CodeIQ vs Booke AI: Which AI Bookkeeper Actually Works?
Two very different approaches to the same problem: automating bookkeeping with AI. One clicks through your accounting software like a fast human. The other connects via API with a multi-layer intelligence pipeline. Here’s how they compare.
By Jack Whitehead, AATQB
The AI bookkeeping category
AI-powered transaction coding is a new category that’s emerged in the last few years. The promise is straightforward: instead of manually coding each bank transaction to an account in Xero, QuickBooks, or Sage, an AI system does it for you. You review the work and post it.
But not all AI bookkeeping tools are built the same way. Two fundamentally different architectural approaches have emerged:
RPA (Robotic Process Automation)
The tool controls a browser and clicks through the accounting software’s interface, mimicking what a human would do. It reads the screen, selects categories from dropdowns, and clicks buttons. Booke AI uses this approach.
API-native integration
The tool connects directly to the accounting platform’s API, reads transaction data programmatically, processes it through classification algorithms, and writes coded results back via API. CodeIQ uses this approach.
This architectural difference isn’t just a technical detail. It determines reliability, speed, accuracy, and how well the tool scales across multiple clients.
How Booke AI works
The RPA approach
Booke AI logs into your Xero or QuickBooks account through a browser automation layer. It reads the bank transactions displayed on screen, categorises them using AI, and selects the appropriate account from the platform’s interface. It can also handle basic bank reconciliation within the platform.
The process is conceptually simple: Booke does what a fast bookkeeper would do, but without the human. It sees the same screen, interacts with the same elements, and produces the same result.
Advantages of RPA
- Works with the existing UI – no API access needed
- Can handle tasks that APIs don’t expose
- Visually understandable – you can see what it does
- Learns from user corrections over time
Inherent limitations of RPA
- Breaks when Xero or QBO update their UI
- Limited by page load speeds and rendering
- Requires browser sessions for each client
- Can’t process in bulk – one transaction at a time
- Has full UI access to your platform (security concern)
- Difficult to run multiple clients simultaneously
How CodeIQ works
The API-native approach
CodeIQ connects to your accounting platform via OAuth (the same secure authorisation Dext, Hubdoc, and other tools use). It reads transaction data through the platform’s API, processes it through a dedicated classification pipeline, and writes coded results back via API.
The classification pipeline has seven layers, each designed to catch what earlier layers missed:
- Layer 1 – Transfer detection: Identifies transfers between your own accounts by matching equal and opposite amounts within a time window across bank accounts
- Layer 2 – Invoice matching: Matches bank payments against outstanding invoices from the platform, handling partial payments and adjustments
- Layer 3 – Historical patterns: Learns from the client’s own general ledger history. If “TESCO” has been coded to “Office Supplies” twelve times in this client’s books, it picks that up
- Layer 4 – Universal patterns: Draws from a crowd-sourced database of 3,500+ anonymised merchant-to-account mappings across all CodeIQ users. New practices benefit from patterns learned by existing ones
- Layer 5 – MCC matching: Uses merchant category codes from card transactions as a classification signal, cross-referenced with the universal pattern database
- Layer 6 – Semantic analysis: An embedding model (BGE-base-en-v1.5) understands what a transaction description means, matching it against enriched account descriptions even if the exact words don’t match
- Layer 7 – User learning: Your manual corrections permanently override all future suggestions for the same merchant. The system doesn’t repeat mistakes
After all seven layers, CodeIQ runs VAT classification using a dedicated engine that maps transactions to platform-specific VAT codes (Xero’s INPUT2, NONE, RRINPUT, etc. – not generic labels). Then it posts coded transactions back to the platform.
Head-to-head comparison
Speed
RPA is limited by browser rendering speed – it has to wait for each page to load, each dropdown to populate. API-native processes transactions in bulk. A typical client with 300 transactions takes CodeIQ roughly two minutes end-to-end. RPA tools take longer because they process sequentially through the UI.
Reliability
This is the critical difference. When Xero or QuickBooks update their interface – which happens regularly – RPA tools break until they’re updated to match the new UI elements. API contracts are versioned and stable. Xero’s API v2 doesn’t change its endpoint structure when they redesign a button. CodeIQ hasn’t broken from a platform UI update because it never touches the UI.
Accuracy
Single-pass classification (read description, pick category) typically achieves 70–85% accuracy. CodeIQ’s seven-layer approach means multiple independent attempts to classify each transaction. If the historical pattern layer doesn’t find a match, the universal pattern layer might. If that misses, semantic analysis catches it. Each layer improves the overall accuracy.
