We tried RPA. Now we have forty bots nobody can maintain.
Correct, and that outcome was structural rather than bad luck: record-and-playback binds a bot to a UI that changes without notice, and unattended logins break at the next MFA rollout. We inventory by execution telemetry, not the register — most estates find bots that have not succeeded in months, and 20–40% get decommissioned outright. Anything with a real interface becomes a typed integration with retries, dead-letter queues and idempotency; what has none is rebuilt as versioned code with nightly contract tests. The decommissioning plan is a deliverable from day one, not a later conversation.
Our data is a mess and the ERP has been customised beyond recognition.
That is the normal starting condition, and it is the reason we mine the log before proposing anything. Customisation is visible in the change-document tables; data quality is measurable as the share of cases with complete timestamps and resolvable IDs. Below roughly 80% we tell you the mining result is unsafe and quote the measurement fix instead of a build — which is a smaller invoice and a worse-looking proposal, and still the right call. What we will not do is write to unreleased tables or the database directly: that violates clean core and the next S/4HANA upgrade breaks it silently.
Vendors promise 90%+ accuracy. Our invoices are Bulgarian, scanned crooked, half of them handwritten.
Then most of the market benchmarks do not apply to you, and one vendor is disqualified before the demo: Amazon Textract supports text detection in English, French, German, Italian, Portuguese and Spanish only, with handwriting in English only. No Cyrillic. Azure AI Document Intelligence, Google Document AI and ABBYY are the shortlist, and each carries a caveat we test rather than assume: the prebuilt invoice models and custom-neural training carry their own supported-locale lists, which are narrower than the underlying OCR’s Cyrillic coverage — so for Bulgarian we bench Read/Layout plus a custom model against the prebuilt path on your documents before we commit to either. On handwriting, be sceptical of all three: handwritten Cyrillic line items are not an OCR problem you can buy your way out of. That is a supplier problem, and we would rather move those twenty suppliers to a portal, a template or a Peppol lane than promise you a model that reads them — we will tell you which twenty they are in week two. Header fields — invoice number, date, total, VAT and EIK — commonly run 95–99% on clean digital PDFs; line items run 75–90% and fall further on degraded scans. We contract on per-field F1 measured on your own sample, with a confidence threshold and the auto-acceptance rate it yields. A single document-level accuracy figure means nobody looked at your mail.
Hours saved never show up in the P&L.
Usually true, and it is a measurement design failure rather than a technology one. Six minutes saved across forty people is absorbed into the working day and never reaches the ledger. We count only benefit that consolidates into whole FTE-equivalents inside one team, or that is cash-visible: headcount not backfilled, contractor and BPO spend reduced, early-payment discounts captured, late-payment penalties avoided, DSO or DPO days moved, audit and rework cost removed. That definition is agreed and signed with Finance before build, and the baseline comes from system timestamps rather than a survey of how long people think a task takes.
We are mid-S/4HANA migration. Everything is frozen.
Usually the right call, and the real constraint is not the code freeze — it is that your Basis team, your functional consultants and your change board have no spare hours until cutover. So the honest answer is that most of this waits. What does not have to wait is read-only: extracting the event log costs your team a database grant and about half a day, and it tells the migration programme which of the processes it is about to rebuild are ones nobody actually uses. If you want to build during the freeze, it can only be the layer outside the core — ingestion, validation, master-data screening, exception queues, orchestration — against released APIs and side-by-side on BTP, and we would still want your migration lead to say yes in writing before we start.
Our IT security review will take six months.
It should be scoped as a supply-chain review, because under NIS2 Art. 21 that is what it is — we hold production ERP credentials, so we sit inside your regulated supply chain, with a 24-hour early-warning and 72-hour notification clock on incidents under Art. 23. We arrive with the pack already built: service accounts scoped to named operations rather than a person’s login, secrets in a managed vault with rotation, documented network paths, logging into your SIEM, a signed DPA with a current subprocessor list, and a data-flow map naming every place personal data lands. That turns your review from an investigation into a checklist.
The people who understand the process are the ones this replaces.
They will not cooperate if that is the deal, and they are right not to. At EUR 12.0 an hour headcount removal is a weak case anyway — the return is cycle time, error rate and volume headroom. Practically: the process expert sits on the build team as the person who defines the exception rules, and the first thing we automate is the task they hate, not the one that demos best.
