AI for Medical Billing Companies: Seven Jobs to Automate First

A medical billing company should automate first the jobs that repeat in a fixed shape and need no patient detail: client onboarding paperwork, template-based denial and appeal letters, client reports and staff procedures. Eligibility exceptions, aged-claim worklists and missed-charge flags come next, once a business associate agreement covers the tool, with a person approving every claim-affecting step.

The order matters more than the list. Billing is a business of small errors with money attached, so the safest sequence starts with documents and reports around the claims and moves towards the claim data only when your controls are working. Nothing here lets AI change a claim by itself; the model drafts, sorts and flags, and a trained biller decides. The seven jobs below follow a client from sign-up to month-end, each with its risk and its human check, and the denial letter gets the full worked example because most firms want it first. The closing section sets the order in which to switch them on.

Follow me on Instagram@sagnikteaches

Before job one: what a business associate agreement does

If your clients are healthcare providers covered by health-privacy law, you are a business associate of theirs, and any vendor that handles protected health information for you becomes your subcontractor. That chain is held together by business associate agreements (BAAs). The practical consequences are these.

Connect on LinkedInSagnik Bhattacharya

• Do not paste patient names, dates of birth, member IDs, claim numbers or diagnosis codes into a consumer chat tool. Consumer plans generally do not come with a BAA.
• If you want an AI tool to touch real claim data, the vendor must sign a BAA covering that specific product. Anthropic's help page, for example, lists Claude Enterprise (once its Primary Owner accepts the BAA) and named API features as eligible, while the Files and Batch APIs and data sent to third-party connectors are not covered. OpenAI's page says it signs BAAs for the API and for sales-managed ChatGPT Enterprise or Edu accounts, and does not offer one for ChatGPT Business. Check the exact product and feature you would use.
• Your client contracts should say whether you may use AI subprocessors at all. Some clients will refuse. Ask, record the answer and respect it per client.

Jobs one, three, four and seven need no patient data if you work from templates and placeholders. Jobs two, five and six touch claim-level information, so they need de-identified exports or a settled BAA position first. Take legal or compliance advice on your own agreements; a tutorial is not legal advice.

Subscribe on YouTube@sagnikteaches

1. Client onboarding paperwork and checklists

Every new practice means the same cycle: contract, BAA, access to the practice management system, clearinghouse enrolment, payer lists, fee schedules, a first-week plan. AI can draft the welcome pack from a template, build a client-specific checklist from a short intake form, and write the chasing emails when documents are late.

Say a company onboards eight practices a quarter and each takes about three hours of document chasing. A good prompt takes the practice type and systems and returns the checklist:

Create an onboarding checklist for a new client: a 4-provider orthopaedic practice using [PM system], enrolling with [clearinghouse], payers: one government programme, two commercial plans, workers' compensation. Include: documents we need from them, access we need, enrolment forms, a day-by-day first-week plan, and a list of things we must NOT receive by email. No patient information is included in this request.

The output is a table with owners and due dates. What you fix: the draft will list generic forms, and payer-specific enrolment forms change, so confirm each against the payer's current page. The saving is real (the chase email alone saves ten minutes a time), and the risk is low. For the contract and access side, the onboarding tutorial for billing firms goes further.

2. Eligibility check follow-up, not eligibility itself

Verifying coverage runs through the clearinghouse or payer portal, using standard electronic transactions (the 270 request and 271 response). That is not an AI job and should stay with your verified systems. What AI can do is handle the exceptions afterwards: group the failed or ambiguous responses by reason (inactive coverage, wrong member ID, coordination of benefits, plan requires referral) and draft the call script or the message to the practice asking for corrected details.

Illustrative: a batch of 120 verifications returns 14 problems. A biller exports the 14 with member details removed and keeps a reference number, then asks for a grouping. The assistant sorts them: six inactive coverage, four ID mismatches, three secondary-payer questions, one unclear. The biller opens each record by reference number, checks the source, and sends the practice front desk a short note per group. The tool never sees the patients' identities, and the person decides what each exception means. If your vendor supplies built-in AI inside the eligibility tool under a BAA, you can go further, but keep the review step.

3. Denial letters and appeal drafts from templates

This is the job most firms want first, and the one where discipline matters. The idea: build a library of appeal letter templates by denial type (medical necessity, missing authorisation, timely filing, bundling, coordination of benefits), then use AI to adapt the wording around the facts, which a person supplies and checks.

A worked example, with illustrative numbers. A company handles 300 denials a month, and 60 per cent fall into five types. Letter drafting takes an average of 18 minutes. If a template plus AI-assisted adaptation brings it to 8 minutes, 180 letters save 30 hours a month. At a biller cost of $25 an hour that is $750 in time, which is modest on its own and large against a team already stretched.

