AI for Investor Onboarding and KYC Documents: Costs and Limits

AI can now read an investor's KYC folder, classify the documents, extract the identity and ownership data, spot what is missing, and draft the chase email. What it cannot do is the AML sign-off, and it still makes too many mistakes on complex tables and multi-step calculations to run unsupervised. This post lays out the measured accuracy, the real costs, and the line between what to automate and what to leave with your compliance officer.

The evidence base is better than it was a year ago. Royal Bank of Canada published a peer-style technical paper on a production KYC pipeline. Gen II Fund Services, one of the largest private-capital administrators, is rolling out an automated onboarding platform globally. And several 2026 benchmarks finally measure extraction on finance documents rather than academic papers. The numbers below come from those sources, not from vendor marketing.

What KYC document work actually involves

Every new investor in a fund arrives as a folder. Subscription document, passports, proof of address, structure charts, source-of-funds evidence, for the entity and often for every beneficial owner. Someone reads each document, works out what is missing, drafts the "please also send" email, re-checks everything when it comes back, runs the names through screening, and then reads the screening hits to separate real matches from false positives. Multiply that by every investor in a close, against a closing deadline.

The RBC team behind the Spectra paper measured the shape of this work in capital markets onboarding. Their numbers give a sense of scale:

• More than 20 documents per client case.
• A taxonomy of more than 50 document types, with jurisdiction-specific requirements.
• Validation against compliance policy documents running to 400 pages.

That combination, high document volume plus long policy text plus zero tolerance for compliance error, is exactly where manual processing is slowest and where a naive "just prompt an LLM with the whole file" approach fails first.

What the measured evidence says AI can do

The Spectra paper is the most useful public data point because it describes a real pipeline, evaluated on real KYC documents, at a bank with genuine regulatory exposure. The system pairs a rules engine, which encodes compliance policy as a queryable database, with LLM agents for classification, extraction, and validation. Each stage is isolated and auditable, rather than one monolithic prompt. The reported results:

• 100 percent classification accuracy on real KYC documents.
• 89.4 percent extraction accuracy on a synthetic test set.
• A 96 percent reduction in human review burden.

Read those numbers carefully. Classification, deciding what a document is, is close to solved on this document mix. Extraction, pulling the right field values out, still fails on roughly one document in ten. That is why the review burden drops by 96 percent rather than disappearing: the human stops reading everything and starts checking a queue of flagged items.

The finance-specific benchmarks from 2026 fill in where the limits sit. FinSheet-Bench, built on synthetic private equity fund spreadsheets, tested ten frontier model configurations and found the best model, Gemini 3.1 Pro, at 82.4 percent accuracy overall, with Claude Opus 4.6 and GPT-5.2 close behind at about 80 percent. On the largest spreadsheet, 152 companies across 8 funds, average accuracy across all models fell to 48.6 percent. The authors' conclusion is blunt: no standalone model has error rates low enough for unsupervised professional finance use, and reliable extraction needs document understanding separated from deterministic computation.

A separate benchmark study of extraction accuracy makes the complementary point: the general document benchmarks are nearly saturated, but on the wrong corpus. Top systems score above 95 overall on OmniDocBench, yet financial reports are one stratum among ten document types in it, and the dominant table metric correlates with human judgement at only r = 0.68. On finance tasks, reading and calculating fail at different rates: the best model in the FAITH benchmark hits 97.0 percent on direct lookup from annual reports but only 80.0 percent on multivariate calculation.

The pattern across all three sources is consistent. Reading, classifying, and looking up values is strong. Extracting from complex tables and doing arithmetic across documents is not, and should be handed to deterministic code wherever possible.

What AI does in investor onboarding today

For fund investor onboarding specifically, the working slice looks like this:

• Classification. What is this document, whose is it, does it match what the subscription document requires.
• Extraction. Identity and ownership data: names, registration numbers, addresses, signatories, beneficial ownership percentages.
• Gap detection and chasing. Comparing the folder against the requirement checklist and drafting the chase email for whatever is missing.
• Screening-hit triage. Clustering the obvious false positives so a human reviews a shortlist instead of hundreds of raw alerts.

