The crown jewel is the back office
I have spent most of two decades inside the operating stack of private markets funds. Fund finance, operations, procurement, a few consulting mandates, and for the last couple of years building AI tools against the same problems I used to own. I have signed capital account statements. I have run vendor selections where the losing bidder had the better product. I have sat on the buy side of an ERP replacement that took three years to pay back. I have watched a quarter-end waterfall break because one side letter lived in an Excel tab that nobody owned.
So when the labs announced they were coming for financial services, I read every release. And when I mapped where they landed, one thing stood out. They went for the easier part first.
This note is about why the private markets operating stack is the hardest AI market in finance, and who I think ends up holding it.
Public filings were the easier start
Anthropic launched Claude for Financial Services in July 2025, added Claude for Excel and connectors into Chronograph, Egnyte and others that October, and on 5 May 2026 shipped ten agent templates. A pitch agent and a KYC screener were among them, but so were a general ledger reconciler, a NAV check against the books, a month-end closer and a statement auditor. On 14 September it shipped Claude for Financial Advisors with Schwab and BlackRock data behind it. OpenAI answered on 10 September with ChatGPT for Financial Services, built on GPT-6 Astra with Morgan Stanley and Evercore as design partners, aimed at investment banking and equity research first.
The specialists moved in the same order. Rogo went from $2M of ARR in 2024 to more than $50M in 2026 and raised $160M in April at a $2B valuation, with Lazard, Jefferies and Moelis on the customer list. Harvey sits above $400M of ARR at a $15.5B valuation. Hebbia, the earlier mover, has not priced a round since its $700M Series B in July 2024.
Rogo and both lab products started on the sell side. Harvey started in legal, Hebbia in the diligence room. The first wave went where the raw material was already organised: EDGAR, XBRL and a decade of transcripts for the banks, the data room for the lawyers. A model can read all of it before anyone signs a contract. It is the Amazon playbook: books first, because books were the easiest category to catalogue.
Private markets have no EDGAR. The LPA is private by contract. So are the side letters, the capital call notices, the NAV packs, the borrowing base certificates and the waterfall model. The SEC’s Form PF data counts $29.6T of gross assets in US private funds at the end of 2025, hedge funds included. Preqin counts more than 29,000 active private equity and venture funds. The ground truth for almost all of it sits in an administrator’s ledger that no model has ever seen. So the labs have not ignored the back office. They have shipped templates for it. A template is a demonstration. The production workflow, with its approvals and its audit trail, is still the customer’s problem.
The gap between a pilot and an operating budget
Last week I wrote about the Anthropic S-1 and its $67B run rate priced at $1.5T, with OpenAI’s confidential filing behind it at more than $850B. Both prices assume the enterprise cheque keeps growing. Menlo Ventures’ State of Enterprise AI, published on 3 November, puts the enterprise AI deployment market at $177B on an annualised run-rate basis, from $33B a year earlier, and the labs took 69% of it. Anthropic and OpenAI together now hold 86% of enterprise API spend.
Read the next lines in the same report. The average enterprise spent 4.7x last year’s level, the median only 2.2x. Spend per enterprise nearly quintupled while adoption rose from 55% to 65%. Menlo calls it the fat head. Apollo’s Torsten Slok measured it from the other side: 95.6% of OpenAI and Anthropic spend sits with the top 10% of firms. And in Menlo’s $77.3B application market, coding takes $34.0B, chatbots $25.0B, healthcare $3.7B and legal $1.8B. Financial services does not get its own line. That is not zero spend, since banks and asset managers buy coding tools and chatbots like everyone else. It does show where the big application budgets sit, and fund operations is not among them.
Outside that top decile the cheque is small, and in my own procurement experience the first-year AI contract rarely clears $100,000. The public data points the same way, though none of it is a market-wide figure. Vendr’s median contract for Glean, across 174 purchases, is $98,890. Rogo’s $50M of ARR across 350 institutions is a rough $140,000 per institution. The Atlanta Fed puts the median US company below $200 of AI spend per employee. In the finance function, 28% of respondents to the CFO AI Leverage Report have no clear AI budget at all.
The reason is the output. McKinsey’s 2026 survey finds 37% of organisations report any EBIT impact from AI and 6% attribute 5% or more. KPMG finds 11% have scaled agents across the enterprise. Bain says 15% to 25% of CFOs have fully scaled AI inside finance. MIT’s NANDA group put the pilot failure rate at 95%.
In a fund I can tell you exactly where that gap sits. A chatbot drafting an IC memo produces a judgment a partner still owns, so deal teams adopted first. A capital account statement is a number the CFO attests and the auditor re-performs. It has to come out the same to the cent every time, with lineage an auditor can re-perform. A probability-weighted answer is the wrong shape for the arithmetic. AI belongs on the intake, reading the side letter and flagging the exception. A versioned calculation engine belongs on the number. A person signs. Menlo’s own test for where agents succeed is an environment they can act in, context they can learn from and a result they can verify. Fund accounting passes the third test better than almost any workflow in the economy. It fails the first two, because the contracts live in PDFs and the procedures live in people. I have not seen anyone build that environment at the fund level. The labs built templates. The incumbents built connectors.
