AI-assisted ICD-10 and CPT coding: cleaner claims without upcoding risk

AI medical coding uses AI to suggest diagnosis and procedure codes, such as ICD-10 and CPT, from what was documented in a visit. The safe pattern is simple. Codes come only from documented facts, each suggestion shows its evidence, and a clinician or coder confirms it. Verto suggests codes this way, on the same record that feeds billing and claims.

Why coding still goes wrong

In the UAE, insurance claims through eClaimLink in Dubai and Shafafiya in Abu Dhabi carry diagnosis and procedure codes that must match the documentation. Coding problems usually fall into a few groups:

• Low specificity. An unspecified code where the note supports a more specific one.
• Mismatch. The code says one thing, the note another, or the diagnosis does not support the procedure billed.
• Missed codes. A procedure or secondary diagnosis documented in the note but never coded.
• Late coding. Codes chosen days later by someone who was not in the room, working from an incomplete note.
• Upcoding. A code that claims more complexity or a bigger service than the documentation supports.

Each one can lead to rejections, rework or audit findings. Our guide to reducing claim denials in UAE clinics covers the wider denial picture.

How AI coding assistance works

1. The scribe transcribes the visit and extracts clinical facts: diagnoses, findings and procedures.
2. A model maps the documented facts to candidate codes, each with a confidence level.
3. Each suggestion links to the evidence in the transcript that supports it.
4. The clinician or coder accepts, rejects or changes each code, and can add codes by hand.
5. Only accepted codes travel onward to the note, the invoice and the claim.

The value is timing. Codes are proposed while the visit is fresh, by the clinician who was there, from facts that were checked a moment earlier.

This changes the coder's job rather than removing it. Instead of reading a free-text note days later and guessing at detail, the coder reviews codes that already point to their evidence. Time moves from hunting for information to checking it, and to the complex cases, payer rules and rejection follow-ups that need an expert. Clinicians benefit too. A suggestion that shows the missing detail, such as which side or which type, is a prompt to document it properly while the patient is still there.

The upcoding risk, and how to avoid it

Any tool that suggests codes can nudge people towards higher-paying ones, whether by design or by accident. Upcoding is not a billing optimization. It is a compliance risk that can lead to recoveries and penalties. A safe AI coding tool follows four rules:

This is where traceability pays off twice. The same link from fact to audio that protects the clinical note also protects the claim. Our article on AI scribe accuracy and traceability explains that chain in detail.

How Verto handles coding

• Suggestions from the visit. Verto suggests ICD-10 diagnosis codes and CPT procedure codes from the transcript and the reviewed note, each with a confidence level and its supporting evidence.
• The chart confirms. ICD-10 suggestions appear in the chart's diagnosis block as AI suggestions, with accept and dismiss buttons. Accepting adds the code to the visit's diagnosis list, tagged as AI-suggested. The chart is the single place a code is committed.
• Only accepted codes move on. Rejected and unreviewed suggestions are left out of exports.
• Manual codes are easy. If Verto misses a code, add it by hand.
• Changes are caught. If the note changes and codes are re-analysed, previously accepted codes return for review.

Because Verto sits inside Helix, accepted codes are already on the record that Helix billing and claims uses. In the UAE, that same record connects to eClaimLink and Shafafiya. See our UAE compliance overview for the full list of connections.

Documentation drives the code

AI cannot code what was never documented. The most useful thing a scribe does for coding is to capture the details that decide specificity while the patient is still in the room:

• Laterality: right, left or both, for eyes, ears, joints and limbs.
• Type and cause: for example, the type of diabetes, or whether a condition is due to another one.
• Acuity and stage: acute or chronic, initial or follow-up, and the stage where it applies.
• Complications and related conditions that change the code.
• What was actually done: the procedure performed, not only the one discussed.

When the scribe captures these as reviewed facts, the code suggestion follows from them, and the evidence is already in place if a payer asks.

What AI coding does not fix

Better codes help, but many rejections have nothing to do with coding. AI coding will not fix a missing prior authorization, a lapsed policy, wrong patient or member details, or a service outside the payer's contract. These need clean front-desk data, an authorization workflow and a claims team that tracks rejection reasons. Keep coding in proportion: it is one link in the revenue cycle, not the whole chain.

A note on Saudi Arabia and other markets

Code sets differ by country and payer. Saudi claims through NPHIES use code sets that differ from the UAE's, so check exactly which code systems any tool supports for your market before you rely on it. For Saudi claims, Helix connects to NPHIES via Waseel; our Saudi Arabia page explains the setup.

Rolling out AI coding safely

1. Involve your coders early. They know your payers' rules and common rejection reasons.
2. Audit a sample. For the first weeks, compare accepted AI-suggested codes with a coder's independent review.
3. Watch the direction of change. If average coding levels rise, check that documentation supports it.
4. Do not bulk-accept blindly. Accept-all buttons save time on routine visits, but only after reading.
5. Feed rejections back. Track why claims were rejected and whether the cause was the code, the documentation or the authorization.