The 6-month fund onboarding problem, and why "just use AI" is not the answer
By Deepak Sheoran, Founder & CTO, DwellFi
TL;DR. Onboarding a new fund client takes most administrators months of manual reconciliation, and it is the single biggest blocker on revenue in fund services. Managers usually know they have outgrown their administrator long before they move, because the fear of a botched, error-prone migration outweighs the discomfort of staying. The instinct in 2026 is to point a general-purpose AI tool at the problem, but that breaks on the hardest, highest-stakes parts of the work. What actually compresses the timeline from months to days is a vertical platform that computes the repeatable work deterministically, paired with fund-data engineers who own the exceptions and sign off on every step. We do not sell an agent. We deliver the conversion as a finished, audit-ready outcome.
Why fund onboarding still takes months
Ask anyone who runs fund operations to name their most painful workflow, and the answer is almost always client onboarding and fund conversion. It is the biggest pain in the business and the biggest blocker on revenue, and it stays broken for reasons that start with the documents themselves.
A single subscription document can run to a couple of hundred pages. It is not a form you sign but a stack that bundles the agreement, the side letter that quietly changes the economics, KYC and AML paperwork, tax forms, accreditation evidence, and entity formation docs. Then the prior administrator hands over the subscription file for the whole fund, which is often one PDF with a dozen or more LP packs merged into it, running to several thousand pages with no bookmarks and no index, and the conversion cannot start until someone has actually read all of it.
The documents are only the entry fee, because years of historical transactions have to move as well, not just current balances. The files arrive in dozens of formats, from scanned PDFs to sprawling Excel workbooks and legacy platform exports, and the GL accounts rarely map cleanly between systems. Topside entries sit in the financials with no matching ledger entry, because someone once made it balance by hand, and now you have to reconcile it. The timing makes all of it worse, since conversions tend to collide with tax season, exactly when the team is already buried closing books for everyone else.
The result is a migration that drags on for months, and the fear of getting it wrong is acute enough to keep managers frozen in place. Industry surveys of fund managers consistently find that the fear of a complex, error-prone migration is one of the main reasons they delay switching, even once they have clearly outgrown their provider. The pattern underneath it is familiar: a manager will tolerate an administrator they have outgrown for as long as the pain stays internal, and the moment it reaches investors, through a reporting error an LP actually notices, is usually when they finally decide to move.
So the tension underneath the whole problem is this. The GP wants to switch administrators now and does not want to wait, while the administrator cannot absorb the migration without freezing capacity, so the deal stalls and growth stalls with it.
Why more analysts cannot fix it
The old playbook was to throw people at it, and the trouble with that is structural rather than a matter of effort. It does not scale with headcount, because every new client needs another set of analysts running lookups and rekeying, so cost and time rise in a straight line. The errors also cluster exactly where they hurt, since a large share of the work is document extraction and manual entry, the same points where a transposed figure or an unmapped transaction type becomes a downstream NAV error. And the work cannot run when you need it most, because it depends on the same scarce capacity that quarter-end and tax season already consume.
This is not hypothetical risk. According to Ignites, the number of asset managers reporting NAV errors rose 29% in 2022 compared with the prior year, and 129 funds required NAV restatements, with turnover and complexity cited as primary causes. Striking a NAV is a daily chain of handoffs between a fund's managers and its service providers, and it takes only one wrong figure, caught too late, to produce an error that reaches investors.
Rules-based automation, the RPA and OCR pipelines, helped at the edges, but it breaks the moment a document format changes or a calculation is undocumented, and the rest still lands on a person.
Why a Claude, Copilot, or ChatGPT subscription cannot fix it either
This is the question every operations leader is actually asking in 2026, so let me answer it directly: if AI in fund services is this good now, why can I not just point ChatGPT or Copilot at my conversion?
The short answer is that fund data breaks general-purpose AI exactly where fund data is hardest, and it is a structural mismatch rather than a prompting problem you can engineer around.
The real-world artifact | What a general-purpose AI tool does | Why it is structural, not a prompt problem |
A single sub doc runs to a couple of hundred pages; the merged LP file runs into the thousands | Processes the easy parts and skims the rest, without telling you which pages it effectively ignored | The one side letter that flips the waterfall is exactly what gets lost, silently |
An Excel workbook with over a hundred tabs and tens of thousands of rows | Reads it partially, applies inconsistent logic across tabs, and returns a clean-looking total | The output looks correct while the number is wrong, the most dangerous failure in a NAV process |
An undocumented topside entry or a bespoke waterfall | Guesses the methodology and does not flag that it guessed | A wrong NAV looks identical to a right one |
Reconciliation across systems never built to talk to each other | Generates a plausible figure, with no deterministic compute layer underneath | Fund administration needs numbers that are computed, not generated |
None of this is a knock on the underlying models, which are genuinely good at what they were built for. They are simply the wrong tool for an exception-heavy, regulated process. A general-purpose chatbot is built to produce fluent output, while fund accounting has to produce output that is correct and reproducible, and in a period when NAV errors are already rising, the gap between fluent and correct is exactly where the damage happens.
