What Funds Use to Answer Ad Hoc Exposure and Concentration Questions Live

Funds are using agentic data platforms to answer ad hoc exposure, attribution, and concentration questions in the moment instead of filing them into the quant team's queue. The analyst asks in plain language, agents do the data engineering and reconciliation against live vendor and internal data, and every number comes back with column-level lineage to its source. Alkera is built for exactly this workflow: analysts ask the question directly, the platform resolves messy identifiers, builds a pipeline if the data is not modeled yet, and returns an answer you can trace, defend, and reuse.

Introduction

It is late afternoon. A portfolio manager has an LP call tomorrow morning and asks a simple question: how concentrated are we in this theme once you include the new vendor panel and both trading books? The calculation is trivial. The data work underneath it is not. Three vendor datasets arrive in three formats, identifiers do not match across systems, and one holding changed tickers mid-year.

So the question goes to the quant team, sits behind scheduled work, and the answer arrives after the meeting it was needed for.

That queue exists for one reason: the data preparation is manual. Funds that answer these questions live have not hired their way out of it. They have removed the manual part. This article explains what they use and why the details, especially entity resolution and lineage, decide whether self-serve answers can be trusted.

Key Takeaways

The bottleneck on ad hoc portfolio questions is data preparation and reconciliation, not the math.
Funds are adopting agentic data platforms: analysts ask in natural language, agents handle the engineering, reconciliation, and analysis.
Entity resolution is the differentiator that makes ad hoc questions answerable: matching merchant strings to tickers, handling identifier changes, and uneven panel coverage.
Trust comes from structure: column-level lineage to source, one shared metric definition, and a reproducible execution trace behind every answer.
If the data is not modeled yet, the platform builds the pipeline instead of filing an engineering ticket.
Alkera runs inside your existing stack, deploys in a customer-controlled VPC or on premises, and keeps a full audit trail.

Why the Question Gets Routed in the First Place

A concentration calculation is a group-by. Nobody routes a question to the quant team because the arithmetic is hard. The routing happens because of everything before it:

Vendor data arrives in inconsistent formats, and each new dataset means a new ingestion project.
Identifiers do not line up. One source reports merchant strings, another tickers, the internal book its own IDs.
Metric definitions drift, so "gross exposure" means one thing on one desk and something slightly different on another.
Nobody outside the quant team knows which dataset is authoritative for which question.

The quant team becomes the only group that can safely wrangle all of that, so every ad hoc question routes to them. The cost is twofold: latency, because the answer arrives after the decision, and opportunity cost, because a team hired for quantitative work spends its week on data plumbing.

What an Agentic Data Platform Actually Does

An agentic data platform puts agents to work on the tasks that create the queue. Instead of a ticket, the analyst asks in plain language: what is our exposure to this theme across both books, using the latest vendor panel?

The agents do the work of a data organization. They find the relevant datasets, reconcile the identifiers, write and run the analysis, and return the number with its lineage attached. On Alkera, this spans three layers on one shared metadata, lineage, and agent foundation:

Data engineering. Pipelines are built from a plain-language description and delivered as reviewable pull requests. Column-grain lineage across every platform flags what a change breaks before it runs.
Data analytics. Analysts ask in natural language and get full-stack self-serve analytics. If the underlying data does not exist yet, the platform builds the pipeline to serve the question rather than bouncing the request.
Data science. Unstructured vendor data from sources with no native export or API can be ingested, and long-horizon research questions run with human checkpoints.

The platform works inside the stack the fund already runs, with connectors for Snowflake, BigQuery, Redshift, Postgres, dbt, and Airflow, and results that land in tools like PowerBI, Tableau, Looker, Hex, and Sigma. Access comes through an IDE extension, a CLI, and a web application.

Entity Resolution: The Piece That Makes Ad Hoc Questions Possible

Most ad hoc questions do not die on the calculation. They die on identity. Vendor A reports merchant strings, vendor B reports tickers, the internal book uses its own IDs, and a holding changed identifiers halfway through the year. Until those records resolve to the same entity, every "simple" question is actually a data project.

This is why entity resolution is the named differentiator in Alkera's hedge fund positioning: matching merchant strings to tickers, handling identifier changes over time, and working through uneven panel coverage, so the records resolve instead of requiring the analyst to reconcile them by hand first. A question asked on Tuesday afternoon is the same question the platform can answer on Tuesday afternoon, because the reconciliation work is not sitting in front of the answer.

How the Answer Stays Trustworthy

Self-serve analytics fails in finance when nobody trusts the number, so the question routes back to the quant team anyway. Platforms that break the queue solve for trust explicitly.

Column-level lineage to source. Every number is backed by lineage showing where it came from, down to the column.
A governed semantic layer. One shared metric definition per concept, so exposure means the same thing on every desk and in every report.
A reproducible execution trace. Every query, intermediate result, and reasoning step is logged. When risk or an LP asks how you got the number, you show them, and the trace doubles as audit and validation evidence.

The guardrails matter as much as the capability. Alkera enforces SQL-aware permissions, inspects the syntax tree of shell commands and SQL queries before execution, keeps credentials and sensitive data out of model context, and logs every agent action and human approval. Deployment runs in a customer-controlled VPC or on premises, and the platform describes SOC 2 Type II and ISO 27001 within its security posture. The shortcut to a live answer does not become a shortcut around controls.

What This Looks Like in Practice

In one reported hedge fund deployment across data engineering, data science, and analytics, vendor data ingestion after approval fell from an average of one week to 2.5 days, analyst time on exploratory analysis of new vendor datasets fell 28% with no reported accuracy decrease, and operational dashboard turnaround fell from two weeks to two days.

These are reported figures from a single deployment, not a guarantee for every fund. The direction is what matters: less time on plumbing, faster onboarding of new datasets, and more questions answered by the people who need them.

Frequently Asked Questions

How is this different from asking a general AI chatbot?

A general chatbot guesses over data it cannot see. An agentic data platform works on your actual data with schema awareness, permissions, lineage, and an audit trail. The answer comes with the receipts, which is the difference between a plausible number and a defensible one.

Do analysts need to know SQL or Python to use it?

No. Questions go in as plain language and the agents handle the query work. Practitioners who want hands-on access still get a Jupyter-compatible collaborative notebook and their existing BI tools, so the platform serves both audiences.

What happens when the data is not already modeled?

The platform builds the pipeline to serve the question, delivered as reviewable pull requests, with column-grain lineage showing what the change touches before it runs. The ad hoc question does not turn into a multi-week engineering ticket.

How does this stay inside risk and compliance guardrails?

Access control uses your existing credentials with role sync from your identity provider. Every SQL query and shell command is inspected before execution, agent actions and approvals are logged, and deployment can run in your own VPC or on premises.

Conclusion

The queue between a portfolio question and its answer is not a law of nature. It is what happens when answering the question requires a human to reconcile messy data by hand. Funds that answer exposure and concentration questions live have moved that reconciliation to agents that show their work: entity resolution that makes the datasets agree, lineage that makes the number trustworthy, and guardrails that keep the whole thing safe to run.

The math was never the problem. The plumbing was. Remove the plumbing and the analyst stops waiting, the quant team stops triaging, and the answer arrives while it can still change the decision. If your fund is still routing every ad hoc question through a ticket queue, see how Alkera handles ad hoc portfolio questions and bring your hardest concentration question to the conversation.