Everyone adopted AI. Almost nobody pointed it at the close.
Reconciliation is where finance still does the reading by hand. ReconcileAI reads every document, pairs it, judges what does not agree, and proposes the resolution — then stops, and waits for a person to decide.
Runs in your own AWS account if you need it to. Complementary to SAP and non-SAP systems — it does not replace your ERP, it reconciles across what you already run.
Three documents, one of which does not agree with the other two. Everything the product does starts here. Illustrative — no customer document is shown anywhere on this site.
None of this is abstract. It is a person opening a spreadsheet.
Every exception in the pile is somebody pulling a statement, calling a supplier, finding the goods receipt that was reversed, and writing down what they concluded so the next person does not have to work it out again. The work is not hard. There is simply too much of it, and it all lands in the same five days.
- ~Days, not minutes. The close waits on exceptions, and exceptions wait on whoever has the context. The tilde is deliberate — that is an estimate, and the product marks every number it is guessing at.
- ›The judgement is the valuable part, and it is the part that gets rushed, because the reading ate the time.
- …And some of it is not yours to finish. A charge nobody can account for is blocked on a bank, a supplier or another team — work that is real, moving, and moving somewhere else.
It reads the document before it matches it.
Any layout, any language. Fields are lifted with a confidence score, and anything below the floor is parked for a person to confirm rather than guessed at. That floor is set per document pack by our delivery team, not left to a default — a pack with no floor refuses to run.
One field is deliberately parked. The bar is a measurement and nothing else — it never means good, passed or safe.
Needs confirmation ›
Tax code read at 0.41, below the 0.60 floor. The document is shown beside the field, and confirming it teaches the pack rather than patching the row.
Ruled out —
Three candidate matches the model considered and rejected, with the reason for each. The ruled-out list is what makes a proposal reviewable instead of a verdict.
Run completed 03:12
Past tense, with a time on it. A figure the system is certain of never wears a hedge, so a number without one is a number you can rely on.
Pick an exception. Watch what it costs.
Three exception types, three different teams, one shape. Something is knowable from the documents, and nobody has had time to go and know it.
The same invoice arrives twice, six weeks apart, under two purchase orders and a supplier name spelled two ways. It clears both times. Somebody finds it in a quarterly review, or nobody does.
Whose move. Accounts payable, if anyone. Nothing in the ledger says the two rows are the same invoice.
Both documents are read, the supplier is resolved across both spellings, and the pair is flagged before the second payment runs — with the two documents side by side and the three near-matches it ruled out.
What the person still does. Confirms it is the same invoice and applies the SOP. The proposal is drafted; the decision is not.
One path, six stages — and a person owns the last one.
Everything in the product serves one of these or governs it. The agent runs every stage; naming it says nothing. A person in the loop is the exception, so that is what gets named.
Ask is a capability before it is a destination. Any question about the period, in words — no SQL, no export request, no ticket — answered from the source systems with a confidence and the cost of the answer. It opens over the register rather than taking you away from it.
Two deployments, and the harder one is the one enterprises ask for.
Your documents are the most sensitive thing in the close. So the question of whose account they sit in is not a deployment detail, it is the first thing a security review asks.
We run it. Fastest to a first reconciliation, and the default for a partner-sold tenant.
Your data sits in our account under your retention and residency settings, and the audit trail is yours.
The product deploys into your account. Your documents, your KMS keys, your Bedrock quota, your bill — and your connector credentials in your own Secrets Manager, which we never see.
The application is served from our CloudFront and your browser calls your own API directly, so tenant data never transits a 1CloudHub account. The code comes from us and executes against your data in your user’s browser — stated plainly here because a security review will ask about exactly that.
- ›Half of onboarding is yours. The key exchange, the model pre-flight, identity federation and the connector credentials all happen inside your account, by your people. We track a state per step so “why are we not live yet” always has an answer.
- ›Your identity provider, your roles. OIDC or SAML, at your own host. There is no ReconcileAI password.
- …Escalating hosted to your own account moves everything — documents, runs, resolutions, approvals and the audit trail. A partial migration is a lost history wearing a better word.
One engagement on the record. Two more being cleared.
Every client here is unnamed at their request, and every figure is one the client agreed to. Where a story is not cleared yet, the slot says so rather than being filled with something weaker.
A 9-step manual payables process, read and matched by an agent.
A premier Philippine real estate conglomerate — 90 buildings, 3.4M sqm, 1,000+ employees. Vendors upload the pack, fields are extracted and three-way matched across purchase order, goods receipt and supplier invoice, and the team reviews only what is flagged.
Extraction held above 85% across 58+ vendor document formats. A two-week proof of concept came first, then production in phases. Engagement period June 2025 — ongoing.
Not written yet.
Reserved for a second engagement. Sector, geography, the exception type it started on, and the figures the client will stand behind. Nothing goes in this slot until they are agreed in writing.
Not written yet.
Reserved for a third engagement, ideally in a different industry and a different deployment mode to the first — so the three together say the product is not shaped around one landscape.
The refusals are the product, as much as the features are.
A reconciliation engine that quietly decides things is worth less than one that stops. These are deliberate, and they are on the record.
- ✕It does not post above an authority limit. The agent proposes; a named person approves, with a required reason. The reason is what an SOP amendment gets drafted from.
- ✕It does not guess a field below the confidence floor. It parks it and asks.
- ✕No colour carries a state. Every status is a word. Nothing in the interface depends on your being able to tell green from amber.
- ✕It does not answer questions it cannot evidence. An answer arrives with a confidence, and a confidence below the floor arrives as a question.
- ✕There is no self-serve sign-up. A tenant carries a deployment mode, a residency, a retention policy and a support-access mode — four decisions with no safe defaults, so a person makes them with you.
Built by 1CloudHub, on AWS.
We built ReconcileAI because we kept meeting the same gap in every finance landscape we were asked to modernise: the systems were current, the reconciliation was still manual, and nobody had pointed the new tooling at the part that actually consumed the week.
Offices in Singapore, India, the United States, Dubai and the Philippines. Delivered by the same engineers who run the onboarding.
Bring us one exception type.
Forty-five minutes on your landscape, your exception types and who needs to see what. The demo runs on a real reconciliation against synthetic documents, not a slide deck.