01 The problem
A private equity fund managing over $13 billion in assets had a bottleneck that lived at the top of the org chart. Before every potential deal, partners spent six or more hours gathering market intelligence, analyzing competitive landscapes, and preparing briefing documents for investment committee meetings.
The work followed the same shape every time. Pull the market data. Read the filings and the news. Map the competitive set. Check what the firm already knew. Then compress all of it into a document the committee could act on. Hours of collection and formatting, wrapped around a few minutes of actual judgment.
Across 50+ investment professionals in multiple offices, that work added up to a serious tax on the firm's most expensive hours. It also produced uneven output. Quality depended on who prepared the brief and how much time they had. A brief assembled the night before a committee meeting did not look like a brief assembled across a quiet week. Manual coverage left occasional blind spots in the market analysis, and nobody could say in advance where the next one would be.
The firm did not need partners doing research. It needed partners in rooms, building relationships and negotiating deals.
02 Who the client is and why this matters
The fund manages over $13 billion from multiple offices. At that scale, the scarcest resource is not capital. It is partner attention. Capital can sit and wait. A partner's calendar cannot. Every hour a partner spends pulling comps and formatting a document is an hour not spent with a founder, a banker, or a management team, and those hours are the ones that actually move deals.
The multi-office structure added a second problem: consistency. Each office had its own habits for preparing intelligence. Different people leaned on different sources, structured documents differently, and went deeper or shallower depending on the week. The investment committee saw all of it. When every office prepares briefs its own way, the committee spends part of every meeting adjusting to the format instead of weighing the deal.
There is a quieter cost too. A committee that has learned to discount rushed briefs starts reading every brief defensively. Partners pad their preparation to protect their credibility in the room, which pushes the six hours toward seven. The whole loop runs on the most expensive time in the building, and it compounds: more deals means more prep, which means either more late nights or thinner coverage. Headcount does not fix it, because the bottleneck is not hands. It is the structure of the work.
So the firm was paying twice. Once in senior hours spent on assembly work, and again in the friction of inconsistent output. Neither cost showed up on a budget line, which is exactly why it had survived so long.
This pattern is not unique to this fund or to private equity. Any partnership where the most senior people personally assemble information before they can use it has the same bottleneck. This case is what it looks like to remove it.
03 What had to be true first
Before we wrote a line of code, three constraints were fixed. They shaped every architectural decision that followed.
Proprietary data never leaves the firm's control. The fund's deal history, internal notes, and models are the firm's edge. They had to stay in stores the firm controls, with model access running under API terms that do not train on client data. And inside the firm, information walls between deal teams are a real obligation, not a courtesy. The system had to be built to respect those walls: one team's data invisible to every other team by construction, not by policy.
Briefs needed committee-grade structure. The investment committee already knew what a good briefing document looked like. The system had to produce that document, with the same sections in the same order every time, not a chat transcript a partner would have to reshape before the meeting. This ruled out the generic chat interface from the start. A tool that returns prose a partner must then restructure has not removed the assembly work. It has moved it later in the evening.
Zero tolerance for uncited claims. A committee cannot act on "the model says". Every statement in a brief had to trace back to a source a partner could check. If a claim could not be traced, it did not ship. This constraint did the most to shape the build, and it is the first thing we now tell any firm evaluating a system like this.
None of these constraints came from us. They came from how the firm already worked: its obligations, its committee, its standards of evidence. That is the right order. A system designed around the firm's rules gets adopted. A system the firm has to bend its rules around gets piloted, praised, and quietly abandoned.
04 What we built
We designed and deployed an AI deal desk: a system that owns the intelligence-gathering layer of deal preparation, end to end.
The system has five working parts.
Live ingestion. The system pulls market feeds from 20+ data sources in real time, continuously, not on request. By the time anyone asks a question, the reading is already done. That is the only way a 20-minute brief is possible.
Synthesis across sources. Proprietary firm data and public sources get combined into one view instead of living in separate tabs. The synthesis layer is where the deal desk earns its keep: it reads across everything the firm is allowed to see and writes one account of what matters, with every line traced to where it came from.
