Read invoices into the accounting system
AI reads incoming invoices, checks them against purchase orders, and records them in the accounting system. Invoices whose amounts differ wait for approval.
- Runs on your own PCs and servers
- Use Azure OpenAI, Amazon Bedrock, Google Vertex AI, or an in-house model
- Its engine is Descarty's open-source Dagu (GPL-3.0)
20 supplier invoices read in 72 seconds and checked against their purchase orders; the 3 that differ, on quantity, unit price, and a freight charge, came with the invoice's own words. Accounting approved, and all 20 were entered through the accounting app's screens in 4 minutes. Kitewell built the workflow 23 minutes after the request.
A real run on October 29, 2026, sped up. The system it works with was built for the demo.
The problem today
Every month, staff open hundreds of invoices and type the supplier, amount, and due date into the accounting system. Every supplier uses a different layout, so checking for typing errors takes time too.
The workflow in Kitewell
- Add invoices from the inbox folder to a batch sheet and extract their textAutomatic
- AI reads the supplier, amount, and due dateAI judgment
- Check the amount against purchase order dataAutomatic
- Accounting reviews and approves invoices with differencesApproval
- Record the invoice in the accounting systemWeb
What AI does
Reads supplier, amount, and due date from invoices in any layout. For each value, you can see the part of the invoice it came from.
Where people decide
Staff check only invoices that do not match the purchase order and fields read with low confidence, then approve the entry.
Done condition
Every invoice received is read, the count read matches the count received, and each invoice is either entered in the accounting system with a total that matches its purchase order or listed as an exception.
Exceptions it lists
- An invoice with no matching purchase order
- An amount that differs on quantity, unit price, or a charge
- A value read with low confidence
- An entry the accounting system did not accept
Prove Kitewell on your own systems
Try it free for 14 days on your own PC. To see it working first, ask for a 30-minute demo. There is also a one-month onboarding (¥500,000 before tax), where we build your first job together.
What to check when you try this work
- Whether the needed fields are read correctly from each supplier's layout
- Whether the matching rules against purchase orders fit your process
- Whether to record entries through the screen or an API
How the onboarding runs
- Week 1Requirements and success criteria
- Weeks 2 to 3Build
- Week 4Verify on production screens
- Final dayReport
One-month onboarding
- Length
- One month
- Cost
- ¥500,000 (tax excluded). Your staff build, with the developer beside them
- If you go on to a license
- The onboarding fee is credited to the first year's contract
- Without the onboarding
- At Team scale, start on Team monthly (¥14,800 per computer a month, tax excluded), and cancel in any month
You can try it first on your own PC, in the free 14-day trial.
Frequently asked questions
- Do our documents and screens leave the company?
- Kitewell runs on your own PCs and servers, and Descarty never receives your business data. You choose the AI model: a cloud API, or an in-house or local model with an OpenAI-compatible API. With an in-house model, no data leaves the company. With a cloud model, what the AI reads is sent to the model's provider.
- What if the AI misreads something?
- A person checks what the AI produced before the work goes on. In a sheet, every value the AI reads shows the quote it came from, and rows below your confidence threshold go to a person. Reviewers can send work back with notes on what to fix, and who approved what is kept in the run history.
- What happens when a screen changes or a run stops?
- When a screen changes, the AI reads it again, relearns the step, and carries on. The relearned step is saved, and later runs repeat it. If a run still stops, ask the assistant right there, and it finds the cause and proposes a fix. A proposal is not saved until a person applies it.
- What happens to the RPA and systems we use today?
- Keep them. Kitewell operates screens the way a person does, so it works with systems that have no API. Move the robots that keep stopping and the work you gave up automating because it needs judgment, and leave fast, fixed processing in your current RPA.
- What does it cost?
- ¥14,800 per computer per month, or ¥148,000 per computer when paid yearly, before tax. A computer costs the same whether it builds or runs. You can try it free for 14 days. The one-month onboarding, where we build your first job together, is ¥500,000 before tax, credited to the first year if you go on.
Related use cases
Reconcile rent payments
Match internet banking deposits against the tenant register. When the payer name differs from the tenant's, AI suggests the matching lease.
Create invoices and send them by email
Invoices are filled into your template from an Excel billing list and emailed once a person approves. AI first points out amounts far from last month's.
Take in emailed invoices and draft the entries
Find unread email with an invoice attached, have AI read the amount, and record a draft ledger entry. Uncertain entries wait for accounting to confirm.
See this work in a demo
A 30-minute demo shows this workflow running. The 14-day free trial lets you check it on your own systems.