Accounted Review Queue: AI-Flagged Items

The review queue is where Accounted's value proposition for accountants crystallises. It is the screen you will spend more time on than any other in the practice portal, and it is designed around a simple principle: you should only see the items that need your professional judgement. Everything else — the correctly categorised, confidently matched, routine transactions — is handled and documented without requiring your attention.

This is not a novel concept in principle. Risk-based auditing has existed for decades. But applying it to day-to-day bookkeeping review, supported by AI confidence scoring, changes the practical economics of managing a client base. Here is exactly how the review queue works, what determines which items appear in it, and how to use it effectively.

How Items End Up in the Review Queue

Every transaction that flows into a client's Accounted ledger goes through Penny's categorisation engine. This is a three-tier process that operates on every transaction, whether it arrives via bank feed, manual entry, or receipt upload.

Tier 1: Rules-Based Matching

The first tier applies deterministic rules. If a transaction matches a known pattern — a recurring payment to a previously categorised supplier, a direct debit with a consistent reference, or a bank transfer between the client's own accounts — the categorisation is applied with high confidence. These transactions almost never appear in your review queue.

For example, if a client pays £45 per month to the same broadband provider, and the first three months were categorised as telecommunications, the fourth month is categorised identically with a confidence score of 0.99. There is no reason for you to see this transaction.

Tier 2: Contextual Matching

The second tier handles transactions that are less clear-cut but still identifiable. A payment to a new supplier whose name suggests a category (for instance, "Screwfix Direct" strongly suggesting building materials for a construction client), or a variable-amount payment to a known supplier where only the amount has changed.

These transactions receive moderate-to-high confidence scores, typically between 0.80 and 0.95. Depending on your firm's threshold settings, they may or may not appear in the review queue. At the default threshold of 0.90, a transaction scored at 0.92 would be auto-applied, while one scored at 0.87 would appear for your review.

Tier 3: LLM Reasoning

The third tier uses large language model reasoning for genuinely ambiguous transactions. An unusual payment to an unfamiliar payee, a transaction that could plausibly fall into multiple categories, or a pattern that does not match anything in the client's history.

These transactions receive lower confidence scores and are almost always flagged for review. Penny provides her suggested categorisation along with her reasoning, but she presents it as a suggestion rather than a determination.

The review queue aggregates all flagged items across your client base, creating a single prioritised list of items that need your attention. You do not need to open each client's records individually, scroll through transactions, and identify which ones need review. The queue does that for you.

What the Review Queue Looks Like

When you open the review queue, you see a list of items sorted by priority. Each item includes:

• Client name — whose records the transaction belongs to
• Transaction details — date, amount, payee or payer, and bank account
• Penny's suggestion — the proposed categorisation, including the account code and VAT treatment
• Confidence score — a percentage indicating how confident Penny is in her suggestion
• Flag reason — why the item was flagged (low confidence, unusual amount, new supplier, possible duplicate, VAT uncertainty, etc.)
• Context — relevant prior transactions with the same payee, the client's typical spending patterns, and any notes

You can filter the queue by client, by flag reason, by date range, or by confidence band. If you want to process all items for a specific client in one sitting, you can filter to that client. If you want to handle all "new supplier" flags across your practice, you can filter by that reason.

For each item, you have three options:

1. Approve — accept Penny's suggestion as-is
2. Amend — change the categorisation, VAT treatment, or other details
3. Escalate — mark the item as requiring further information (from the client or from a senior team member)

When you approve or amend an item, the decision is logged and feeds back into Penny's learning for that client. If you amend a categorisation, Penny learns from the correction and applies it to similar future transactions. Over time, this means fewer items are flagged for each client as the AI learns their specific patterns and your preferences.

Types of Flags and How to Handle Them

Understanding the different types of flags helps you process the queue efficiently. Each flag type has a typical resolution pattern.

Low Confidence

These are transactions where Penny is uncertain about the correct categorisation. The most common causes are unfamiliar payees, transaction descriptions that are vague or truncated, and payments that could legitimately fall into multiple categories.

Resolution is typically quick: review Penny's suggestion, check the amount and payee, and either approve or amend. Most low-confidence items have a reasonable suggestion that just needs confirmation.

New Supplier

Any transaction involving a payee or payer that has not appeared in the client's history before is flagged as a new supplier. This is a precautionary flag. Penny may still have a high-confidence categorisation based on the supplier name, but the newness itself triggers a review to ensure nothing unusual is happening.

These flags reduce naturally over time as a client's supplier base stabilises. In the first few months of a client using Accounted, you will see more new supplier flags. After six months, they become infrequent unless the client's business is genuinely changing.

Unusual Amount

When a transaction to a known supplier deviates significantly from the historical pattern, it is flagged. If a client usually pays their accountant £200 per month and a payment of £2,000 appears, that is worth reviewing even though the supplier is known and the categorisation is clear. It might be a year-end fee, or it might be an error.

VAT Uncertainty

Transactions where the VAT treatment is unclear are flagged separately. This often occurs with mixed-supply purchases, items that could be zero-rated or standard-rated depending on context, or transactions with suppliers who sell both VATable and exempt goods. VAT flags require particular care because errors have direct financial consequences on VAT returns.

Possible Duplicate

Penny identifies potential duplicate transactions by comparing amounts, dates, and payees. If two transactions of the same amount to the same payee occur within a few days, one may be flagged as a possible duplicate. This catches accidental double-payments and data import issues.