What we built instead: a managed outcome, not an agent
So we did not build an agent and hand you the keys. We built a way to deliver the conversion as a finished, audit-ready outcome, and it works because two parts operate together.
The first is a vertical platform that does the heavy, repeatable work deterministically. It ingests the raw data from the prior administrator, whether that is historical portfolio and LP data, transactions, or subscription PDFs in mixed formats from several legacy platforms, and it builds a complete inventory of what is there and what is missing. It maps every GL account and every LP across source and target, reconciles the activity, and produces a GP-ready pack of trial balance, partner schedules, investment schedules, and reconciliation notes. The math compiles to code and is cross-validated, so there are no silent guesses inside the books.
The second is a team of fund-data engineers, our AI Pilots, who own the exceptions the platform surfaces. When a topside entry will not tie, or a forty-tab model will not reconcile, the platform flags it rather than guessing, and a specialist resolves it. That combination of a vertical platform and deployed expertise is what a general-purpose subscription structurally cannot give you. The agent is the engine, and the outcome is the product.
Is AI safe inside a regulated NAV process?
The objection I hear most is fair: speed is great, but I cannot put AI guesswork inside a NAV filing. I agree, and that is exactly why AI for a regulated environment has to be built differently from a consumer chatbot. It is deterministic by construction, so the math compiles to code and is cross-validated and there are no silent guesses inside the books. A human sits at every gate, which means the platform does the labor while your team owns every approval and sign-off. Every figure carries its provenance, so a reviewer can trace exactly how a number was derived. And it is sovereign by design, running against your existing accounting platform with no rip-and-replace, so your data never leaves your environment.
People are not the cost being cut here; the tedious, error-prone labor is. Remove that, and the same team onboards far more clients without giving up accuracy.
Who handles the mess, and why it still scales
A fair worry about any deployed-engineer model is whether it just means throwing bodies at each account, and it does not, which is the distinction that matters. The platform handles the overwhelming majority of the work deterministically and repeatably, and the engineers step in only where judgment is genuinely required, on the undocumented and the genuinely ambiguous. That is a small, high-value slice rather than the bulk of the job. Software carries the scale while people carry the edges, and that is what lets one team take on far more clients without re-hiring for every conversion, which is the opposite of a headcount-per-account model.
What faster onboarding changes for the business
When onboarding compresses from months to days, the economics of the whole business change. Revenue stops stalling, because you can say yes to the next client without freezing a quarter. Capacity multiplies, since administrators running this approach free up meaningful time and create real room to grow the book. And the tax-season collision fades, because conversions no longer compete with close cycles for the same hands. Just as important, you remove the single fear that keeps most managers frozen with an administrator they have already outgrown, the fear of migration errors, because when the migration is deterministic and human-signed, with every figure traceable, that fear stops being a reason to wait.
The bottom line
The six-month onboarding problem is not something you optimize your way out of with more analysts, and it is not something a general-purpose AI subscription can quietly solve in the background. It is a structural ceiling. Breaking through it takes a vertical platform that computes rather than guesses, run by specialists who own the exceptions and sign off on every step, delivered as a finished outcome on your data and in your environment.
Nobody wants to wait six months. With the right approach, nobody has to.
Frequently asked questions
How long does fund onboarding take today?
For most administrators it runs anywhere from a couple of months to half a year, depending on the number of LPs and how fragmented the prior administrator's data is. It is widely considered the most painful workflow in fund services, and the fear of migration errors is a big part of why many managers delay switching at all.
Can I just use ChatGPT or Copilot for a fund conversion?
Not reliably. General-purpose AI tools break on the hardest parts of fund data: merged document bundles running into the thousands of pages get skimmed rather than fully read, large multi-tab workbooks produce clean-looking but incorrect numbers, and undocumented methodologies get guessed at rather than flagged. Fund accounting needs numbers that are computed and reproducible, not generated.
Can AI actually do fund administration work, or just assist?
Done correctly, it does the work. A deterministic platform ingests, maps, reconciles, and produces a GP-ready pack, while fund-data specialists resolve the exceptions and a human approves every step. That is different from a chatbot that drafts text and leaves the reconciliation to you.
Is AI safe for regulated fund operations?
It can be, under specific conditions. The output has to be deterministic and reproducible, every figure has to be cited and traceable, a human has to sign off at each gate, and the system has to run in your own environment. Without those properties, AI does not belong near a NAV, especially now that NAV errors and restatements are on the rise industry-wide.
Do I have to replace my accounting platform?
No. The approach is designed to run against your existing accounting platform with no rip-and-replace, inside your own environment.
Want to see a fund conversion run on your own data? DwellFi will walk through it end to end and leave you with the output and the audit trail.