The prompt works without patient identifiers, using placeholders that your system fills at merge:

Draft an appeal letter for a denial of type "missing prior authorisation". Facts I will supply as placeholders: [DATE OF SERVICE], [PROCEDURE DESCRIPTION], [DENIAL CODE AND REASON], [AUTH STATUS: authorisation requested on DATE / retro-authorisation request attached]. Tone: factual, firm, courteous. 250 words. Do not cite clinical evidence, medical policy numbers or laws. Mark [CITATION NEEDED] where a payer policy reference would strengthen it.

Sample output (illustrative): a short letter requesting reconsideration, summarising the timeline in neutral terms, listing the attached documents and requesting a written decision. What you fix: the draft sometimes asserts "this service was medically necessary" without any clinical support, which is a statement only the provider can make; delete it or replace it with "the treating provider's documentation is enclosed". The citation flags tell the biller where to look up the payer's actual policy. For the full denial workflow, grouping, appealing and tracking denials has the grouping step, and ChatGPT for appeals without patient information covers the no-identifiers method in more depth.

4. Client reports and month-end commentary

Practices want a monthly report: collections, days in AR, denial rate, top denial reasons, actions next month. The numbers come from your system; the commentary is the time sink. Paste the summary figures (no patient-level data) and ask for a one-page narrative and three suggested actions.

Say the summary shows collections of $412,000 against $398,000 the month before, days in AR at 38 down from 41, and a denial rate of 9.2 per cent up from 8.1. The draft will note the rise in denials and suggest looking at the top reason. The check: confirm every figure against the system export, and rewrite any suggestion that assigns blame to the practice without evidence. A line such as "front desk staff are entering insurance details incorrectly" is a guess unless your data shows it. Replace it with "ID mismatches account for 31 of the 74 denials; we suggest a check at registration".

A table the client sees every month helps:

Figures are illustrative. Build the table in your reporting tool and let AI draft only the final column.

5. Aged-claim worklists

With a BAA position settled, or using de-identified exports, AI can help rank an accounts receivable list. Give it fields such as age in days, payer, amount, last action date and denial code, and ask for a prioritised worklist with a reason for each rank. The value is consistency: instead of each biller sorting differently, everyone follows the same rule, such as oldest and highest value first, but never let anything approach a payer's timely-filing limit unchecked.

A realistic mistake: a tool ranks claims by amount alone and a $90 claim two weeks from its filing deadline sinks to the bottom. Fix it by encoding the rule in the prompt (anything within 30 days of the payer's filing limit goes to the top, with the limit supplied by you for each payer) and by spot-checking the top and bottom 20 each week. For the full method, see prioritising aged claims with AI.

6. Charge-capture and missed-charge flags

This is the first job that touches clinical documentation, so it sits later. AI can compare visit notes against charges and flag possible gaps for a person to review: a procedure mentioned in the note but not billed, a time-based service without the time recorded. Treat every flag as a question, never a charge. Coding decisions belong to certified coders and the provider, and adding a code a note does not support is a compliance problem however confident the tool sounds. Spotting missed charges from visit notes shows a safe review set-up, and the audit-trail question is covered in the further reads below.

The control is a sample audit. Each month, have a second coder review 30 flagged and 30 unflagged encounters, record the result and track false positives and misses. If misses exceed what you accept, switch it off.

7. Staff procedures, training and quality checks

The least glamorous and often the most valuable. Billing firms depend on tacit knowledge: how to handle a particular payer's portal, what to do when a remittance does not match a deposit. Record a biller explaining the steps and have AI turn the transcript into a numbered procedure, a one-page quick reference and five quiz questions for new starters. Use screenshots with no patient data. A reviewer confirms every step by following it. This makes you less dependent on the one person who knows, and it makes onboarding staff faster.

A before and after: the "payer X appeals" knowledge sat in one person's head and took a new hire three weeks to absorb. A 12-step procedure, reviewed once, cuts that to three days of supervised practice. The numbers are illustrative; time your own.

A filled-in client report checklist for the first quarter

A new automation earns trust when you can show a client what changed. After the first quarter, a billing company can send a short note built from three measures. Illustrative figures for a four-provider practice: letters drafted in 8 minutes instead of 18 (about 25 hours a month back across 150 letters), appeal letters sent within three working days of denial instead of seven, and onboarding for the next new client finished in 9 working days instead of 14. Ask the assistant to turn those three bullets into a client-friendly paragraph, then strike any sentence that implies a result you cannot prove, such as "we now recover more revenue". Recovery is measured from payments, not from speed. If you offer to share the numbers, share the ones you can reproduce from your system.