Automated briefings. The system generates M&A briefing documents with competitive analysis built in, formatted for the investment committee. Same structure, same depth, whichever office asks. The brief a partner gets in one city is built the same way as the brief a partner gets in another, which is the consistency problem from section 02 solved as a side effect of the architecture rather than by a memo asking everyone to try harder.
Proactive risk flags. Emerging risks and market shifts get surfaced as they appear, before they become surprises. The monitoring loop never stops watching the feeds, so a development that lands between committee meetings reaches a partner when it happens, not when someone next runs a search.
On-demand reports. Executive-ready output, delivered when a partner asks for it. Not a search box that returns links, but an answer with its sources attached. The difference matters in practice: a search result hands the assembly work back to the person asking. A cited answer finishes the job.
The design principle was simple: every hour a partner spends assembling information is an hour the system should own. Judgment stays human. Collection and first-pass analysis do not.
05 The technologies behind it
Three tools carry the build, each chosen for a specific reason.
Claude, for long-document synthesis. Deal intelligence is a long-document problem. Filings, transcripts, news archives, and internal memos do not fit in a context the size of a sticky note. Claude is our default for reading across long material and producing summaries that stay faithful to the source, and its API terms do not train on client data, which constraint one demanded.
Supabase, for deal-team scoping. The firm's data lives in Postgres with row-level security doing the access-control work. The walls between deal teams are rows and policies, enforced by the database on every query, not a sharing setting someone has to remember.
Modal, for continuous ingestion. The 20+ feeds run on scheduled pipelines with no servers for anyone to babysit. Adding a source or changing a schedule is a small, reviewable change, which matters when the honest answer to "how often does this run" has to be "continuously".
Just as telling is what the build does not use. There is no fine-tuned model, because retrieval over the firm's own data solves the problem without creating a copy of that data inside someone else's weights. There is no agent framework that hides behavior from the client, because a fund that demands citations from its briefs should be able to demand them from its vendor.
This is the same stack we run on every build, for reasons we publish. The full list, with the reasoning attached, is on the stack page.
06 The engineering decisions that made it reliable
A demo of this system is easy. A version a $13 billion fund treats as its default first step is not. The difference is a handful of decisions made early, none of them glamorous, all of them about what happens on a bad day rather than a good one.
Citation-first synthesis. The pipeline retrieves first, then writes only from what it retrieved. Every claim in a brief carries a reference to its source, and a statement that cannot be traced is dropped rather than shipped. This sounds like a feature. It is actually an architecture: bolting citations onto freeform generation after the fact does not work, because the text was never anchored to the sources to begin with.
Per-deal-team access scoping. Access control lives at the database layer, with audit logs on access. A query from one deal team cannot touch another team's data, no matter how the question is phrased, because the rows are invisible to it. Built to respect information walls means the wall holds even when a person makes a mistake.
Continuous ingestion instead of on-request. We chose to ingest constantly rather than fetch when asked, for two reasons. Latency: a 20-minute brief is only possible if the reading happened before the question. Coverage: a system that only looks when asked can only find what someone thought to ask about, and the whole point of risk monitoring is catching what nobody asked about. The trade-off is real, since continuous pipelines cost more to run and more to maintain than a fetch-on-demand script. For a firm whose partner hours are the constraint, the trade is not close.
Structured brief templates over freeform generation. Every brief uses fixed sections: market view, competitive landscape, risk flags, sources. Freeform output reads differently every time, which forces the committee to re-orient on every document. A fixed structure means partners know exactly where to look, and it makes weak sections obvious instead of buried.
Graceful degradation when a feed dies. External data sources fail. Feeds change formats, rate-limit, or go quiet without notice. When that happens, the system says so: the brief states which source was unavailable instead of papering over the gap, and the failure is flagged for repair. A partner reading the document knows exactly what the system did not see. Partial coverage presented honestly is useful. Partial coverage presented as full coverage is how trust dies.
