--- title: "Reconciliation hell: why your systems don't agree with each other" description: "To reconcile two systems, name one source of truth per record, check every copy against it automatically, and send only the mismatches to a person." url: "https://automancer.uk/field-notes/reconciliation-hell-why-your-systems-dont-agree/" canonical: "https://automancer.uk/field-notes/reconciliation-hell-why-your-systems-dont-agree/" type: "BlogPosting" author: "Waseem Ilyas" datePublished: "2026-09-28T00:00:00.000Z" --- Your rota says a support worker did four shifts last week. Payroll says three. Both systems are working perfectly, which is exactly the problem. ## Why nobody notices until the meeting Systems that are meant to mirror each other drift apart quietly. Nothing crashes and nothing sends an alert. Invoices still go out, orders still ship, the rota still gets filled. The gap only shows up when two people pull the same figure from two screens in the same meeting and get two different answers. That is why it is one of the most common hidden costs in a small business: it never appears as a line in any budget. It appears as someone's Thursday afternoon, spent with two browser tabs open, working out which number is real. Then it happens again next month, because nothing about the cause has changed. The usual suspects: - **Rota and payroll.** A shift swapped on the rota and never updated on the timesheet. - **Stock and orders.** A product code edited in one system and not the other, so the portal happily sells something the warehouse no longer has. - **CRM and finance.** A customer marked active in one and archived in the other, and a revenue-by-customer report nobody quite trusts. None of these is a software bug. Each system is faithfully reporting what it was told. Nobody told them how to agree, and nobody built anything to notice when they don't. ## How to reconcile data between two systems To reconcile data between systems, name one system as the source of truth for each type of record, treat every other copy as a mirror, compare the mirror against the source automatically on a schedule, and send only the mismatches to a person to resolve. That is the whole framework, in four parts: 1. **A source of truth, per record type.** Not one master system for the whole business: one per entity. The rota might own who was scheduled, the clock-in system who actually turned up, the finance package what a customer owes. If you cannot say which system wins an argument about a given record, you don't have a data problem yet. You have a decision nobody has made. 2. **Mirrors follow, they don't argue.** Every other copy of that record is a mirror. It takes updates from the source. When staff edit the mirror directly because it happens to be the screen they have open, you now have two masters, and drift is guaranteed. 3. **A reconciliation check, not an audit.** A scheduled comparison of mirror against source that outputs only the differences. Records that match are nobody's business. Records that don't come out as a short, specific list: missing on one side, different value, different date. 4. **An exception queue with an owner.** That list has to land somewhere a named person works through it, fixes the source rather than the mirror, and clears it. A check whose output goes to an unread inbox is just a more expensive way of not knowing. When two systems stay in step for months without anyone touching them, it looks like a spell someone cast once and forgot about. It isn't magic. It's very good engineering, and most of it is deciding in advance what happens when the numbers disagree, instead of hoping they won't. ## What this looked like on a live build On the core platform we built for [a ~600-person supported-living care provider](/work/care-provider-transformation/), the new system's report mirrors were reconciled against the client's real exports before anyone was asked to trust a number coming out of it. That check covered 500 visits, 175 medication alerts and 416 names routed to review. Care makes the stakes easy to see. When the records are visits and medication alerts, "roughly right" is not good enough. Every workstream on that engagement was reconciled against the client's real exports and board-gated before it went anywhere near live service-user data. The core platform is live as a production pilot the client is actively using. The same discipline applies whether the mismatch is a medication alert or a pallet of stock. If a system's numbers have never been checked against a source you already trust, you are taking them on faith. If your problem is less "two numbers disagree" and more "the same person exists twice under two spellings", that is a close cousin: identity matching. We have written up [what identity matching actually fixes](/field-notes/duplicate-data-what-identity-matching-fixes/) separately. ## What to do yourself, and when to pay someone A lot of this you can start this week with a spreadsheet and some honest conversations. **Do it yourself:** - **List your pairs.** Write down every place two systems hold the same thing: people, shifts, stock, customers, invoices. - **Pick the winner for each.** One line per pair: "for this record, system A is right". Get whoever runs each system to agree to it out loud. - **Agree the matching key.** Decide what links a record in one system to its twin in the other. If the honest answer is "the name, spelled however someone typed it", fix that first. - **Run one manual comparison.** Export both, compare them in a spreadsheet, count the mismatches. That count tells you whether you have a tidy-up or a structural problem. **Worth paying someone when:** - **It's weekly.** The comparison needs doing every week and the export-and-compare routine is eating real hours. - **Matching gets fuzzy.** Typos, nicknames, changed addresses, merged records: the kind of matching a spreadsheet lookup gets confidently wrong. - **You want it to run itself.** A scheduled check, an append-only audit trail of every correction, and an exception queue inside the system rather than in someone's inbox. - **A miss is expensive.** Payroll, invoicing, or anything an inspector will read. ## Where to start If your team spends time every month working out which system is telling the truth, that time is the cost, and it is usually bigger than anyone has written down. An [Automation Opportunity Audit](/services/), from £450, is where we would start: map which systems hold what, find where they disagree, and give you a costed plan for what to fix first. The fee is credited against any project you go on to do. A real person, Waseem, reads what you send and replies personally, and we promise a meeting within one week.