AI Reconciliation for Fund Administration: Costs and Limits

Reconciliation consumes about 40 percent of a fund accountant's working time, and most of that time goes to a small residue of breaks that rules engines cannot close. The right fix is not an LLM matching every line. It is a deterministic matching engine for the bulk, plus a language model that reads, classifies, and drafts resolutions for the breaks a human would otherwise chase. This post lays out how that architecture works, what it costs per break at current API prices, what the research says about LLM entity matching, and when you should skip AI entirely.

What reconciliation actually costs a fund team

The numbers here are bigger than most operations leads expect.

• Share of accounting time. Operational data from CIBC Mellon, published by ActiveOps, shows reconciliation consumed roughly 40 percent of total fund accountant time, more than 10,000 hours per month. Cash, asset, and capital stock reconciliations alone were about 75 percent of reconciliation activity and 30 percent of total daily fund accounting effort.
• Oversight on top. A SEI and Cutter Associates survey of asset managers running at least $2 billion found 55 percent maintain an internal accounting book of record alongside their administrator's, and 43 percent say more than half of non-investment staff spend time on replication or oversight. For 47 percent of firms, that replication delays valuation and reporting cycles by three or more business days.
• Regulatory exposure. In July 2024 the Federal Reserve fined Citigroup $60.6 million for insufficient progress on remediating data quality management, part of roughly $135.6 million in combined penalties with the OCC. Breaks that linger are not just an efficiency problem. They are the kind of data quality failure regulators now price directly.

The pattern across these sources is consistent: most lines match cleanly, and the cost concentrates in the small percentage that do not.

What an AI reconciliation agent is, and is not

An AI reconciliation agent is not a language model comparing two million-row files. That would be slow, expensive, and wrong, because LLMs are poor at arithmetic and exact comparison at volume.

The working architecture is a hybrid:

• Deterministic engine first. Exact keys, tolerances, date windows, and rule-based matching close 95 to 99 percent of lines. This part is boring, fast, auditable code.
• LLM for the residual. The remaining breaks go to the model, which does what a good analyst does: reads the break, pulls the security master record and prior resolutions, classifies the cause, and drafts a proposed resolution with evidence attached.
• Humans keep authority. Anything above a defined risk threshold goes to a reviewer. Low-risk break classes, once their playbooks are proven, can auto-resolve.

This matches the design principle in Anthropic's own guidance on building effective agents: find the simplest solution that works, use predefined workflows for well-defined tasks, and reach for agentic behavior only where flexibility and model-driven judgment are genuinely needed.

The pipeline, step by step

A production reconciliation agent runs as a fixed sequence of stages, with the model appearing only where judgment is required.

1. Ingest and normalize. Custodian position files, bank statements, and the accounting book of record land in one schema. Most reconciliation pain starts here: one custodian sends SWIFT, another sends CSV over SFTP, a third sends a PDF.
2. Deterministic first pass. Match on security identifier, currency, quantity, and value date, applying the fund's tolerance rules. Every matched pair is logged with the rule that closed it.
3. Break queue. Unmatched and out-of-tolerance items enter a queue, enriched with context: the security master row, the prior day's status for the same account, and how similar past breaks were resolved.
4. Cheap classification. A small, fast model sorts each break into known classes: timing difference, corporate action pending, price source divergence, missing trade, mapping error, or genuinely new. Classification is high volume and low stakes, so it belongs on the cheapest capable model.
5. Deep triage on hard breaks. Classes that need investigation go to a stronger model. It retrieves the relevant records, forms a hypothesis, and drafts a resolution: what to adjust, which system, and the evidence chain, with links back to the source records.
6. Risk-tiered review. Breaks below a value and risk threshold can auto-resolve if their class has an approved playbook. Everything else lands in a reviewer's queue with the draft attached, which turns a 20-minute investigation into a two-minute approval.
7. Learning loop. Every accepted resolution is a candidate new matching rule. Over months the deterministic engine absorbs what the model taught it, and break volumes fall.
8. Audit trail. Every match, classification, and resolution is logged with inputs, model version, and reviewer. In a regulated environment this logging is not optional, and it needs the same governance discipline we set out in our working guide to LLM model risk documentation for banks.

