Loading
Loading
Product concept
Documents that read themselves, and a human who only sees the doubtful ones.
Context
The concept assumes a fictional bookkeeping firm of fifteen people processing a few thousand documents a month for small businesses. No data team, no appetite for experiments that put a tax filing at risk. So Helix has to be explainable from day one.
Challenge
AI in bookkeeping is only worth something when you know exactly when not to trust it. A model that reads 95 percent of invoices correctly and quietly misposts the rest costs more time than it saves. The task was never recognition; it was making doubt visible.
Approach
work.projects.helix.solution
Research
We counted where the time actually goes, and it was not retyping but searching: which document belongs to which client, which VAT rule applies, who has to approve this. That is why Helix classifies and routes first, and only then fills in the fields.
User flow
The essential path through the product, step by step.
Mail, scanner and portal converge into one queue, whatever the format.
Amounts, VAT, counterparty and date are recognised, each with its own confidence score.
Above the threshold Helix posts on its own; below it, the document goes to review.
The reviewer sees the proposal beside the source and corrects with one click.
Every correction becomes a rule, so the same case passes cleanly next time.
Design direction
Graphite, hairlines and monospaced figures: the screen should read as an operations console, not a chat window. Mint means the system handled it, amber means a human is needed. There are no other colours, because this is the only decision the user makes all day.
Interface preview of Helix with illustrative demo data, no client information.
Architecture
The layers that make it dependable, from client to data.
Mail, scans and uploads normalised into a single document model.
Field-level recognition carrying a confidence score and the source location.
Thresholds, routing and approvals, editable without a developer.
Posted documents and corrections pushed back to the accounting package, with an audit trail.
Responsive
Processing happens on a wide screen, so the queue keeps its density there. On a phone what remains is the part that matters in transit: approving open reviews and seeing whether the flow is stuck.
Interface preview of Helix with illustrative demo data, no client information.
Outcome
The concept shows how we put AI into an existing process: not as a bolt-on assistant, but as a layer that measures its own doubt and hands that doubt to a person. The extraction layer and the rules engine are our reference for administrative automation.
Stack
06
Next case study
Internal showcasePlant software that states the facts and gets out of the way.