AI Agents in Finance: From Cited Reports to Finished Deliverables

Most AI research agents are graded on their prose. Analysts are graded on what lands in the model, the memo and the deck.

This post covers a large update driven entirely by watching how people actually used it on our Finance open source app.

Finance is open source.

TL;DR

• 23 curated deep research workflows across investment banking, private equity, hedge funds and GTM — each with a depth chip, a runtime estimate, and deliverable badges stated before you spend a credit.
• The app now reads your freeform query and pre-fills the deliverables picker with the format it implies. Conservative, visible, editable, and it never blocks your run.
• Generated files moved to the top of the report, above the body.
• Every claim resolves to a clickable primary source.
• It's open source on github and live at finance.valyu.ai

The Missing Deliverable

Ask a junior analyst for "a peer comps table for Alphabet vs Microsoft and Meta as an Excel spreadsheet" and you get a spreadsheet.

Ask most AI research tools the same thing and you get a beautifully written essay — sometimes with formatted tables — about peer comparables, followed by silence on where the downloadable file should be.

That silence is the missing mile. Every leaderboard measures retrieval accuracy, citation quality and reasoning depth. Very few measure whether the thing you asked for arrived in a form you could open.

Which workflows ship today

Workflows are organised by vertical, because a hedge fund analyst and a GTM lead are not looking for the same artifact. The catalogue currently holds 23 templates across four domains:

• Investment banking — company profiles, IC memos, comparables
• Private equity — commercial due diligence, market maps
• Hedge funds — thesis work, screening
• GTM — account intelligence briefings

A fifth lens, Popular, sits first and cuts across all four, surfacing the most-used templates.

Each card carries three signals before you commit a credit:

Pick one and you get an auto-generated form. The variables come from the workflow definition itself, so a template that needs a company gets a company field with real ticker examples, and a template that needs a thesis gets a textarea instead of a single-line input.

The agent now reads your query

Alongside the templated workflows there's a freeform research box. Type anything, pick a depth, hit enter, get a cited report.

A deliverables picker sat under that box: pick a format, describe the contents, get a file. It worked perfectly. It also required you to know it was there.

So what happens when someone doesn't know which format is right, or simply forgets to open it? Finance now runs a small model over your settled query and pre-fills the picker with a format and description drawn from the query itself.

Type the peer comps request, pause for a moment, and the panel opens on its own with an Excel row already filled in:

What it infers, and when it stays quiet

The design constraint was that inference must never make things worse than the old manual flow.

• It suggests, it does not decide. Suggestions appear in an open panel, badged, before launch. Every one is editable and deletable. Nothing is attached without being shown to you first.
• It fails open. No API key, a timeout, or a malformed response all resolve to "no suggestions" — your research launches exactly as it did before, with nothing visible going wrong.
• It is schema-bound. The response is strictly validated; a bare-array answer is rejected rather than guessed at.
• It caps out at five deliverables per run across .xlsx, .pptx, .docx, .csv and .pdf.

Deliverables now lead the report

Small in the diff, large in practice.

Generated files used to render last — below the full report body and the chart gallery. On a report running to several thousand words, the spreadsheet you asked for sat several screens down, past every table and footnote.

They now sit directly under the title, above the collapsed activity feed and the body.

The reasoning is straightforward: the file is often why a finished report gets opened at all. Charts stayed where they were, because they illustrate the analysis and read alongside it.

Why every claim carries its source

Templated workflow or freeform query, the output is cited inline. Claims resolve to favicon pills you can click through to sec.gov, finance.yahoo.com, nvidianews.nvidia.com, and whatever else the research actually touched.

The reasoning isn't hidden either. Every run keeps an activity feed — the model's plan, the searches it ran, the sources it found, the code it executed — live while it runs, collapsed by default once it finishes.

For financial research this isn't a nice-to-have. A number you cannot trace is a number you cannot put in a memo.

Three lessons for anyone building finance agents

The deliverables picker was fully functional the entire time. It produced excellent spreadsheets. Its problem was that reaching them required a user to know it existed, open it, choose a format, and write a good description — four steps of friction guarding a capability they had already described, in plain English, one input box above.

1. Grade your agent on the artifact, not the prose. If an analyst has to retype your output into Excel, the run finished at 80%.

2. Inference belongs where users already speak. If people describe a control in natural language before failing to find it, that control is a candidate for inference. Show what you inferred, let them change it, never let it block the run.

3. Provenance is the product. Depth of reasoning is worth nothing if the number cannot be traced to a filing with a date on it.

The query already contained the answer. All that was missing was something small, cheap and conservative to read it.

Try it

Open finance.valyu.ai and run a workflow on a name your team already knows well. Compare it against what you would have produced manually.

Check the citations. Do they resolve to the actual filing, or to an article about it?

Then type a query that names a file, like peer comps for GOOGL, MSFT and META as an Excel spreadsheet, and watch the panel open before you hit enter.

Finance runs on the same Search and DeepResearch APIs your own agents can call. Grab a key at platform.valyu.ai — $10 in free credits, $20 with a work email, no card — and see the full coverage at docs.valyu.ai.

FAQ

What is an AI agent in finance?
One that plans and executes a multi-step research task — deciding which sub-questions to ask, fanning out across filings, market data and news, cross-checking what it finds, and returning a synthesised, cited output. It differs from a single search call, which answers one question against one index, and from a chatbot, which has no persistent task or artifact.

What is a deep research workflow?
A versioned, parameterised research template that runs as an async task and returns a cited report plus downloadable files. The methodology, source strategy and section structure are fixed; the inputs — company, thesis, category — are variables anyone on the team can fill in.

Does the app decide what files to generate on its own?
No, it suggests. Suggestions appear in an open panel, badged, before you launch, and you can edit or delete every one. Nothing is attached without being shown to you first.

What happens if the suggestion model is unavailable?
Nothing visible. The feature fails open — no key, a timeout, or a bad response all resolve to "no suggestions," and your research launches exactly as before.

Which model does the deliverable extraction?
gpt-5.6-luna by default, overridable with the DELIVERABLE_SUGGEST_MODEL environment variable. Whatever you pick must support strict structured outputs, since the response is schema-validated and a bare-array answer is rejected. Without an OPENAI_API_KEY the extractor stays off.

How long does a deep research run take?
Depth is the knob, not speed. Estimates range from 4–8 minutes for an Account Intelligence Briefing to 10–25 for a Market Map, IC Memo or Commercial Due Diligence. Runs continue server-side if you navigate away — pick them up again from Reports.

What file formats can it produce?
Excel (.xlsx), PowerPoint (.pptx), Word (.docx), CSV and PDF, up to five per run. Office formats are produced through code execution, which the app enables automatically when you request one. Every run produces a PDF regardless.

Can I run this without a Valyu account?
Not on finance.valyu.ai — that needs a signup. But it's open source, so you can deploy it yourself by switching the mode to self-hosted in the env variables.

Can I build my own workflow?
You can build workflows on the Valyu platform, and the app links you there. Note that the in-app browser currently requests the curated catalogue only (scope=valyu), so custom templates won't appear in the domain lenses yet.