Platform support
Booke AI supports Xero and QuickBooks. CodeIQ supports Xero, QuickBooks, Sage, and Pandle. For multi-platform practices, this matters.
VAT handling
CodeIQ has a dedicated VAT classification engine that understands the difference between standard-rated, reduced-rate, exempt, zero-rated, and reverse charge supplies. It maps to the correct platform-specific VAT codes. Booke AI relies on the platform’s own VAT rules, which means the accuracy depends on how well the platform’s built-in rules are configured.
Network effect
CodeIQ’s universal pattern database gets smarter as more practices use it. When a practice in Manchester codes “BRITISH GAS” to “Gas and Electric,” that pattern becomes available (anonymised, with consent) to every other CodeIQ user. Booke AI’s learning is per-client only.
Feature comparison table
| Feature | Booke AI | CodeIQ |
|---|---|---|
| Architecture | RPA (browser automation) | API-native (OAuth + REST) |
| Classification method | Single-pass AI | 7-layer pipeline |
| Transfer detection | No | Yes (cross-account matching) |
| Invoice matching | No | Yes (partials + adjustments) |
| Historical pattern learning | Per-client | Per-client + universal network |
| Universal pattern database | No | 3,500+ patterns |
| Semantic analysis | No | BGE-base embedding model |
| VAT classification | Platform rules | Dedicated engine + platform mapping |
| Platform support | Xero, QuickBooks | Xero, QuickBooks, Sage, Pandle |
| Posting to platform | Yes (via UI) | Yes (via API) |
| Bulk processing | Sequential (UI-limited) | Batch (API bulk) |
| Breaks on UI updates | Yes | No |
| Bank statement upload | No (platform only) | CSV, PDF, Excel, OFX, QBO |
| PDF converter | No | 17+ UK banks supported |
| User learning | Yes | Yes (permanent overrides) |
Pricing comparison
Booke AI’s pricing is per-client, typically starting around $20/month (roughly £16). For a practice with 30 clients, that’s $600/month.
CodeIQ uses a credit-based model starting at £5/month (Starter), with practice plans at £39/month (50,000 credits) and £79/month (150,000 credits). Credits are consumed per transaction processed. A typical client with 300 monthly transactions uses roughly 300–600 credits. A practice with 30 clients would typically fit within the £39 or £79 plan.
The credit model means you pay for what you use. Months with fewer transactions cost less. Per-client pricing charges the same regardless of transaction volume.
Which should you choose?
Choose Booke AI if:
- You only use Xero or QuickBooks (not Sage or Pandle)
- You have a small number of clients (under 10)
- You’re comfortable with the RPA reliability trade-off
- You prefer seeing the automation work through the UI
Choose CodeIQ if:
- You need reliability – can’t afford the tool breaking when Xero updates
- You use multiple platforms (Xero, QBO, Sage, and/or Pandle)
- You want a multi-layer classification pipeline for higher accuracy
- You need dedicated UK VAT classification
- You want network effects – patterns learned across all users
- You process bank statements in PDF or other formats
- You have more than 10 clients and need scalable pricing
See the seven-layer pipeline in action
Upload a bank statement, watch CodeIQ code every transaction, review the results, and post. About two minutes per client.
Try CodeIQ Free →Frequently asked questions
What is the difference between RPA and API-native bookkeeping automation?
RPA controls a browser to click through the accounting software’s UI, mimicking human actions. API-native tools connect directly to the platform’s API, reading and writing data without touching the interface. RPA is quicker to build but fragile – any UI change can break the automation. API-native tools are more reliable because APIs have stable, versioned contracts.
Does Booke AI work with Sage or Pandle?
No. Booke AI supports Xero and QuickBooks Online. CodeIQ supports Xero, QuickBooks, Sage, and Pandle.
How accurate is AI bookkeeping compared to manual coding?
Single-pass AI categorisation typically achieves 70–85% accuracy. Multi-layer pipelines like CodeIQ’s achieve higher accuracy because each layer catches what earlier layers missed. Both approaches require human review before posting.
Can AI handle UK VAT classification correctly?
Generic categorisation tools often struggle with UK VAT. CodeIQ has a dedicated VAT classification engine that maps to platform-specific codes (Xero’s NONE, INPUT2, RRINPUT, etc.). Booke AI relies on the platform’s own VAT rules.
Which is better for a multi-client practice?
CodeIQ is better suited for multi-client practices: four platform support, universal pattern database across all clients, and API-native integration that doesn’t require individual browser sessions. Booke AI’s RPA approach requires per-client sessions, which limits scalability.