AI in Accounts Receivable: Use Cases, Costs, 90-Day Plan

AI in accounts receivable is no longer a pilot topic for large enterprises. In a September 2026 survey of more than 600 businesses in the United States, Canada and Mexico, credit insurer Atradius found that seven in ten companies deal with late payment and that overdue invoices average 23% of B2B receivables. Nearly a quarter of the money your customers owe you is, on average, already late. The same survey found that 43% of sales are made on credit terms. That is the problem AI in accounts receivable is built to attack, and it is a cash problem before it is a technology problem.

This guide is written for CFOs, controllers and founders who run a credit-based B2B business and want to know what AI can actually do in the order-to-cash cycle, what it costs, where it fails and how to roll it out in 90 days without breaking the controls your auditors rely on.

I have founded companies across very different industries over the last twenty years. The pattern I see in receivables is always the same: the team is busy, the aging report is long, and nobody can say with confidence which ten customers will pay late next month. AI does not fix bad credit policy. It does make good policy enforceable at scale.

What AI in Accounts Receivable Actually Means

Accounts receivable automation has existed for decades. Rules engines send a reminder 7 days after the due date, lockbox files post payments that match an invoice number exactly, and dunning letters escalate on a fixed calendar. That is automation. It executes what you told it to do and stops at the first exception.

AI in accounts receivable is different in one specific way: it handles the exceptions and it predicts. A machine learning model reads a remittance that lists three invoices with a short payment, figures out which deduction belongs to which invoice, and proposes the posting. Another model looks at a customer's payment history, order pattern and open disputes, and estimates the probability that the next invoice will be paid late.

In practice, AI in AR covers four families of capability:

• Document understanding: reading remittance advice, emails, PDFs and portal statements, and turning them into structured data.
• Matching and reconciliation: linking incoming cash to open invoices when references are missing, partial or wrong.
• Prediction: payment date forecasts, late payment risk scores, dispute likelihood and credit risk signals.
• Generation: drafting collection emails, dispute responses and customer statements in the right tone and language.

The first two reduce manual work. The last two change decisions. Most companies get their fastest payback from the first two and their biggest strategic gain from the last two.

Why receivables is a good fit for AI

Receivables has three properties that make AI work better here than in many other finance processes.

1. The data already exists. Every invoice, payment, dispute and reminder is recorded in your ERP or billing system. You do not need to collect new data to start.
2. The feedback loop is fast. You know within days whether a predicted payment date was right. Models improve quickly when they get frequent, unambiguous feedback.
3. The value is measurable in cash. Days sales outstanding (DSO), overdue balance and bad debt write-offs are hard numbers. You do not have to argue about soft productivity gains.

The State of AI Adoption in Finance

Finance teams have moved faster than many people assume. Gartner surveyed 121 finance leaders in June 2024 and found that 58% of finance functions were using AI, up from 37% a year earlier, as reported by the Journal of Accountancy. The most common use cases were intelligent process automation (44%), anomaly and error detection (39%), analytics (28%) and operational assistance (27%).

Gartner's follow-up in November 2025 described adoption as steady rather than accelerating after that jump. My reading: the easy wins have been claimed, and the next wave depends on companies wiring AI into specific processes with clear owners, which is exactly where receivables sits.

The payment environment is also getting riskier. The 2025 AFP Payments Fraud and Control Survey, summarized by Nacha, found that 79% of organizations experienced actual or attempted payments fraud in 2024, that 63% were hit by check fraud, and that business email compromise touched 63% of organizations. Receivables teams handle bank details, remittance emails and customer master data every day. Any AI you introduce has to make that surface safer, not wider.

What the research says about productivity

The strongest controlled evidence on generative AI at work comes from outside finance, but it transfers well to collections and dispute handling. A study of 5,179 customer support agents by Brynjolfsson, Li and Raymond, published as NBER Working Paper 31161, found that access to an AI assistant raised productivity by 14% on average, with the largest gains among less experienced agents.

That matters for AR because collections is a customer conversation. A junior collector with a good assistant writes like a senior collector. That is a staffing and training advantage, not just a speed advantage.