The Palantir turn
The labs noticed. Lightcast counts forward deployed engineer postings up more than 1,000% between January and August 2026 against the same months last year, and 4,600% against 2023, at a median salary of $188,000. LinkedIn’s January report had FDE roles growing 42-fold between 2023 and 2025, against 13-fold for AI engineers. OpenAI put $1B into a dedicated FDE organisation in June. AWS matched the $1B on 30 June. Anthropic launched a $100M academy on 2 October to train 10,000 of them by the end of 2027. Rogo’s Series D release says the money funds embedded engineering teams at customer sites.
Then came the joint ventures. On 4 May 2026 Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a $1.5B services firm, named Ode in July, built on the Fractional AI team with 100 engineers at launch. A week later, on 11 May, OpenAI unveiled the Deployment Company. Axios reported $4B at a $10B pre-money, TPG leading with Advent, Bain Capital and Brookfield, a guaranteed 17.5% minimum return for investors, and Bain & Company, McKinsey and Capgemini writing cheques into the vehicle built to compete with them. On 23 October the FT reported Anthropic and Hg will open a European arm in London and Berlin before year end. Hg already employs about 150 AI specialists across its portfolio.
Palantir ran forward deployment for roughly fifteen years before the software scaled on its own, and its US commercial revenue grew 149% to $764M in the second quarter of this year. The model delivers day-one revenue and sticky accounts, at a services margin. At a median advertised salary of $188,000, an engineer dedicated to a customer paying $100,000 a year loses money before overheads, so the economics only work when the second and third fund implementations are easier than the first. The bet is that deployment data compounds into product. Gartner doubts it. Its August playbook expects 70% of enterprises to abandon vendor-built agentic AI by 2028, and fewer than 20% of FDE engagements to turn a recurring customer need into a feature of the vendor’s product. The labs copied the staffing model inside a single year. Whether they copied the patience is the open question, and the IPO price does not answer it.
Who holds the unfair advantage
Three groups claim it. The labs have the models and the capital. The consultancies have the access and the change programme, and they are leaning in hard: BCG expects 40% of its revenue to be AI-tied in 2026, Bain targets 50%, and McKinsey now prices about a quarter of its fees on outcomes. Source Global Research puts technology consulting at $420B in 2026, up 8%, so the pool they are defending is real. The operators have the judgment. They know which spreadsheet the waterfall actually runs on and which side letter breaks it.
Get Mohsin Abdul Qadir’s stories in your inbox
Join Medium for free to get updates from this writer.
My answer is that none of the three wins alone, and private equity is the one buyer that has already assembled all of them. Ownership is not integration, though. A sponsor’s stake in a vendor does not give it the right to use a portfolio company’s data, and it does not skip the security review. Blackstone co-owns Dynamo Software with Francisco Partners and co-founded Ode. Vista owns Allvue, with ARR above $200M, and hired Evercore and Barclays in August to test a sale at $2B to $3B. Hg owns a software portfolio of more than 60 companies and now owns a lab-backed consultancy to transform it. The strategy firms funded their own disintermediator rather than fight it. The same balance sheet now sits on the customer, the incumbent vendor and the challenger.
The poaching tells the same story. FTI’s 2026 survey names talent as the primary constraint for 35% of PE firms. Coller Capital’s summer barometer of 108 LPs managing $2.045T found 67% expect AI to widen the gap between the best GPs and the rest. The people AI was meant to replace are the people everyone is now hiring.
Eight versions of the same number
Here is the part of the argument I have not seen anyone make, because it is unsexy. The attention goes to the front office. Pitchbooks, IC memos, screening. That is where the demos happen and where Rogo and the lab products sell. The middle and back office get called plumbing.
The plumbing is where the stickiness lives. In every firm I have worked in, the same portfolio company EBITDA sat in at least eight places. The deal team’s model. The value creation team’s dashboard. The IR deck going to LPs. The valuation pack. Treasury’s cash forecast. Fund finance’s capital accounts. The administrator’s ledger. The GP’s own management accounts, and at a listed GP, the plc numbers on top. Eight versions, eight owners, eight refresh dates, reconciled by email at quarter end. The fund operating metrics never quite agree with the GP house metrics, and the lag between them runs to weeks.
That mismatch is the cost nobody budgets for, and it is also the prize. Whoever holds the book of record gets read by all eight teams. Menlo’s own finding is that incumbents hold their ground where they are the system of record. A fund’s ledger is as sticky as a system gets. The front office tool gets replaced when a better demo comes along. The record does not.
Why the operating stack is the crown jewel
First, the size of what the stack administers. Preqin’s five-year forecast, out last week, has private markets AUM above $34T by 2031, with private equity at $11.6T, private credit past $5T, venture at $3.79T and infrastructure at $2.74T growing 9.8% a year. Bain counts about 32,000 buyout-backed companies. PitchBook counts 13,325 unsold US portfolio companies. Opimas notes the top 7 PE firms hold more than 35% of AUM and the top 5 managers took 51% of 2025 fundraising, so the buyer base is consolidating and would rather run one vendor than twelve.
