A question lands in your inbox. You just got access to Claude Code, Codex, Claude Finance, or a similar model, and a dangerous (but natural) thought crosses your mind: Why am I paying for all these data tools? I have intelligence bottled; I will just tell the model what I want and build my own research stack from the bottom up.
Owning your full stack feels like leverage: the moment you stop renting, start to own, and build up equity.
If you’ve had this thought before, this may be the moment you quietly start a transformation from research analyst to data engineer… But three questions later, you’ve received an unrequested promotion to senior data engineer. Compensation package: frustration and token bills.
Two Research Prerequisites
In fairness, these models are some of the most incredible advances we've seen, and their value is real. Using them well isn't about burning fewer tokens or swearing off what's genuinely magic in an API. It's about aiming that capability at the work that amplifies your job, not the plumbing you never had to think about before.
AI hype aside, analysis and research still rest on two prerequisites: good data and a knowledge rationale. AI has changed who does parts of the reasoning in exchange for tokens. Despite model performance enhancements, output is capped by the quality of data going in and the work spent getting that data ready and enriched before a question can be asked.
A quick dose of reality: adding a model to the equation does not remove the prep cost; it simply makes it your responsibility. The hours spent by data vendors, your library team, and your junior analysts to clean, tag, and connect documents do not disappear into the collective experience of AGI. The cost is still there: waiting, taxing, and bringing in your own documents. But now you pay in terms of tokens, latency, and derived hours every single time you ask a question.
The Knowledge Trap
To be clear, I am no Luddite and enjoy using these models probably more than I should. Bringing your own document to a capable model to build a bespoke research stack is genuinely a functional design. Download S-1 from EDGAR, drop it into the context window, and in less than two minutes, you get a sharper summary than a first-year analyst could produce in an entire day. That first query gives you a response that feels like magic. And that first success is exactly why this trap works.
Why is this a trap? Good research is never one query. Reading the S-1 does its job, which then elicits subsequent questions. Each better question starts the meter without consideration of cost. That meter has a name: query-time recomputation.
Hidden Tax in Every Incremental Question
Consider the SpaceX IPO. Thousands of research analysts globally wanted to understand it. Thousands of corporate strategy and investor relations teams were tasked with evaluating what this historic IPO meant for their business. Perhaps for similar reasons you read the S-1. Maybe you asked your model to summarize the S-1. This is just the beginning of the real work.
For better context, you wonder: ‘How do the launch economics compare to the rest of the market?’ ‘Who are their peers, what are their weaknesses and strengths, and how is management projecting growth?’ ‘Who else is foiling, raising, or guiding on satellite or AI tech?’ ‘What did relevant expert calls say about pricing pressure, long-term margins, and customer satisfaction?’
In a bring-your-own-infrastructure workflow, each new set of questions is a set of fresh extract-transform-load (ETL) jobs that you now own end-to-end. You go to EDGAR and hope the web fetch landed on the right filing instead of an exhibit index. You wrestle the 300-page doc through the AI PDF parser and pray that the financial tables survive contact with it. You broaden the scope to every company that touches satellites.
Suddenly, you are burning through billions of tokens and real hours just to find the actual documents, download them, parse them, format them, and save them as residual metadata. You do all this before spending a single token (human or model-made) toward the analysis you actually care about. And every incremental piece of information sucks you back into dealing with your pipeline: Discover, fetch, validate, parse, validate, deduplicate, store, repeat.
As anyone who has ever run ETL pipelines knows, failure is inevitable. When your role is research and analysis, you aren't thinking about observability in your pipelines. Your failures are silent because nobody installed observability in the hobby research platform you accidentally built between earnings season and lunch. You miss a filing; the crawler silently grabs the wrong year; the file download is corrupted; the parser times out; and your system doesn't throw an error because you didn't know that it should. In bringing your own documents, the tools have conscripted you into logistics, and the logistics fail in ways you cannot see and won't know how to fix.
Compounding Taxes
The tax of accidentally becoming a data engineer is worse than it first seems because it's not a one-time tax. Instead, it compounds in three ways. First, the corpus gets re-fetched, re-processed, and re-extracted for every new angle you take. Every time you have a new question, data is discovered, processed, and extracted again.
