AI reconciliation: the files go in, only the breaks come back
By Deepak Sheoran, Founder and CTO, DwellFi
AI reconciliation reads your source files directly, infers how they relate to each other without templates or mapping rules, runs the match, and returns only the exceptions, each one scored by risk. Your reviewer looks at the forty lines that disagree instead of the forty thousand that already tie. The matching itself finishes in minutes.
Key takeaways
- Reconciliation is high in volume, repetitive, and unforgiving. That is the exact profile an agent handles well and a skilled person is wasted on.
- You drop the files in a folder. The system maps the relationships across sources itself, no templates to build first.
- Only the breaks come back, ranked by risk, so the reviewer starts where the money is.
- Accuracy is the price of entry. The return that changes your week is the time you get back, every cycle.
Why is reconciliation such a heavy task?
Nothing about reconciliation is intellectually hard. That is what makes it so expensive.
Pull the statements from a dozen accounts. Match every transaction against the general ledger. Tie out the broker positions. Build the exception report, then chase the handful of variances that refuse to balance, which is the part that always takes longer than the plan allowed. Cash gets reconciled daily. Positions and capital accounts run on their own cadence, monthly or at each NAV strike. Either way the work eats the better part of a day for most teams, and it does not require a qualified accountant's judgment. It requires a qualified accountant's afternoon.
It is also unforgiving in a way that most back-office work is not. One transposed digit and nothing balances, and the failure gives you no clue where it started. So you go back through the sheet looking for a mistake you already know is there somewhere, which is a genuinely miserable way for a skilled person to spend a Wednesday.
That profile, enormous and rules-bound and merciless about detail, is precisely what an agent is good at and a person is wasted on.
How does AI reconciliation work, step by step?
You drop the files into a folder. That is the interface.
The system reads each one at source, in whatever shape it arrived: the custodian's PDF, the bank's export, the ledger extract, the broker's position file. Then it works out the relationships across those sources on its own, without you building templates or writing mapping rules first. This bank line corresponds to this ledger entry, which corresponds to this position. Nobody configures that. The system infers it from the documents themselves, which is the part that used to require a six-week implementation project before anything reconciled at all.
Then it runs the match. What ties, ties, and disappears from your day. What does not tie gets isolated and scored, and comes back to you as a short list of breaks with the two conflicting values side by side and a link to the exact page each one came from. Every match it made and every exception it raised is written into the audit trail as the work happens, so the derivation exists before your auditor asks for it rather than being reconstructed afterward.
The reviewer's job changes shape at that point. They are not re-deriving the reconciliation to satisfy themselves it ran correctly. They are working the exceptions, which is the only part that ever needed them.
What does "risk-scored exceptions" mean?
A flat list of breaks is only half a gift, because it hands you back the ranking problem you were trying to escape.
Consider two exceptions. One is a rounding difference of a cent on a fee accrual. The other is an unexplained variance on a capital movement that shows up in the custodian file and not in the ledger. Both are technically breaks. Only one of them can ruin your quarter. Risk scoring ranks them so the reviewer opens the second one first, rather than working a flat list from the top and reaching the thing that mattered at four in the afternoon.
That ordering is doing more work than it appears to. It is the difference between an exception report you triage and an exception report you simply work.
So why isn't accuracy the headline?
Because a reconciliation that does not tie within tolerance is worthless, which means accuracy is the price of the ticket rather than the show.
Every credible vendor in this category clears that bar, and the ones that do not are not really in the category. So when accuracy leads the pitch, it is telling you something about the vendor's marketing rather than about the product's value. The return that actually changes your operation is the day you get back, every cycle, forever.
Here is the honest concession. A capable team with good spreadsheets and enough hours will reconcile accurately too. They have been doing it for decades, and they are not doing it wrong. The structural gap is not that people match badly. It is that people match slowly, and slow matching quietly sets a ceiling on how often you are allowed to look.
Curious how this holds up on your own break patterns? Reconcile a real account set in minutes.
Can reconciliation move from monthly to continuous?
This is the part that gets missed, and it is the reason the time argument matters more than it sounds.
Your reconciliation cadence was never really an accounting decision. It was a budget decision wearing an accounting costume. You reconcile positions monthly because a day of a skilled person's time is what the process costs, and a day a month is what you could afford. Nobody sat down and concluded that thirty days was the correct interval at which to discover that a wire went to the wrong account. Thirty days is just what the labor bill would bear.
Change the cost and the cadence stops being a constraint. When reconciliation takes minutes of machine time and a short review, you can run it weekly, then daily, then continuously, without hiring anyone. And running it continuously is a different activity from running it monthly, even though it has the same name. You catch a break while the transaction is fresh, the counterparty still remembers it, and the reference data behind it has not yet been overwritten. You find it in hours instead of in thirty days, when the trail has gone cold and the reconstruction costs more than the error did.
Reconciliation stops being an event on the calendar and becomes a property of the system. That is the shift. Everything else is throughput.
Where DwellFi fits
DwellFi runs reconciliation in two layers, and the distinction is the whole point. First, the agent infers how your sources map to each other, so you skip the mapping project. Then the match itself is computed by an explicit, deterministic rule against validated inputs, not by a clever guess, so the same inputs produce the same output this quarter and next.
The records underneath are event-sourced and append-only, which is what lets every matched line and every exception trace back to its own source page with its own timestamp. Where a break genuinely needs judgment, a named reviewer signs off, so accountability for the resolved figure sits with the person who signed it.
It runs inside your environment, on your data, which for most regulated administrators is the part that decides whether any of this is adoptable at all.
Get this right and reconciliation stops being the thing that eats the first week of every month. It becomes the verified base that everything else can be built on, from fraud flags on vendor payments to tax reporting rendered per domicile, none of which is safe to automate on top of numbers you cannot defend. That is the real prize. The day you get back is just the first installment.
Frequently asked questions about AI reconciliation
Have more questions about AI reconciliation? Don’t worry, we got them answered:
How long does AI reconciliation take?
The matching itself runs in minutes rather than hours. Human time goes to the flagged exceptions instead of the full ledger, so the reviewer's effort scales with the number of breaks rather than the number of transactions. Confirm specific timings against your own data volume, since file count and source quality both move the number.
Does it handle multiple currencies and multiple sources?
Yes. Multi-currency and multi-source reconciliation across bank, ledger, and broker positions is a core case rather than an edge one, and the relationships across those sources are inferred rather than configured.
Is the output audit-ready?
Every match and every exception carries a trail back to its source document and page, so the reconciliation is defensible when an auditor or an examiner asks how a figure was reached. The trail is generated as the work happens rather than assembled later.
What happens when the system cannot map something?
It stops and flags it rather than guessing. An agent that invents a match to keep the run clean is worse than no agent, so anything it cannot map cleanly becomes an exception for a person to resolve.
How is this different from the rules-based reconciliation software we already run?
Traditional tools need you to build and maintain the mapping rules before anything matches, which is where the six-week implementations and the ongoing config debt come from. AI reconciliation infers the mapping from the source files themselves, then runs a deterministic match on top. You keep the repeatability of rules without paying to write and babysit them.
Go deeper: start with AI agents in fund administration, then AI tables for structured financial data, why the audit trail is the real product, and the DwellFi platform.
Reconcile a real account set in minutes. Book a demo.