The notable production proof point is Gen II Fund Services rolling out Fenergo for client and investor AML/KYC onboarding, announced in November 2025. The rollout started in Europe, with all locations expected live by the end of 2026, and the headline gain is reuse of verified KYC data across multiple funds, so an investor already verified for one fund does not start from zero on the next. Gen II hosts on EU servers for GDPR and cites ISO 27001 certification and SOC I and SOC II compliance.

Two honest caveats. First, the Gen II figures are vendor-announced; there is no independent measurement of the AI layer in production. Second, an administrator adopting a platform is a different decision from a fund building its own pipeline, and the right answer for most funds is the former, as the build-versus-buy section below argues.

This is the same division of labour we implement when deploying Claude for financial services teams: the model reads and drafts, deterministic systems match and calculate, and named humans decide. Anthropic's own financial services page shows the same pattern at larger scale, with FIS building an agent intended to compress AML investigations from days to minutes, and identity data from Dun and Bradstreet used to keep agent outputs anchored to verifiable records. For what the packaged end of this market looks like, our breakdown of Claude finance agents and what the ten templates do and cost covers the off-the-shelf option before you consider anything custom.

What must stay with a human

The regulatory boundary is not blurry. The screening-hit disposition and the source-of-funds judgment belong to the money laundering reporting officer. That is a named person's regulatory responsibility, and a model output saying "this looks like a false positive" is an input to that decision, not the decision itself. Treating it as the decision is exactly the failure an examiner will find.

FATF's guidance on digital identity makes the same point for the identity layer: reliable digital ID can make customer due diligence easier, cheaper, and more secure, but the assessment of whether a given digital ID is suitable sits with the regulated entity. Automation moves the human effort; it does not remove the accountability.

A practical way to draw the line inside your own process:

• Automate fully: classification, completeness checks against a checklist, chase-email drafting, data entry into the register.
• Automate with mandatory review: field extraction, screening-hit clustering, risk-rating suggestions.
• Keep human: screening dispositions, source-of-funds and source-of-wealth judgment, the final AML sign-off, and any decision to decline or exit an investor.

What it costs: a worked example

Cost figures in this space are scattered, so here are the sourced unit economics followed by arithmetic at fund scale.

Per-check figures, from a pricing analysis of KYC automation: a manual KYC check costs roughly $5.10, an automated check roughly $0.14 at scale. Manual processing carries a 4 to 8 percent error rate, with each error costing $22 to $40 to correct. On time, one KYC cost analysis puts a mid-market operation processing 10,000 applications a month at $50,000 to $100,000 monthly on compliance salaries, and manual review at 15 to 45 minutes per application; an automated pipeline brings the machine-handled portion down to minutes.

Now scale that to a fund. Take a manager running three closes a year with 40 new investors each: 120 investor files annually, and at the RBC figure of 20 or more documents per case, roughly 2,400 documents to classify, read, and extract from.

• Manual baseline. At 15 to 45 minutes of review time per file before chasing and re-checks, that is 30 to 90 analyst hours for the first pass alone, and the real total is several times that once chase cycles and screening triage are included.
• With a good pipeline. At Spectra's 89.4 percent extraction accuracy, about 250 of those 2,400 documents will still need human correction. Budget for that queue. If your vendor promises 99 percent extraction on your document mix, ask for the evaluation set.
• Platform route. Onboarding platforms in the Fenergo, Anduin, and Passthrough class typically run to five figures a year at fund scale, priced per investor or per seat.
• Build route. A real build, extraction and triage wired to your register and screening with the compliance gate and data protection work included, runs five to six figures. Published project ranges for KYC automation put a pilot at $10,000 to $50,000 and a departmental rollout at $50,000 to $150,000.
• Ongoing run cost. LLM-heavy workflows carry token costs of $300 or more per month per workflow at scale, while the deterministic parts, completeness checks, format validation, watchlist cross-referencing, cost almost nothing. Architecture choice matters more to run cost than volume does.

One more cost that never appears in vendor decks: errors you do not catch. At a 4 to 8 percent manual error rate, 120 files a year means 5 to 10 files with an error regardless of whether you automate. The argument for the pipeline is not perfection; it is that the errors get cheaper to find.