Second, it compounds across people: That analyst one desk over is at this very moment duplicating your questions and work on SpaceX. So is every other fund in the world. Third, this tax compounds over time. Nothing that was used today was saved as a reusable asset. You plumbed the data, some of that plumbing worked, and when you were done, you threw the plumbing stack away.
A Proprietary Intelligence Warehouse
This isn’t a new problem, and data warehouses figured out a proven approach years ago: Decouple data, storage, and compute, allowing analysts to focus on the message in the data instead of the pipelines. This is what has propelled the likes of Databricks, Snowflake, and ClickHouse to their billions in revenue and thousands of happy customers. Nobody looks at Snowflake and asks if it would be better if everyone rebuilt the warehouse on their laptop every morning. AlphaSense follows the same principles to create its intelligence warehouse, building on top of the world's research context.
Within AlphaSense, the preprocessing happens once at the platform level, before any question is asked. While you are sleeping, eating, commuting, or doom-scrolling, our platform is continuously discovering content and transforming each unique piece into the most research-ready form possible. Raw documents are enriched with extracted entities, topics, and reliance signals. This is the most important part of the whole process: Each new document has platform-level context ingrained within it. Fillings, broker research, earnings call transcripts, expert calls, and IP arrive in a structured format ready for analysis and synthesis.
Each subsequent question lends the whole corpus at your fingertips and is deeply connected. Widening the scope from SpaceX to every company in the satellite supply chain is not a new ETL job you have to design, plan, and implement; it's just a query. The recall set, which the models get to reason over, is narrower, richer, and cleaner. This makes it not just tidier, but cheaper and faster. Time to insight is bounded by your questions and reasoning — no pipelines needed. AlphaSense’s intelligence warehouse is the modern Library of Alexandria: allowing researchers to gather, organize, and connect information in an instant and ultimately devote more time to higher-value analysis.
The Cheapest Token Is the One You Never Spend
A recent argument contends that Wall Street is buying AI the wrong way: metering and billing by the token. It points out that tokens consumed bear "little relationship" to outcomes delivered, and elucidates that "tokenmaxxing is not transformation.” I agree with this sentiment deeply. Where I diverge is in the prescriptive fix, the billing layer. Selling intelligence by the unit of value and routing the task to the cheapest adequate model requires building model brokers whose margins measure success.
A model broker that routes the ten-thousandth re-parse of the same S-1 to a 50% cheaper open-source model instead of a single Fable call is dollar-foolish. It is still parsing the same S-1 for the ten-thousandth time. Wrap this inefficiency in pass-through pricing or dress it in a nice suit and fiduciary language, but you will still be paying for discovery and ETL, which should have been done once and never again.
The cheapest token is not the one you bought at a 10x discount by learning how to find a foundational model doppelgänger. It is the one you never had to spend.
This is why the content layer sits upstream of the pricing layer, and why arguing about how to bill for intelligence before you have removed the rote work is optimizing the wrong end of the pipe. An intelligence warehouse collapses marginal cost: Transform the corpus once, and the thousandth question costs a fraction of the first.
If the real alignment requires "going deep into specific domains and living inside customers' workflows and economics over many years,” this is not an innovation in billing, but the job description for building an intelligence warehouse. The domain depth described is achieved by turning raw content into structured, reusable intelligence for thousands of customers over many years. This domain expertise is exactly what allows an analyst to arrive at a question and not have to remanufacture the data from a bunch of PDFs. The warehouse is what makes outcome-driven pricing possible in the first place.
You Are Not Cheaper Than the Warehouse
Cutting out the middleman to save time and money seems like a good idea at first. But someone still has to turn raw content into something usable, and by signing up for the job, you become the middleman on every question, paying the same tax over and over, forever. Unlike other solutions, AlphaSense is not a model wrapper with access to precomputed answers. It continuously turns the flow of raw content into a structured, reusable intelligence warehouse. Tokens are spent on analysis rather than discovery and data prep. AlphaSense is the intelligence warehouse that allows you to function as a researcher or analyst instead of an ETL engineer.
You can absolutely bring your own documents into a model and get great responses, but the tax will only be deferred on every query. To make it work, you will just keep getting conscripted into the pipeline business until you realize your job has quietly become the one you were trying to avoid.
No CFO would sign off on rebuilding the warehouse every time. Neither should a researcher.