AI in Accounts Receivable Use Cases That Pay Back First

Not every use case deserves your first 90 days. Here are the seven I see delivering value, ordered roughly by how quickly they pay back.

1. Cash application

Cash application is the process of matching incoming payments to open invoices. It is where most AR teams lose the most hours, because customers pay in bulk, short pay, take unearned discounts and send remittance details by email, portal or not at all.

AI cash application tools read the remittance wherever it arrives, propose a match with a confidence score and route low-confidence items to a person. The goal is not 100% automation. The goal is that people only touch the payments that genuinely need judgment.

What to measure: auto-match rate, unapplied cash at month end, time from receipt to posting.

Where it fails: when customer master data is dirty. If the same customer exists under four IDs, the model will learn the mess. Clean the master data first.

2. Collections prioritization

Most collection teams work the aging report from the top: biggest balance, oldest invoice. That feels logical and wastes effort. Many large, old balances belong to customers who always pay, just late, while a mid-sized new customer quietly slides toward default.

An AI model scores each open invoice by the probability of late payment and the expected loss if it is not collected. Collectors start their day with a worklist ranked by risk and value, not by age alone.

What to measure: percentage of overdue balance recovered within 30 days, collector touches per dollar collected, roll rate from 30 to 60 days past due.

3. Payment date prediction and cash forecasting

If you know when each invoice is likely to be paid, your cash forecast stops being a spreadsheet of hopes. AI models predict a payment date per invoice based on the customer's history, the invoice size, seasonality and open disputes. Rolled up, that gives treasury a 13-week forecast grounded in behavior instead of payment terms.

This is where receivables and finance planning meet. If you want the forecasting side in depth, I covered it in the AI cash flow forecasting guide.

4. Dispute and deduction management

In industries like distribution, consumer goods and wholesale, deductions are a silent margin leak. Customers deduct for pricing errors, damaged goods, promotions and late deliveries, and many of those deductions are invalid. Nobody has time to chase them all.

AI reads the claim, classifies the reason code, gathers proof of delivery and pricing documents, and recommends whether to accept or contest. Invalid deductions that used to be written off now get disputed with evidence attached.

What to measure: deduction recovery rate, average days to resolve a dispute, write-offs by reason code.

5. Credit risk monitoring

Credit decisions for new B2B customers and credit limit reviews for existing ones are usually done once a year, if at all. AI makes credit monitoring continuous. It watches internal signals such as slowing payments, shrinking order sizes and rising disputes, and external signals such as credit bureau changes and public filings.

The output is not an automatic decision. It is an alert that tells the credit manager which accounts to review this week. If you lend money rather than extend trade credit, the logic and the regulation are different, and I covered them separately in AI in lending.

6. Collection communications

Generative AI drafts reminders and follow-ups that reference the specific invoice, the customer's history and any open dispute. It adapts tone by segment: a polite nudge to a strategic account, a firmer note to a chronic late payer, a reply in the customer's language.

The rule I insist on: a person approves every message type before it goes live, and high-value or disputed accounts always get a human review. Generated text is fast. A wrong sentence to your biggest customer is expensive.

7. Anomaly and fraud detection

AR is a fraud target. Fraudsters impersonate customers to request refunds to new bank accounts, or impersonate your company to redirect customer payments. AI flags unusual patterns: a refund request to an unfamiliar account, a remittance email from a lookalike domain, a credit memo issued outside normal limits.

Given the AFP numbers above, this belongs in every AR program, even if it is not where you start.

The Business Case: How to Size the Return

You do not need a vendor's ROI calculator. You need three numbers from your own ledger.

1. Average receivables balance. Your total open AR, averaged over 12 months.
2. Daily revenue on credit. Annual credit sales divided by 365.
3. Your cost of capital or borrowing rate. What it costs you to finance working capital.

Every day of DSO you remove releases one day of credit sales in cash. Here is a worked example with round numbers, not a benchmark:

The cash release is a one-time gain. The financing saving recurs every year. Add the hours saved in cash application and collections, and the reduction in write-offs, and you have the full picture.

Be conservative. Assume a modest DSO improvement in year one. If the project only works with heroic assumptions, it is not a good project.