Build, buy, or hand it to your administrator

For most funds this is the actual decision, and the honest default is: do not build.

Use your administrator or transfer agent if they already run investor onboarding as a service, which most do. For a fund with a handful of closes a year and modest investor counts, this is almost always the better buy than any build.

Buy a platform if you are an administrator, a large manager with repeat closes, or a group running many funds where verified-KYC reuse pays for itself, as it does in the Gen II rollout. Wins at scale and on repeat closes.

Build only if you have an unusual document mix, strict data-residency requirements that rule out the platforms, or volume that makes five-figure annual platform fees look small against the manual baseline. If you build, copy the architecture the evidence supports: a rules engine for policy, staged LLM agents for classification and extraction, deterministic code for anything numeric, and an auditable queue for human review. That is the design behind both the Spectra paper and the private deployments we run under our enterprise AI practice for banks, insurers, and funds, where the whole pipeline can sit inside your own cloud so investor documents never leave your tenancy.

A person with a good checklist beats any integration project for a single small close. If that is you, stop reading vendor decks.

Common mistakes

• Treating model output as a decision. The screening disposition and the AML sign-off stay with a named human. Build the workflow so that is structurally true, not just policy-true.
• One monolithic prompt. The RBC work shows why staged pipelines win: errors are localizable, each stage is independently measurable, and the audit trail maps to specific policy clauses.
• Trusting general benchmarks. A 95 on a general document benchmark tells you almost nothing about your credit file or cap table. Test on your own document mix, and treat finance-table extraction as unsolved.
• Aiming for 100 percent automation. The working target is around 80 percent: machines handle the low-risk, high-volume files, senior compliance staff handle flagged exceptions with context assembled for them.
• Skipping the data protection work early. Where documents are processed, under what retention policy, and with what certifications (SOC 2, ISO 27001, EU hosting for GDPR) decides which architectures are even available to you.

For a broader view of where document AI fits alongside OCR and extraction tooling, see our guide to AI agents for document processing automation. And if your interest is the other heavy fund-operations workload, the same build-versus-buy logic applies to AI reconciliation for fund administration.

Frequently asked questions

Can AI make the AML sign-off decision?

No. The sign-off, the screening-hit disposition, and the source-of-funds judgment are a named compliance officer's regulatory responsibility. AI can assemble the file, cluster the false positives, and draft the rationale, but the decision and the accountability stay human.

How accurate is LLM extraction from KYC documents?

On the best public evidence, classification of KYC documents reached 100 percent in RBC's evaluation, while field extraction reached 89.4 percent. On complex financial tables the best frontier models sit around 80 to 82 percent and fall below 50 percent on large multi-entity spreadsheets, so anything numeric should be checked by deterministic code.

What does KYC automation cost for a fund?

An onboarding platform typically runs to five figures a year at fund scale. A custom build runs five to six figures including compliance and data protection work, with published pilots starting around $10,000. Running costs depend on architecture: LLM-heavy workflows cost $300 or more per month at scale, deterministic checks almost nothing.

Should a small fund build its own KYC pipeline?

Almost never. If you run a few closes a year with modest investor counts, your administrator's onboarding service or a platform subscription will beat any build on cost and regulator familiarity. Building only makes sense with unusual volume, document mix, or data-residency constraints.

Does sending investor documents to an LLM create data protection problems?

It can, which is why the deployment model matters. Options include enterprise API tiers with zero data retention, EU-hosted platforms for GDPR exposure, and fully private deployment inside your own cloud. Anthropic, for example, holds SOC 2 Type II certification, which is the kind of evidence to request from any vendor in this chain. Vendor risk is now political as well as technical: the US government designating Anthropic a supply chain risk is the kind of event your procurement and compliance teams should have a written answer for, whichever model provider you standardise on.

Where to go from here

The short version: automate the reading, sorting, and chasing; keep the judging human; and buy before you build unless your volumes or data constraints say otherwise. If you want help scoping that for a fund or administrator, our Claude for financial services practice designs exactly this kind of pipeline, from document intake to the compliance gate, with the audit trail examiners expect.