The software spend against that base is small and growing. Research and Markets sizes alternative investment software at $3.04B in 2026, reaching $4.82B by 2030 at a 12.2% CAGR, and wider definitions run to $5.5B. Gen II alone administers $1.5T. Allvue’s 500 clients against more than $200M of ARR works out at about $400,000 each. Intapp’s cloud ARR is $495.7M, with 142 clients paying more than $1M. Dynamo’s GP survey this summer found 78% of GPs expect to raise technology spend, with AI the top line item.
Those are the incumbents, and they sit in layers. DealCloud and Affinity on the pre-deal CRM. Intralinks, TermGrid and Leverest on execution. Allvue, eFront, SS&C, Yardi and Dynamo on fund accounting and administration, with BlackRock folding Preqin into eFront Insight in February. Chronograph, iLEVEL and 73 Strings on portfolio data and valuation. Oxane on private credit, where $1.5T of hosted notional and 950 staff make it a forward-deployed model that predates the labs by a decade. Hazeltree on treasury. iCapital and InvestCloud on the wealth channel. Carta on 9,000 funds and $203B. Several have already wired into a lab: Chronograph into Claude in October 2025, Carta’s Claude plugins this July, DealCloud’s Celeste into ChatGPT in September.
So the TAM is a few billion, which looks small next to the $177B the labs are chasing, and the administration fee pool behind it is larger than the software line. One commercial catch sits inside that. If AI removes 30% of the hours in a process the administrator runs, the GP does not get 30% off the administration fee unless the contract changes. The budget owner and the beneficiary are not the same person, and most AI business cases I have reviewed skip that. The moat is what makes it the crown jewel. Menlo found almost no evidence of enterprises ripping out Salesforce, Workday or NetSuite for AI, and a fund’s ledger will be at least as hard to displace. The data is contractual, the output is attested, the customer base is consolidating, and the renewal belongs to whoever holds the record all eight teams read. No one has shown that across the whole stack yet.
A question for the people who sign the cheques
If you are the CFO, COO or CIO of a GP, I want your answer to one question. If a vendor could validate your capital calls, NAV and waterfall against the approved terms and the books, with AI on the intake, versioned engines on the arithmetic and evidence your auditor could re-perform, would you buy it? What would you pay for it against your software and administration spend? And would you buy it from your administrator, an incumbent vendor, a frontier lab, or a specialist run by people who have closed a quarter? Before any of that, I would want a parallel run across two historical quarters and one live cycle, amended LPAs and side letter exceptions included, and I would want to see which cases the system refused to resolve without a person.
I am collecting the answers for the next note, which covers the wedge, the go-to-market, who the first customers should be, and what a right to win looks like against Aladdin, eFront, Yardi, Allvue and SS&C. It also carries my census of the 629 vendors in the Private Markets Technology Directory, including the two stages of the GP lifecycle that no software category covers at all. Reply here, message me, or leave a comment.
Sources: Anthropic and OpenAI product releases, Menlo Ventures State of Enterprise AI (3 November 2026), Apollo Daily Spark (24 October 2026), Lightcast and LinkedIn labour data via Fortune and CNBC, Axios on the Ode and DeployCo ventures, Forbes (29 October 2026), Gartner’s Forward-Deployed Engineering Playbook abstract, SEC Private Fund Statistics (Q4 2025), Preqin Private Markets in 2031, Opimas, Research and Markets, the Private Markets Technology Directory, company filings and press releases. Per-institution and per-client figures are rough averages from disclosed totals. Personal views only.
Questions this note answers
Why are private markets harder for AI than public markets? Public-market research starts from EDGAR, XBRL and transcripts that exist before a customer signs. A fund’s LPA, side letters, capital calls and NAV packs are private by contract, and the ground truth sits in an administrator’s ledger no model has seen.
What is the forward deployed engineer model, and why are the labs copying it? Palantir’s model embeds engineers inside the customer to build the workflow, with day-one revenue at a services margin. OpenAI, AWS and Anthropic each committed $100M to $1B to it in 2026, and Ode and the Deployment Company put private equity capital behind it.
Who are the incumbents in the private markets operating stack? DealCloud and Affinity on CRM. Intralinks, TermGrid and Leverest on execution. Allvue, eFront, SS&C, Yardi, Dynamo and Carta on fund accounting and administration. Chronograph, iLEVEL and 73 Strings on portfolio data and valuation. Oxane on private credit, Hazeltree on treasury, iCapital and InvestCloud on the wealth channel.
What should a GP CFO ask before buying an AI operating layer? Whether it removes a control step or adds a review queue, how much configuration each new fund needs, which exceptions still need a person, who keeps the saving between the GP and the administrator, and who is accountable when the number is wrong. Then run it in parallel across two historical quarters and one live cycle.
Originally published at https://mohsintcs.substack.com on November 10, 2026.
