysts. We built the two agents that answer both in plain English, grounded in a live query every time, and cited like everything else on the site.
- Client
- OnlyFacts.io, Australia
- Timeline
- 2025 → ongoing
- Status
- Live in production
- What we built
- Two agents on a dedicated data layer
Change the subject matter. The job stays the same.
- Research desks
- Regulators
- Investment committees
- Compliance teams
- Clinical registries
If your organisation can be held to its numbers, your agent has to be traceable to a source, not merely fluent.
One report. Two days, or a few minutes.
Before
6steps
Every figure OnlyFacts holds was reachable only through a page somebody decided on in advance. So a subscriber report on a single topic started from scratch, with an analyst.
- 1Work out which pages hold the answer
- 2Pull the figures out of each dataset
- 3Combine them by hand
- 4Cross-check every number
- 5Chart it and lay it out to brand
- 6Write it up in the OnlyFacts voice
One to two days of an analyst's time, or more, for one report on one topic.
After
2steps
The question goes in the way an analyst would say it out loud.
- 1Ask in plain English
- 2Report, charts and citations
- Pull the figures out of each dataset
- Combine them by hand
- Cross-check every number
- Chart it and lay it out to brand
The report comes back in minutes, in their voice, with every figure cited.
The same change landed on the free side of the site. Someone who wants to know how their state compares to the national trend no longer has to work out which pages hold the answer, or read the charts themselves.
Why the bar was set where it was
What was at stake.
OnlyFacts publishes under CC BY 4.0. Anyone may republish its figures, with attribution, without asking first. A wrong number does not sit on the site waiting to be corrected. It travels. And for a nonpartisan fact tank, neutrality is the entire product.
Every output had to be:
Traceable
Named source, down to the slice of the dataset the figure came from.
Queried, not recalled
Produced by a database query, never from the model's own memory.
In house voice
Australian English, no judgement, no filler, and no word that turns a data point into a story.
Honest about limits
Able to say “we don’t hold that” rather than approximate.
Discovery to production
How we worked.
- 013 days
Discovery
Learned the domain and the climate areas it covers, how readers already use the charts, tables and articles, and how the data is stored and served today.
- 02Under 3 weeks
A narrow pilot
One agent. Two areas: emissions and transport. Their real database, the site's own visualisations, tight iteration loops. Nothing else in scope.
- 03Their call
They decided it was worth it
The pilot was built to be assessed, not admired. It answered real questions on their own data, which was enough for OnlyFacts to judge the direction and commit to building it out.
- 04Ongoing
Production
Two agents live on a dedicated data layer: the everyday one on onlyfacts.io, and Deep Research behind the paid product.
In production
What it does now.
Everyday agent
Live on onlyfacts.io. Single-area questions, national down to postcode.
Deep Research agent
Multi-step questions over sizable data, returned as one report.
190 data tools
One per published view. It cannot ask what the site cannot answer.
A data layer built to be read
Our own server sits in front of the database and hands back only what the question needs.
No mental arithmetic
Every sum, even a monthly total, goes through a calculator call.
Cited to the slice
Not just the source name. The dataset and the filter behind the figure.
No open web
Search, uploads and code execution are all switched off.
Shows its working
Readers see the query that produced the number, by default.
“I'm sorry — this agent is designed to provide only factual, data-backed information but I cannot generate unsupported projections.”
The models underneath
Both agents run on Claude.
The reasoning behind every answer comes from Anthropic's Claude family, called through the Claude API. Which Claude model sits under which agent is settled by the job rather than by preference: the cheapest one that clears the evals wins, and the two agents have no reason to run the same one.
Both cleared our internal evaluation set on OnlyFacts' own data, and did it at the best intelligence per dollar of the options we tested. That combination is the whole reason the split works: a cheaper model is only a saving if the answers still pass, and the checks further down this page are how we prove they did.
The chart is chosen by the shape of the data.
Left to itself, a model picks the chart it feels like, and a reader ends up looking at a stacked area made of things that do not stack. So it does not get to pick. The result decides, every time, by rule.
- ColumnSeparate categories, or time grouped into distinct intervals
- BarLong labels, or more than six categories
- AreaOne category, over time
- Stacked areaSeveral categories over time that add up to a total
- LineSeveral categories over time that do not add up to a total
The other half of the same problem is classification. A plug-in hybrid is not an electric vehicle, and it does not get to wander into that series, because what counts as what is settled in the data layer before anything is drawn.
One real question, start to finish.
Watch the order, not the answer.
An illustrative example, not a live chat. It replays one real conversation, rebuilt in the live widget's own theme, and is not interactive. Every figure was read from onlyfacts.io on 21 August 2026 under the source listed.
The long answer is a document, not a wall of chat.
A deep research question runs as a sequence of steps over sizable data and comes back as one synthesised report, set in the OnlyFacts brand: their fonts, their colours, their logo, laid out the way their published work is laid out. A subscriber can pass it on without reformatting it first.
“Provide a comprehensive analysis of how Australia is progressing toward net zero emissions. Compare national trends with state performance, identify the biggest contributing sectors, and assess whether current trajectories will meet 2030 targets.”
Minutes
for a report that took days
One to two days of analyst time, or more, became a report that arrives while you wait. Same voice, same citation standard.
75%
less data per report
A full deep research report costs cents rather than dollars.
<3
weeks to a working pilot
From first discovery session to something they could open and use.
“AvestaLabs has been an indispensable partner for OnlyFacts. Their development of AI agents on our proprietary datasets combined deep technical expertise with innovative problem-solving and sound strategic advice. They have become a collaborative and trusted extension of our team.”
Juliette O'Brien OAMCEO and Co-Founder, OnlyFacts.ioSix months from now
Most AI projects quietly get worse.
The model updates, the wording drifts, and nobody notices until a reader does. So nothing ships until it clears five checks.
- 1Does the number match the database?
- 2Do the totals, units and rounding agree between the words and the chart?
- 3Is that the right chart for this shape of data?
- 4Is the citation there, and does it point at the right slice?
- 5Has a banned word crept back in?
- Ships
Those same checks are why OnlyFacts can pick up a newer Claude model the week it lands. We can show the answers did not change, and the saving goes to them.
Start with one narrow question your data should already answer.
A three-day discovery workshop, then a four-week pilot. You finish with a working agent on your own data, the evals that keep it honest, and a straight answer on whether it is worth going further.
Book an AI KickoffOr email hello@avestalabs.ai