Automate the work rules-based tools never could

Robotic process automation was sold as automating any repeatable process, and it genuinely worked for structured, stable inputs. Its failure mode was always the same: it required the world to hold still. Templates keyed to coordinates, screen-scraping keyed to element positions, rules enumerated for every variation someone thought of in advance. Every supplier redesign, portal update and new document format broke something, and maintenance eventually consumed the savings.

A vision-language model reads a document approximately the way a person does — it locates the invoice number because it understands what an invoice number is, not because it was told where to look. A layout it has never encountered is not an error condition. That is the whole difference, and it is why workflows that resisted a decade of automation attempts now automate in weeks.

The trade-off is that the model can be wrong in ways a template cannot. A template either finds the field or fails loudly. A model can return a plausible, incorrect value with no signal of trouble. Everything else in this page is about engineering that signal back in.

For clients with existing RPA estates, the right move is usually augmentation rather than replacement. The bots keep doing what they are good at — moving structured data between systems reliably — and an AI extraction layer in front of them supplies clean, structured input. This preserves existing investment and delivers value in weeks rather than a rebuild.

An extraction system that reports 97% accuracy and cannot tell you which 3% are wrong is not usable in finance, insurance or compliance. The value comes from knowing which outputs to trust.

We generate confidence at the field level using several independent signals rather than the model’s own self-assessment alone, which is unreliable on its own. Token-level probability from the extraction pass, agreement between two different model passes on the same field, validation against the document’s own arithmetic (do the line items sum to the stated total?), format and range checks, and cross-reference against master data — an unrecognised supplier tax ID is low confidence regardless of how cleanly it was read.

Those signals combine into a per-field confidence score, and thresholds decide the route. Above the threshold on every field: straight through. Below on any field: to the review queue with the uncertain fields highlighted. Thresholds are set per field by consequence, not uniformly — a misread invoice date costs a phone call, a misread bank account number costs real money, and they should not share a threshold.

The thresholds are tuned during parallel run, where we can measure precisely what STP rate each setting produces and what error rate slips through. That curve is a business decision, not a technical one, and we present it as such.

Every intelligent automation deployment has a human queue, and its design determines whether the whole system feels like a gain or a burden. We have seen technically excellent extraction pipelines fail adoption entirely because reviewing an exception took longer than processing the document manually would have.

The rules we build to: the document image and the extracted values sit side by side, with the source region for each field highlighted on hover so the reviewer verifies visually rather than hunting. Only uncertain fields are surfaced for attention; confident ones are shown but do not demand it. Correction is keyboard-first with tab-through, because the people doing this work are fast typists and a mouse-driven interface halves their throughput. Approval is one keystroke.

Queues are prioritised by business urgency rather than arrival order — an invoice approaching its discount deadline outranks one that arrived first. And every correction is captured as labelled data, feeding the evaluation set so the same class of error can be measured and reduced.

A well-designed queue turns reviewers into exception handlers rather than data-entry staff, which is both a better job and a much cheaper one per unit of work. Teams that would have resisted automation generally become its strongest advocates once they stop typing.

We do not ask clients to trust a benchmark. Before cutover, the pipeline processes real live volume alongside the existing manual process for two to four weeks, and every disagreement between the two is adjudicated by a subject-matter expert.

This produces three things. First, the genuine accuracy figure on your documents, including the scanned-on-a-phone tail that never appears in a vendor demo. Second, a categorised list of failure modes, each of which becomes either a fix or a known limitation with a routing rule. Third — reliably, in every parallel run we have conducted — a measurement of the manual baseline’s own error rate, which is typically between 1% and 4% and which nobody had ever quantified.

That third finding changes the conversation. The question stops being "is the AI perfect?" and becomes "is it better than what we do now, and are its errors more detectable?" On both counts a confidence-scored pipeline generally wins clearly, and the decision to cut over becomes straightforward rather than an act of faith.

Cost per document processed by a well-architected pipeline is typically a few rupees, against a fully loaded manual cost measured in tens or hundreds depending on complexity and geography. The payback calculation is usually straightforward, but it is worth being precise about the components.

Inference cost per document depends heavily on page count and whether images require a vision model. Routing matters: sending every document to a frontier model is the most common way to destroy the economics, when the majority are clean and handled perfectly by a smaller one. Caching helps where documents repeat. And the human review tail must be included — at an 88% STP rate you are still paying for review on 12% of volume, and that cost belongs in the model.

The volume threshold where this pays back within a quarter sits at roughly 500 documents per month for moderately complex work, lower where the documents are complex or the manual cost is high, higher for simple ones. Below that, the integration and change-management effort usually outweighs the saving, and we will tell you so rather than build it.

The benefits that are harder to quantify tend to be the ones clients value most after a year: month-end that no longer requires overtime, volume spikes absorbed without temporary staff, and complete rather than sampled compliance coverage.