For a broader method to build the business case across finance, the AI for CFOs guide walks through how to present it to a board.

If you want a second pair of eyes on your numbers before you commit budget, you can request a consultation through the consulting request page on this site. I will tell you honestly if the case does not hold.

What It Costs to Implement AI in AR

Costs depend more on your data and your ERP than on the AI itself. Here is how to think about the budget lines rather than a price list, because pricing changes fast and varies by volume.

Software. Most AR platforms price by invoice volume, by number of customers or by module. Cash application, collections and credit are often separate modules. Ask for pricing at your current volume and at twice your volume, so you know what growth costs.

Integration. Connecting to your ERP, bank feeds and email is usually the largest one-time cost. Modern cloud ERPs have standard connectors. Older on-premise systems often need custom work.

Data cleanup. Customer master data, payment terms and historical dispute codes almost always need work. Budget internal time for this, not just vendor time.

Change management. Collectors and credit managers need to trust the scores. That takes training, side-by-side runs and a few weeks of patience.

Ongoing ownership. Someone has to monitor model performance, review exceptions and update rules. If nobody owns it, accuracy decays quietly.

Build, buy or extend your ERP

You have three paths:

For most companies under a few hundred million in revenue, buying or extending the ERP beats building. Build only where your receivables process is genuinely unusual and is a source of competitive advantage.

Self-Assessment: Is Your AR Ready for AI?

Score each statement from 0 (not true) to 2 (fully true). Be honest. This takes ten minutes and saves months.

How to read your score:

• 16 to 20: You are ready. Start with cash application and collections prioritization in parallel.
• 10 to 15: Ready with caveats. Fix master data and dispute codes in the first 30 days, then pilot one use case.
• Below 10: Fix the process first. AI on top of a broken process automates the mess. Map the order-to-cash flow, then come back. The order-to-cash process guide is a good place to start.

A 30/60/90-Day Roadmap for AI in Accounts Receivable

This roadmap assumes a mid-sized company with a finance team of five to twenty people. Scale the timing to your size.

Days 1 to 30: Baseline and foundations

• Measure the baseline. DSO, overdue percentage by bucket, auto-match rate, unapplied cash, dispute cycle time, bad debt rate. Without a baseline, you will never prove the result.
• Clean the customer master. Merge duplicates, fix payment terms, confirm billing contacts.
• Standardize dispute reason codes. Ten clear codes beat forty vague ones.
• Pick one use case. Usually cash application if your volume is high, collections prioritization if your volume is moderate but overdue balances are large.
• Shortlist vendors. Three at most. Ask each for a pilot on your own historical data.

Days 31 to 60: Pilot on real data

• Run the pilot on a subset. One business unit, one region or one customer segment.
• Run side by side. The AI proposes, people decide, and you compare. This builds trust and exposes errors before they matter.
• Track exceptions. Every item the model gets wrong is a data quality signal. Fix the root cause, not just the item.
• Define approval rules. Which matches post automatically above which confidence level? Which messages need human sign-off?

Days 61 to 90: Decide and scale

• Compare to baseline. If the pilot moved the numbers, expand. If it did not, find out why before spending more.
• Document controls. Write down what the AI does, what people approve and how exceptions are logged. Your auditors will ask.
• Extend to the next use case. Usually payment prediction feeds cash forecasting next.
• Assign ongoing ownership. One person reviews model performance monthly.

Measuring Success: The AR Scorecard

AI projects in finance die when nobody can show what changed. Before you sign a contract, agree on the scorecard you will report every month. Keep it short enough that the CFO actually reads it.

Two rules for the scorecard. First, compare like with like: seasonality distorts DSO, so compare the same month year over year, not just month to month. Second, report the gap between actual DSO and best possible DSO. That gap is the part of DSO your team can influence. The rest is your payment terms, which is a commercial decision, not a collections problem.

Leading and lagging indicators

DSO and bad debt are lagging indicators. They tell you what already happened. For an AI program you also need leading indicators that move within weeks:

• Model accuracy on payment dates: the share of invoices paid within a few days of the predicted date.
• Worklist adoption: the share of collector actions that follow the AI-ranked list.
• Override rate: how often people reject the model's suggestion, and why.