What the research says about LLM matching

The part of reconciliation where LLMs genuinely earn their keep is entity matching: deciding that "XYZ Corp Global Bond 4.25% 2032" in the custodian file and "XYZCP-GB 4.25 32" in the accounting system are the same instrument. The academic evidence is encouraging and specific.

A peer-reviewed study published at EDBT 2025 benchmarked LLMs against fine-tuned encoder models like BERT and RoBERTa on entity matching:

• The best LLMs, with zero or few examples, matched the performance of encoder models fine-tuned on thousands of training pairs. That matters in fund operations, where nobody has a labeled dataset of their matching decisions.
• LLM matchers were more robust to entities they had never seen, which is the normal state of a security master that changes daily.
• GPT-4 could produce structured explanations of its match decisions and identify error classes in its own output, which is exactly what a break-triage workflow needs.
• The honest caveat: no single prompt worked best everywhere. Prompts must be tuned per model and per dataset, so plan an evaluation phase instead of assuming a prompt from a blog post will transfer.

The practical reading: use the LLM where names, descriptions, and context diverge, and keep identifiers, quantities, and dates in code.

A worked cost example

These figures are a model, not a quote. Assumptions first, then the arithmetic.

Assumptions

• Team of 10 fund accountants at $80,000 per year fully loaded, so $800,000 of annual cost.
• Reconciliation takes 40 percent of their time, per the CIBC Mellon data: the equivalent of 4 full-time staff, roughly $320,000 per year of labor.
• Volume of 30,000 cash and position lines per day, with a deterministic engine closing 97 percent: about 900 breaks per day.
• Each LLM triage call sends about 3,000 input tokens (the break, security master context, prior resolutions, instructions) and returns about 250 output tokens.

API cost, at the current Claude API prices

• Triage on Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens: $5.40 of input and $2.25 of output per day, $7.65 in total, about $168 per month across 22 business days.
• First-pass classification on Claude Haiku 5.5 at $0.10 per million input tokens: under $10 per month for the same volume.
• At ten times the volume, 9,000 breaks a day, the Sonnet line is still only about $1,700 per month.
• One analyst-month costs roughly $6,700. The entire model line item is about 2.5 percent of one analyst-month. Prompt caching cuts input costs further: static instructions and the security master sit in cache, and cache reads price at a fraction of standard input.

The costs that actually matter

• Integration engineering: connecting custodians, the accounting platform, and the break queue.
• Data access and entitlements, which take longer than any model work.
• Reviewer time on the risk-tiered queue.
• Validation and ongoing monitoring, including re-testing when custodians change file formats.

The conclusion buyers should draw is not that AI reconciliation is cheap. It is that the model is the cheap part, so vendor pricing that scales with line volume deserves scrutiny, and the budget conversation should be about integration and review capacity.

Build, buy, or augment

You have three realistic options, and they are not mutually exclusive.

Buy a reconciliation platform for standard volume. For custody and cash reconciliation at scale, platforms like Duco and SmartStream are proven. CIBC Mellon ran a measurement phase first, then implemented Duco, and reported more than 1,400 fewer hours of manual reconciliation effort, a 15 percent reduction in reconciliation-related workload, and a 30 percent reduction in daily reconciliation effort per person, per the ActiveOps account. An SS&C Advent case study describes a $1 billion manager with about 850 accounts where one person spent more than a week on manual month-end reconciliation from paper statements; after straight-through automation every account reconciled daily in about three hours.

Augment with a custom agent for the unstructured residue. Reconciliation platforms handle structured feeds well. They struggle with the PDF dividend advice, the broker email disputing a corporate action, the private markets capital call notice. That unstructured 20 percent is where a Claude agent with retrieval earns its keep: reading documents, matching entities, and drafting the resolution a human approves. Anthropic's Claude for Financial Services announcement lists pre-built MCP connectors to data providers including FactSet, S&P Global, Morningstar, PitchBook, Snowflake, and Databricks, and its implementation partner ecosystem includes firms doing entity resolution work specifically. We broke down what the ten finance agent templates do and cost in this analysis.

Build only if you have unusual constraints. Funds with exotic instruments, unusual data residency requirements, or existing engineering teams sometimes build the whole pipeline. That is a multi-quarter commitment, and the build-versus-buy math rarely favors it unless a platform genuinely cannot reach your data.