Requires Receipt

While not strictly a categorisation flag, items that require a receipt for HMRC compliance but do not have one attached appear in the queue. This ensures that you are aware of gaps in the client's records during your review, rather than discovering them at year-end.

Setting Your Firm's Threshold

The confidence threshold is the most important configuration decision in the practice portal. It determines the boundary between "auto-apply" and "flag for review," and it directly controls how many items appear in your queue.

Accounted's default threshold is 0.90, which means transactions with a confidence score of 90% or above are auto-applied, and those below are flagged. This is a sensible starting point, but you may want to adjust it based on your risk appetite and workload.

Conservative practices (threshold 0.95) will see more items in the queue. This is appropriate if you are new to AI-assisted review and want to verify accuracy before trusting the system with more autonomy. It is also suitable for clients with complex transactions or high regulatory exposure.

Moderate practices (threshold 0.85-0.90) will see a balanced queue that catches genuine exceptions without including too many straightforward items. This is where most practices settle after a few months of use.

Efficient practices (threshold 0.80) will see a lean queue focused only on genuinely ambiguous items. This is appropriate for clients with well-established patterns and for accountants who have verified the AI's accuracy over multiple review cycles.

You can set different thresholds for different clients. A straightforward sole trader with regular transactions might have a threshold of 0.85, while a construction subcontractor with complex CIS deductions and variable income might have a threshold of 0.95.

The ICAEW's guidance on AI in accountancy supports risk-based approaches, noting that professional judgement should focus on areas of material risk. Setting appropriate thresholds is an exercise in professional judgement itself: you are defining what "material risk" means for each client in terms of categorisation accuracy.

The Review Queue in Practice: A Worked Example

Consider a practice managing eighty clients, with quarterly MTD reviews due. Without the review queue, you would need to review every transaction for every client — potentially thousands of line items across the client base.

With the review queue, the numbers look different. Assuming an average of 200 transactions per client per quarter, that is 16,000 transactions across the practice. At a typical flag rate of 12%, the review queue contains approximately 1,920 items.

Of those flagged items:

• Roughly 40% are new supplier flags that take 15-30 seconds each to resolve (approve Penny's suggestion after a quick check)
• About 25% are low-confidence items that take 30-60 seconds each
• Around 15% are unusual amounts that need 1-2 minutes of investigation
• The remaining 20% are VAT queries, possible duplicates, and complex items that may take 2-5 minutes each

Working through the maths, the total review time for 80 clients comes to approximately 30-40 hours per quarter, compared to 160-240 hours for comprehensive line-by-line review. That is a time saving of 75-80%.

This is not theoretical. Practices using Accounted's review queue routinely report similar figures. The case study of a practice managing 200+ clients provides real-world data on review time reduction.

Integrating the Review Queue into Your Workflow

The review queue works best when it is part of a structured workflow rather than something you dip into sporadically. Here are the approaches that work for the practices we see managing the highest client volumes.

Daily Triage

Spend 15-20 minutes at the start of each day scanning the queue for urgent items: transactions close to a filing deadline, high-value items, or escalated queries. Handle these immediately and leave the routine flags for your dedicated review sessions.

Weekly Batch Review

Dedicate a block of time each week (many practices use Monday or Tuesday morning) to processing the review queue systematically. Work through clients in order of deadline urgency, approve or amend items, and clear the queue for the week. This pairs well with the automation strategies that handle routine tasks between your review sessions. It also complements effective client communication practices, ensuring that when you do contact a client, it is about a substantive issue rather than a routine query.

Pre-Submission Review

Before any filing deadline, do a targeted review of the relevant client's flagged items. The queue can be filtered to show only items within a specific date range, so you can focus on the quarter being submitted.

Making the Most of Feedback Loops

Every decision you make in the review queue feeds back into Penny's categorisation model for that client. This creates a virtuous cycle: the more you review, the more accurate the AI becomes, and the fewer items appear in future review cycles.

To maximise this effect:

• Be consistent — if you categorise similar transactions differently for the same client, the AI receives contradictory signals
• Use the amend function rather than deleting and re-entering — amendments teach the AI what the correct categorisation should have been
• Review promptly — the sooner you process flagged items, the sooner the learning applies to subsequent transactions

Over three to six months, a client's flag rate typically drops from 15-20% in the first quarter to 5-10% as Penny learns the specific patterns of their business. For clients with highly regular transaction patterns, the rate can drop below 5%.

Getting Started with the Review Queue

If you are new to Accounted's practice portal, the review queue is available from the moment you connect to your first client. There is no separate setup or configuration required beyond setting your preferred confidence threshold.

For a walkthrough of the initial setup, the accountant portal setup guide covers everything from creating your account to configuring your first client's review settings. The HMRC guidance on digital record-keeping provides the regulatory context for what records must be maintained, which helps you decide how to configure your review thresholds.

The review queue is, ultimately, the mechanism that makes exception-based review practical. Without it, you would need to manually identify which transactions need attention — which would take almost as long as reviewing them all. With it, the AI does the identification, and you apply your expertise where it matters. It is the difference between reading every email in your inbox and reading only the ones that are not handled by filters and auto-replies.

If you are ready to see the review queue in action, sign up for your free practice portal account and connect to a client's records. The queue populates immediately with any items that are below your confidence threshold, giving you an instant view of what needs attention.

Accounted gives accountants a free practice portal — manage all your clients, file to HMRC, and let Penny handle the routine work. See the accountant portal →