AI tools for IMP's portfolio team: initial thoughts
Danny Quick, H20 Studios
One caveat up front: I'm not a catalog expert. What I know is how to build AI products that people trust with real work. Where I've made incorrect assumptions, correct me.
A lot of your team's time goes to work that comes before the actual investment thinking: reading contracts, matching payments to statement lines, working out which period a payment belongs to. AI is now good at exactly that work. An engineered system, with evals, permissions and checks wrapped around a model, can read a long contract or a royalty statement in minutes and show the clause or line behind every number it pulls out.
It can't make the calls your team makes. The point is to give the people who do make them more time, across more songs. That increases your operating leverage by orders of magnitude.
Why now
- Answer investor questions quickly and accurately, on schedule or on demand - BlackRock's Q4 seasonality question is a preview of what investors will keep asking, and answering it means tying each payment back to the period it was earned.
- Cover 30,000 songs without growing the team at the same rate - The tracking and reconciliation work grows with the catalog as you move beyond Fund I's three investments.
- Go into the next bid with your underwriting rules tested - Fund I now gives you actual results to compare with the sync discount and growth assumptions.
- Get ahead of the firms bidding against you - These tools only became good enough for this work in the past 12 months so most competitors haven't built it yet.
How the same team gets ahead: AI you can rely on
The fair objection to AI in finance is that models make things up. Two years ago that ruled out this kind of work. It doesn't anymore, and the reason matters.
The models did get better. But the bigger change is that hallucination turned into an engineering problem, and it's one we know how to solve. What you'd be getting here is not just a model. It's a system built around the model, with checks, evals and security, and the system is what makes the output trustworthy:
- Every value needs a source - The model has to quote the exact clause or statement line a number came from. Software then confirms that text really is in the document. No match, no entry.
- The model doesn't do the math - It reads and tags. Totals, growth rates and comparisons are calculated by ordinary code, the same way every time.
- The numbers have to add up - Extracted lines are checked against the statement's own totals. If they don't reconcile, the statement is held.
- Uncertain items go to a person - When the model isn't sure, for example about which song a line refers to, it asks instead of guessing.
- Accuracy is measured with evals - Before anything goes live, we test the system against examples your team has already checked, and keep testing as data and models change.
- Security is part of the system - Everything sits inside IMP's AWS account, with access controls and a full audit log of every model call and data change. More on that below.
Put together, this gets the error rate on what reaches a report down to near zero, and when something is wrong, the source is one click away. Building those checks is most of the work on a project like this, and it's what we do.
What this could look like
Four features, each with a mockup you can click through. All of the data is made up. After a discovery call, we'd sort these and others with you by impact and urgency.
No one has to remember a contract date
The AI reads every contract and pulls out the dates. Each one links to the clause it came from, and alerts go to whoever owns it.
Upcoming dates · Fund IAlerts at 90, 30 and 7 days
Click a date to see the clause.
The AI reads the statement. You approve what it's unsure about.
Most lines match on their own. The few that don't come to a person, with the source next to the AI's reading.
Administrator statement · H1 2026
Approve or answer to move through the queue.
Ask a question, see where the answer came from
Anyone on the team can ask in plain English. Every answer shows the contracts or statement rows behind it.
Ask the portfolioAnswers from approved data only
Click an example question to see another answer.
See what drove the gap between the model and actuals
The original underwriting model stays frozen. Collections are compared against it by income type, with a short explanation for each gap.
Deal A · model vs collected, 3 yearsModel frozen at close
Click an income type for the explanation.
Other things we'd consider:
- Seasonality and cash timing, with cash shifted back to the quarter it was earned. This is what would answer the Q4 question from IMP's own data.
- A check of each underwriting rule against Fund I history.
- A weekly email brief with upcoming dates, statements received or overdue, and items waiting for review.
- First drafts of the quarterly investor update, written from approved data.
- Song-level views: which songs are growing, which are slipping and which earn the most sync.
Security
Security decides whether this project can happen at all, so it comes first.
The usual worry is that using the best AI models means sending your data to an AI company. It doesn't have to. Amazon hosts Anthropic's Claude models inside AWS, through a service called Bedrock, and IMP can use them from an AWS account it owns. Here's what that gets you:
- IMP owns the account - It's opened in IMP's name and billed to IMP. We work through access IMP grants and can revoke. Everything stays with IMP when the project ends.
- No training on IMP's data - AWS's terms say prompts and outputs aren't used to train any model.
- Anthropic doesn't see the data - The model is hosted inside AWS. Anthropic doesn't receive or store IMP's contracts, statements or results.
- Data stays in one region - IMP picks the AWS region, and data moves over private network connections, not the public internet.
- Full audit log - Every login, model call and data change is recorded in IMP's account and can be shown to investors during due diligence.
- Model flexibility over time - Bedrock offers models from several providers behind the same account and security setup. IMP can move to a newer or better-suited model as they come out, without a rebuild or a new security review.
- How we handle data - Real client data stays in IMP's AWS account. It doesn't go on our laptops, into personal AI accounts or onto outside file-sharing tools. Demos built outside IMP's environment use sample data.
What the AI costs
IMP pays AWS for usage on its normal monthly bill, with no seats, minimums or separate AI contract. For Fund I, expect hundreds to low thousands of dollars a month. We'd firm that up after seeing real statements. We'd set spending alerts in the account from the start.
Next steps
- A discovery call, mostly for us to listen: how the work gets done today, what the contracts, statements and models look like, and what matters most to you.
- After that, a detailed proposal with scope, timeline and pricing.
About H20 Studios
H20 Studios is an AI product studio in New York City, run by Danny Quick. We partner with teams to build zero-to-one AI products.
Here's how we work:
- Prototype quickly - We build working prototypes early, so your team can react to something real.
- Test with evals - We measure the AI's output against examples your team has checked, so we know how accurate it is before anyone relies on it.
- Guards, checks and citations - Every figure links back to its source, and automated checks catch errors before they reach a person.
- Make it reliable - We keep testing as the data and models change, so results stay consistent.
- Simple interfaces - Review screens and reports are built around how your team already works.
- Turnkey - We handle the technology and infrastructure. Your team gets a tool that works and doesn't have to manage any of it.