Choosing a first tool without overbuying

The seven jobs do not all need a specialist product. Jobs one, four and seven run on a general assistant on a business plan. ChatGPT Business and Claude Team both cost around $20 to $25 per user per month on current list prices, with a two-seat minimum on both, so a company with five billers on the pilot pays roughly $100 to $125 a month. Neither entry plan is a home for patient data: OpenAI says it does not offer a BAA for ChatGPT Business, and Anthropic's BAA page names Enterprise and the API rather than Team. Jobs two, five and six need either de-identified exports or a vendor that signs a BAA for the specific product, which usually means an enterprise plan, an API route or a billing-platform feature. Ask your existing practice-management or clearinghouse vendor first, because they may already include an AI feature under an agreement you hold. Buying a second tool for a job your current system can do wastes money and creates another subprocessor to manage.

A short scorecard for any vendor: Will you sign a BAA for this product and plan? Where is data processed and for how long retained? Is client data used to train models? Can staff access be limited by client? Can I export logs showing who used it on what? A vendor that answers four of five in writing is workable; fewer than that, pause.

The human review points, written down

Every claim-affecting step needs a named human. Put it in a table your auditors can read.

Write-offs and adjustments get no AI involvement in this plan, because they move money and are the easiest place for an error to hide.

A day of denial work, step by step

Here is how the denial job runs end to end for one biller on a normal morning, as an illustration. At 8:30 she opens the denial worklist exported from the billing system and filters to the five template types. By 8:45 she has grouped 22 denials: nine missing authorisation, six eligibility, four bundling, three timely filing. For each group she opens the template, fills the placeholders from the record (dates, codes, authorisation status) and, where the template is adapted by AI, pastes only the de-identified fact lines. By 10:00 twelve letters are drafted. She reads each against the claim, adds attachments, and flags two where the practice must supply a clinical note. At 11:00 the letters are sent and the system records the date and the next follow-up in 30 days.

Look at where the checks sit: before drafting (the grouping is verified against the record), after drafting (a full read against the claim), and at sending (attachments and dates). If any stage is skipped under time pressure, that is the point to stop, because a wrong date on an appeal can forfeit it. A weekly sample of ten sent letters, read by a lead, gives you a quality rate you can show a client who asks how AI is being used.

Four billing failures and the first sign of each

• A letter with a made-up policy reference. Shows up when an appeal is rejected for citing a policy that does not exist. Prevent it with the [CITATION NEEDED] rule and a lookup against the payer's site.
• Patient data in the wrong tool. Shows up as a screenshot in a team chat. Prevent it by blocking consumer AI sites on work devices if you can, and by giving staff an approved option so they are not tempted.
• Template drift. A template is edited once to fix one case and quietly becomes the default. Version-control the library and review it quarterly.
• Silent automation creep. A feature turned on in your practice-management software adds AI summaries nobody reviewed. Read release notes and keep a register of AI features by system.

Switch-on order for the seven jobs, and what to measure

Begin with jobs one, four and seven for a month, which need no patient data, and add job three once the template library is built and a reviewer has signed it off. Add jobs two, five and six when the BAA position is settled and the audit sample is running. Measure hours per task before and after, appeal rejection rate for AI-assisted letters against the rest, and the error rate found at review. If you are weighing whether to keep billing in-house, the outsourcing versus in-house decision rule may help.

If you would rather map the order with someone, that is what my AI implementation consultation covers; get in touch.

Questions medical billing owners ask before automating

Do I need a BAA before using AI on claim data?

If the data includes protected health information and your clients are covered by a health-privacy rule, the AI vendor handling it needs a business associate agreement with you, and you need one with each client. Consumer plans do not offer one, and OpenAI says it offers none for ChatGPT Business either. Check the vendor's own page, and ask your clients' compliance contacts.

Can AI submit or change claims on its own?

It should not. Anything that alters a claim, a code, a payment posting or a write-off needs a trained person to review and approve it. Use AI to draft, sort and flag, then keep a named human as the decision maker and the audit trail.

Which job saves the most time first?

Usually denial letters and appeal drafts, because they repeat in a few standard shapes, and client onboarding because the same documents are chased every time. Start where the work is repetitive, the text contains little or no patient detail, and a mistake is caught by review before it leaves.

How should I charge clients if AI makes work faster?

Many billing companies charge a percentage of collections, so speed lifts margin without changing the fee. If you bill per claim or hourly, tell clients plainly what changed and what stayed the same. Do not cut human review to protect margin.

Further reads

• AI Medical Coding Software: Audit Trails and Upcoding Risk — Audit trails and upcoding risk before automating coding.
• Can AI Do Medical Billing and Coding? What Still Needs a Human — What still needs a human in billing and coding.
• Medical Practice AI Budget: What to Fund First and What It Costs — What to fund first and what it costs.
• Medical Billing Proposal: Draft Practice Quotes With AI — Draft practice quotes faster.
• Best AI Tool for Medical Billing: Scrubbing, Coding, Denials — Tool types matched to scrubbing, coding and denials.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic business associate agreements help page; OpenAI help article on getting a business associate agreement; ChatGPT Business and Claude pricing pages; a clearinghouse's 270/271 eligibility API documentation; checked 4 November 2026.