One caution on vendor numbers: AIG reports that generative AI compressed its underwriting review timelines more than fivefold and raised data accuracy from 75 percent to over 90 percent, in the same Anthropic announcement. That is document review, not reconciliation, and it is the customer's own claim. Treat it as directional evidence that document-heavy finance workflows respond well to this architecture, not as a reconciliation benchmark.

When AI reconciliation is the wrong choice

Skip it, or fix something else first, when any of these hold.

• Low volume. A few hundred lines a month reconciles fine in Excel with a disciplined checklist. The integration cost will never pay back.
• No defined tolerances. If the fund has no written matching policy, no tolerance thresholds, and no break taxonomy, an AI project will stall. Write the policy first; it is a spreadsheet exercise, not a software project.
• Broken upstream data. If the custodian feed itself is unreliable, the model inherits the mess. Fix the feed or change the custodian relationship before automating anything downstream.
• Expecting the model to do arithmetic. Final quantities, cash movements, and postings belong in deterministic code with sign-off, never in a probabilistic system.
• Unacceptable data handling terms. Check your agreements, but note that Anthropic's API does not use customer data to train models by default, per its financial services announcement. For funds with stricter residency requirements, a private deployment of open-weight models is the fallback, at a real cost in capability and operations. Vendor due diligence now includes political risk as well: when the US administration moved to declare Anthropic a supply chain risk, buyers with Anthropic-dependent workflows had to reassess concentration, a scenario we walked through in this analysis.

Implementation checklist

In rough order:

• Measure first. Record hours by reconciliation type for at least a month. CIBC Mellon measured before it bought anything, and the measurement is what made the business case and the later results credible.
• Catalogue your breaks. List the classes, their frequency, their average value at risk, and the current resolution playbook for each.
• Start with cash reconciliation. It is the highest volume, the most rule-friendly, and the easiest to validate.
• Run shadow mode. For the first month the agent classifies and drafts, but humans resolve everything. Compare the drafts against what reviewers actually did.
• Set risk tiers in writing. Which break classes and value thresholds may auto-resolve, and who signed that off.
• Wire audit logging on day one. Inputs, outputs, model version, reviewer, timestamps. Our enterprise AI work for banks and insurers treats this governance documentation as a deliverable in its own right, because examiners ask for exactly this.
• Evaluate weekly. Classification precision, draft acceptance rate, and breaks per 1,000 lines, tracked by custodian so you can see feed problems early.

Frequently asked questions

Can an LLM do the matching itself? Not reliably, and you should not want it to. Exact comparison at volume belongs in deterministic code that is fast, cheap, and auditable. The research shows LLMs excel at entity matching across inconsistent names and descriptions, which is the residual problem after deterministic rules have closed the easy 95-plus percent.

What does AI reconciliation cost per month? The model line is small: roughly $170 per month to triage 900 breaks a day on Claude Sonnet 5.5 at current API prices, and single digits for first-pass classification on Haiku 5.5. The real budget is integration engineering, data access, and reviewer time. Any vendor whose pricing scales steeply with your line volume is charging for the cheap part.

Is our fund data used to train the model? No. Anthropic does not use customer data submitted through its API to train generative models by default. Funds with contractual or regulatory residency requirements beyond that should look at private deployment of open-weight models and accept the capability and operations trade-off.

How long does an implementation take? Plan in quarters, not weeks. A focused cash reconciliation pilot with one custodian can reach shadow mode in four to eight weeks if the data plumbing exists. A multi-custodian rollout with auto-resolution tiers and validated controls is typically a two-to-three-quarter program.

What accuracy should we expect? Measure against your own break classes, not a vendor benchmark. The published research shows LLM matchers are competitive with models fine-tuned on thousands of examples and more robust to unseen entities, but prompts need tuning per dataset, which is why the evaluation phase is not optional.

Where to go from here

Reconciliation is a good first AI workload in fund operations because the inputs are bounded, the outputs are checkable, and the baseline cost is measurable. Start by measuring your own hours by reconciliation type, then pilot on cash reconciliation in shadow mode. If you want help scoping or building that pipeline, our Claude for Financial Services practice implements finance agents for fund administration, research, and investment teams, from the data connectors through to the audit trail.