07 Before and after
The cleanest way to see the change is to put the two timelines side by side.
Before, a partner or a senior analyst worked through the same manual sequence for every deal: gather the raw material, analyze it, assemble the document. Six or more hours, mostly spent on the first and third steps, which involve no judgment at all.
After, the partner's role starts where it should: with a question. The deal desk does the gathering and the first-pass analysis it has already been doing continuously, assembles the brief in the committee's format, and hands it back in about 20 minutes. The partner reads, checks the cited lines that matter, and walks into the meeting having spent their time on the deal instead of the document.
Nothing about the firm's judgment moved. The committee still decides. The partners still own every position. What moved is the assembly work underneath the judgment, which was never the part anyone was paid for.
08 How the firm runs it
The system is not a side tool. It became the default first step in deal preparation.
The deal desk now handles 100+ queries daily across the firm. A partner heading into a committee meeting asks for a brief and gets it in minutes, with consistent structure and consistent depth, regardless of which office or which team is asking.
Adoption did not need a campaign. The system won by being the fastest route to a better document: once the first briefs held up in committee, asking the deal desk became the obvious first move, and the old way started to look like what it was, six hours of manual assembly for a worse result. Coverage is now comprehensive across all deals rather than rationed by whoever had research capacity that week.
The division of labor is clean. The system reads, watches, assembles, and cites. People ask, check, decide, and own the outcome. When a risk flag arrives, a partner judges whether it matters. When a brief lands, a partner verifies the lines the decision will turn on, which the citations make a minutes-long job instead of an act of faith. Nothing reaches the committee without a person who can defend every line in it.
That consistency is its own result. The firm no longer has good briefs and rushed briefs. It has one standard, and the committee can feel the difference in every meeting.
09 Results
Intel preparation dropped from six hours to twenty minutes per deal, a 95 percent reduction. Partners regained 12+ hours per week for the work that actually moves deals: relationships and negotiation.
- Intel prep: 6 hours down to 20 minutes per deal
- 12+ hours per week back per partner
- Consistent, comprehensive coverage across every deal
- Risks flagged proactively instead of discovered late
- 100+ queries handled daily across the firm
The numbers came with receipts. Because every brief cites its sources and the prep workflow was measured before the build, the firm can defend both halves of "6 hours to 20 minutes".
The pattern transfers. Any firm whose senior people spend hours assembling information before they can exercise judgment has the same system waiting to be built.
10 What to know before building this
If you run a fund and this case reads like your firm, three lessons from this build are worth taking with you.
Start where senior hours meet assembly work. Not the flashiest workflow, and not whatever a vendor happens to sell. Find the place where your most expensive people personally collect, format, and re-key information before they can exercise judgment. At this fund it was deal prep. At yours it might be LP reporting or diligence summaries. The economics are the same: the system buys back the hours that cost the most.
Demand citations from day one. If a system, or a vendor demo, cannot show you exactly where a claim came from, walk away. Citation-first synthesis is a build decision, not a setting, and it is nearly impossible to retrofit. It is also the only honest answer to a committee that will not act on a black box, and yours should not.
Measure prep time before you start. You can only say "6 hours to 20 minutes" if someone measured the six hours first. Baseline the workflow before any build begins: how long, how often, and whose hours. Without it, the after state is a feeling. With it, the after state is a number your CFO can check.
There is a fourth lesson hiding in section 03: write your constraints down before you talk to anyone. Data control, output structure, citation standards. The firms that get good systems are the ones that show up knowing what has to be true.
11 Related work
This build is our custom agents service running in production at fund scale, and its reporting layer is the same pattern as the operations hub: numbers from systems, narrative assembled around them, humans approving the output.
For the sector view, including how builds like this are designed around information walls, LP confidentiality, and Canadian privacy law, see AI for private equity. For another build where the constraints did the designing, this time under solicitor-client privilege, read the legal document automation case. And the reasoning behind every tool named above lives on the stack page.