Anequityanalyst'sweek,runthroughanAIagent

An AI agent can take the first pass at most of an equity analyst's recurring week: the Monday market note, the earnings flash, an initiation draft, thesis work and the Friday monitoring. We arranged seven of those workflows as a working week and priced each one with the reference costs Bigdata.com publishes on its pricing page. The full week of data comes to $8.39. The useful part is not the low price, but seeing what actually drives the cost, what the agent can prepare, and where the analyst still has to make the call.

Key takeaways

• Bigdata.com's reference costs put each workflow between $0.57 and $1.82 per run, and one run of all seven comes to $8.39.

• Cost follows the evidence retrieved, not the length of the note: a 1,560-token short-thesis memo costs more than a 2,600-token initiation report.

• Pricing is per token of content retrieved, so an agent pays only for the passages that answer the question, not for whole documents.

• The agent prepares the first pass; the analyst still decides what is trusted, changed and published.

Monday: what does the weekly market note look like when an agent drafts it?

A good Monday note tells you what changed over the weekend and what could matter this week, not every headline that ran. The questions are simple. What moved across rates, currencies and the coverage list? Which data releases and events are coming? What deserves a closer look before the open?

An agent can do the gathering. It checks approved news, macro and company sources, groups the developments by theme, and drafts a short note with a citation behind every claim. The analyst then reads it and decides which items actually touch the book.

That review matters because importance is not uniform. A move that changes one thesis may be noise for another. Bigdata.com's reference cost for a weekly market commentary, a roughly 1,040-token macro note, is $0.82 per run.

Tuesday: how does an agent handle the post-earnings flash?

Earnings suit an agent because the first questions repeat every quarter and the clock is short. Pull the reported numbers, compare them with the prior period or the estimates you supply, isolate guidance changes, summarise management's explanation and flag anything that does not reconcile.

The agent gets the first pass out fast. The analyst should still open the release behind the key figures and check the transcript once it lands. A post-earnings note is exactly where a wrong period, a wrong unit or a hurried calculation produces a confident-looking error.

It is also the cheapest job of the week. Bigdata.com's reference example, an RH post-earnings flash of about 520 tokens, costs $0.57 per run.

Wednesday: can an agent draft an initiation report?

It can draft one, and the important word is draft. An initiation pulls together work that suits an agent well: company description, segment history, management commentary, key risks, peer evidence and the chronology behind the thesis.

The analyst still sets the framework. Which peers are truly comparable? Which KPI deserves the most weight? Which parts of management's story are credible, and what is already in the price? Those are not formatting questions. They are the research.

Bigdata.com's reference example is an NVIDIA initiation report of about 2,600 tokens at $1.70 per run. The pricing page also breaks that run down by content type, which shows where the money actually goes:

Premium news, holdings, hiring and ESG data make up more than 90% of the cost. Filings and transcripts, the backbone of most initiations, cost a few cents. The value of the workflow is less the headline figure than the repeatability: once the questions and source rules are set, the same structure reruns when a new filing or event changes the evidence.

Thursday: the long thesis memo and a short-thesis pressure test

A thesis memo asks the agent to organise evidence around a view. A pressure test asks the opposite: find what would make that view fail. The second is often more useful, because it forces the process to hunt for contradictions instead of rewarding a prompt that already leans one way.

For a long memo, the agent collects the evidence behind the main drivers, risks and catalysts, then drafts a structured case with citations. For a short pressure test, ask for the strongest evidence against the short as well as for it. That gives the analyst something to challenge, not just agree with.

Bigdata.com's reference examples are a Tesla long-thesis memo of about 1,950 tokens at $1.77 per run, and a Delta short-thesis test of about 1,560 tokens at $1.82. The shorter note is the most expensive run of the week, a reminder that what you ask the agent to look for shapes the bill more than how much it writes.

Friday: competitive landscape and regulatory monitoring

Friday is for the work that keeps the coverage current rather than the work that makes headlines. Two jobs fit an agent well: a competitive landscape brief that tracks what peers announced, hired for and said this week, and a regulatory and policy monitor that flags rule changes, filings and consultations touching the coverage list.

Both are monitoring jobs with a fixed set of questions, which is where a reusable workflow pays off. The agent runs the same checks each week against fresh evidence; the analyst decides which changes alter a thesis or need a client note.

Bigdata.com's reference costs are $1.08 per run for a competitive landscape brief of about 1,300 tokens, and $0.63 for a regulatory and policy monitor of about 780 tokens.

What did the week cost in data search and retrieval?

One run of each workflow comes to $8.39 in data, using the per-run reference costs on Bigdata.com's pricing page (as of October 2026).

Nore that these are Bigdata.com's illustrative estimates for specific example workflows, not fixed quotes: your own cost moves with the sources and the amount of content your agent retrieves.

These figures cover the data your agent retrieves. The assistant's own model usage is billed separately by its provider.

Why does cost follow content, not length?

Because Bigdata.com charges for the content tokens an agent retrieves, not for seats or for the size of the note it writes. In July 2026 the company launched what it calls the tokenization of content: every source is split into the same tokens an AI model reads, and each provider sets its own price per token.

When the agent asks a question, retrieval selects only the passages that answer it, and those tokens are counted and settled at each provider's rate. RavenPack's example compares pushing a whole document of roughly 20,000 tokens into the model with retrieving the roughly 200 tokens that carry the answer, a reduction in paid context of up to 100 times.

That explains the pattern in the table. A short note that draws on premium news, holdings and hiring data costs more than a longer one built mainly on filings and transcripts. It also keeps the assistant's own bill down, since smaller context means cheaper inference on every step an agent loops through. Each retrieved passage carries its licence and citation with it, which matters when the output goes to an investment committee.

What did the agent save time on?

The pattern holds across the week: the agent takes the first pass at work that is repetitive but evidence-heavy. It collects, normalises and assembles, and on recurring work the same research plan runs again without rebuilding the process.

• Scanning markets and the coverage list against the same questions each Monday.

• Pulling reported figures and management commentary into a standard earnings template.

• Rerunning an initiation structure when new primary sources arrive.

• Keeping a thesis memo tied to the evidence behind each claim.

• Running a deliberate challenge against an existing view.

• Tracking peers and regulators with a fixed weekly checklist.

The time saved only counts if the review stays in place. Taking the analyst out of the last mile swaps visible manual work for less visible model risk.

Where did the analyst still have to step in?

At every point where the output moved from "what the sources say" to "what we should believe." The analyst chose the questions, decided which sources counted, checked the figures, challenged the interpretation and decided what could go into a client or investment-committee document.

That split is the right way to think about an AI agent for equity research. It is not an autonomous stock picker. It is a reusable research worker that keeps the evidence moving, at a data cost measured in dollars per week, while the analyst keeps the judgement.