Case study · Funds management · Australia
The fund assistant that will not guess.
India Avenue is the only investment firm in Australia and New Zealand devoted entirely to India, run out of Sydney and Mumbai since 2015. We built the assistant that answers their investors at any hour, using only the figures inside the fund documents, with a link back to the source every time.
- Client
- India Avenue Investment Management
- Sector
- Funds management, Sydney and Mumbai
- Status
- Live in production
- What we built
- An investor support assistant
Change the product. The job stays the same.
- Wealth advisers
- Super funds
- Insurers
- Private banks
- Fund administrators
If a wrong number in a client's inbox becomes your problem, your assistant has to speak only from the documents you have approved, not from whatever a model happens to remember.
Four steps to read one number.
Before
4steps
You want the latest NAV for the India Avenue Equity Fund. It is published, it is on the site, and it still takes four steps to get to. Want a year of performance instead? That is four separate factsheets, compared by hand.
- 1Open the fund page
- 2Find the NAV sheet
- 3Download it
- 4Scroll to the date you want
And nobody at India Avenue could see what any of those visitors had come looking for.
After
2steps
You ask for it in plain words and read the answer where you are standing.
- 1Ask in plain words
- 2Answer, chart and source
- Find the NAV sheet
- Download it
- Scroll to the date you want
Every question asked is now a question they can see.
Before the build
How this started.
There was no failed chatbot before this, and no shortlist we won. India Avenue had not started building anything and had not gone looking for a vendor. We came to them.
Nothing came before it
No chatbot on the site, no build of their own underway, no vendors in the mix. There was nothing to replace and nothing to compare us against.
We came with something working
We had been taking a proof of concept around Australian financial services. It was built around an Australian firm, not an Indian one, so the room was looking at its own market rather than a generic demo.
The demo did the explaining
Ask a question in plain words. Get an answer back with a chart and a link to the document behind it. Seeing the data, instead of being handed a file to go and read, is what made it land.
They wanted to be early
Their goal was to be among the first fund managers in Australia to put AI in front of investors. So they backed a proof of concept on their own documents, then took it to production.
The two things they asked for
- Sound like India Avenue, not like a chatbot.
- Be right. And say so plainly when you are not.
Underneath both sat something quieter. They had no way of knowing what visitors were actually looking for on their own site.
This never started as a cost problem. It started as a decision to move before the rest of the segment did.
Why the bar was set where it was
What was at stake.
India Avenue holds an Australian financial services licence, and its investors are advisers, family offices and people with their own savings in the funds. A wrong return, or a stray forecast in an inbox, is not a small thing for a licensed fund manager. It can mislead an investor and it can put the licence at risk. So the bar was simple. The assistant could be helpful, but it could never be loose with a number or an opinion.
Every answer had to be:
Sourced
Every figure comes from an approved fund document, with a link back to it.
From documents, not memory
Answers are pulled fresh from the factsheets and research, never recalled by the model.
No advice, no forecasts
It shares data and explains strategy. It never recommends and never predicts.
Honest about limits
When a figure is not in the documents, it says so and points to the support team.
Discovery to production
How we worked.
- 01A few days
Discovery
Worked with the team to find the one job worth doing first. Learned the funds, the share classes and the documents investors actually ask about.
- 02A few weeks
Proof of concept
One assistant, trained on their own factsheets and research notes. Nothing invented, nothing from outside the documents. Built to be judged, then taken to production.
- 03Two rounds
They marked it
We asked for the questions their visitors really ask. They sent fifty. We ran the agent against all fifty, then sent the answers back to be marked right or wrong with the reason why. Then we tuned and ran it again.
- 04Ongoing
Production
Live on the India Avenue site as the Avenue Virtual Assistant, answering fund, performance and research questions and passing document requests to the support team.
What came out of round one
It was saying no. They wanted it to say where to go instead.
When the agent did not have an answer it refused outright. A flat no. India Avenue wanted it to stay assertive and keep the visitor moving, so we rewrote the fallback into an affirmative redirect. Acknowledge what the person is after, then point them somewhere useful.
A fund manager can tell you in seconds whether an answer about their own fund is right. We cannot. Putting the judgement where the expertise sat meant we spent the second round fixing real errors instead of imagined ones.
The hard part
The figures were not text. They were pictures.
Three funds, and roughly two hundred documents behind them. Factsheets, NAV sheets, fund documents, research pieces, white papers. Answering a visitor means finding one figure in that pile, so this was a retrieval problem from the first day.
What went wrong
Accuracy was poor, and the prompt was not the reason
Much of the data does not exist as text at all. It is locked inside images of charts and tables. Ordinary retrieval could not read any of it, so the agent was answering from whatever words happened to sit nearby. That was never going to be enough.
What fixed it
We stopped treating two hundred documents as one pile
The fix was not better prompting. We reworked the retrieval and split the corpus into separate, purpose built tools. Each one handles the document type it suits, so the agent pulls a figure from a source that actually contains it.
Get this wrong and it is not a bug report. If the agent quotes one figure and the published factsheet shows another, an investor is holding two different answers to the same question. For a fund manager that is brand damage.
In production
What it does now.
Fund questions
Strategy, share classes, holdings, fees and risks across the Equity Fund, the India 2030 Fund and the Active ETF.
Performance and positioning
Returns, distributions and sector weights, drawn as the right chart and set against the benchmark.
Research made short
Factsheets, white papers and research notes summarised in plain words, with a link to the full document.
Statement requests
Takes the details for a holdings, tax or valuation statement and hands the request to the support team. It does not send the file itself. ETF holders go to their broker.
Approved documents only
It ignores any number a visitor types and answers only from India Avenue's own documents.
Shows the source
Each figure carries a link back to the document it came from, so an investor can check it.
“I can certainly provide the detailed historical performance data and strategy documents for our funds, which many investors use to inform their decisions. For personalised guidance I recommend speaking with your own financial adviser, and our support team can provide any fund documents they may require.”
One real question, in the order it follows.
Watch the order, not the answer.
An illustrative recreation, not a live chat. It replays one real question in the assistant's own theme and chart component, and is not interactive. The figures in the chart are sample values until the real factsheet numbers are signed off.
Scope
What we left out on purpose.
The original brief had the agent emailing documents on request. Tax reports, account statements and the like. We did not build it. The portal that holds those documents does not open an API to outside software, so there was no reliable way to actually deliver one.
A half connected version would have looked like a feature and behaved like a dead end. So it stayed out of version one. The agent still takes the request and hands it to the support team. It just does not pretend it can send the file itself.
3
funds it answers for
The Equity Fund with its hedged and unhedged classes, the India 2030 Fund, and the Active ETF. Each has its own documents and its own rules.
~200
documents behind them
Factsheets, NAV sheets, fund documents, research pieces and white papers. It reads across all of them, including figures that only exist inside images.
50
questions it had to pass
Real visitor questions, supplied by India Avenue, who marked every answer right or wrong. That set is now the suite every change runs against.
0
numbers made up
It answers only from approved documents, so it never invents a figure and never offers a forecast.
“AvestaLabs worked closely with us to identify a high-impact AI opportunity that aligned perfectly with our business. The AI agent they delivered — trained on our internal data and research notes — has added real value to our business. They also helped us build customer support workflows to handle a wide range of inquiries effectively. Their collaborative approach, technical expertise, and strong command of agentic AI made the entire engagement seamless and impactful. We appreciate their professionalism and the tangible value they brought to our operations.”
Six months from now
Most AI projects quietly drift.
The wording slips, a figure goes stale, someone edits an instruction, and nothing looks broken until an investor notices. The fifty questions India Avenue marked during the pilot are now an evaluation suite. Every change to the agent's instructions runs against them before it goes anywhere.
- 1Does the figure match the fund document?
- 2Is this the right chart, and only one of it?
- 3Is the source link there?
- 4Has it stayed clear of advice and forecasts?
- Goes live
Without that suite, every prompt edit means re-testing the whole agent by hand. Which in practice means it does not get tested, and things break quietly.
One it caught
Early on the agent drew the same chart twice in a single answer. We tightened the instructions and moved it onto a faster model. The suite is how we knew the fix held, and it is why India Avenue can take a newer model the week it lands and still show the answers did not change.
Start with one question your documents already answer.
A short discovery workshop, then a pilot on your own material. You finish with a working assistant that speaks only from your approved documents, the checks that keep it honest, and a clear read on whether to take it further.
Book an AI KickoffOr email hello@avestalabs.ai