A high override rate is not a failure. It is information. Either the model is wrong and needs better data, or your team has knowledge the model does not see yet. Both are worth knowing.

Choosing a Vendor: Ten Questions to Ask

Vendor demos all look the same: a clean dataset, a high match rate and a dashboard full of green. The questions below separate products that work on real receivables from products that work in demos.

1. Will you run a pilot on our historical data before we sign? If the answer is no, walk away.
2. What match rate do customers with our ERP and our remittance mix actually reach? Ask for references in your industry.
3. How does the model explain a match or a risk score? A score without a reason cannot be audited or challenged.
4. What happens to low-confidence items? You want a clear exception queue, not silent guesses.
5. Where is our data stored and processed, and is it used to train models for other customers? Get the answer in the contract.
6. How are bank detail changes and refunds protected? Look for out-of-band verification and role-based limits.
7. Which connectors exist for our ERP, bank and email, and who maintains them? Integration is where projects slip.
8. How is pricing structured as our volume grows? Ask for the price at two times and three times current volume.
9. What does the audit trail look like? Ask to see the log for a single automated posting.
10. What does a typical first 90 days look like, and what do you need from us? Vague answers here predict a long project.

AI in Accounts Receivable by Company Size

The right first move depends on your size and complexity more than on the technology.

Small businesses with a few hundred invoices a month. You probably do not need a specialist platform. Use the AI features in your accounting software for reminders and payment matching, set clear payment terms and require electronic payment. Your biggest lever is policy: deposits, shorter terms for new customers and prompt invoicing.

Mid-market companies with thousands of invoices a month. This is where AI in AR earns its keep. Cash application and collections prioritization usually pay back within the first year when volume is high and the team is stretched. Integration with the ERP is the main project risk.

Enterprises with multiple entities and currencies. The complexity is in shared service centers, intercompany flows and dozens of remittance formats. AI adds most value in cash application across entities, deduction management at scale and continuous credit monitoring of large customer portfolios. Governance and model ownership matter as much as the tool.

Controls, Compliance and Audit Trail

Finance leaders worry, correctly, that AI weakens controls. It does not have to, but you have to design for it.

Segregation of duties still applies. The AI should not be able to create a customer, change bank details and issue a credit memo. Treat it like a user with a limited role.

Every automated posting needs a trail. Who or what posted it, on what evidence, with what confidence score. If you cannot reconstruct why a payment was applied to an invoice, your auditors will not accept it.

Bank detail changes need out-of-band verification. No AI-drafted email, and no email at all, should be enough to change where a customer sends money or where you send a refund. Call back on a number you already have.

Collection communications have legal limits. In the United States, the Fair Debt Collection Practices Act governs consumer debts, defined in 15 U.S.C. 1692a as obligations primarily for personal, family or household purposes. Pure B2B trade receivables generally fall outside it, but if you also bill individuals or sole proprietors for personal purposes, your AI-drafted reminders must respect those rules. Check with counsel before you automate consumer collections.

If you sell into Europe, know the AI Act basics. Since 2 February 2025, Article 4 of the EU AI Act asks companies that use AI to take measures that support the AI literacy of their staff. It does not prescribe a fixed course or duration. Creditworthiness assessment of natural persons is listed as high-risk in Annex III, with those obligations applying from 2 December 2027. Credit scoring of companies is a different case, but if your model scores individuals, involve your legal team early. My broader view on governance is in the AI compliance guide.

The Human Side: What Changes for Your AR Team

The fear in every receivables team is the same: AI will replace us. In my experience, the opposite risk is more common. Companies buy the tool, keep the old job descriptions and wonder why nothing improved.

Here is how roles change when AI in accounts receivable works:

• Cash application specialists move from keying payments to resolving exceptions and fixing the root causes behind them.
• Collectors spend less time sending reminders and more time on the accounts that actually need a conversation.
• Credit managers move from annual reviews to continuous monitoring, with a weekly list of accounts that deserve attention.
• The AR manager becomes the owner of a system, not just a team: accountable for model performance, data quality and policy.

Train for judgment, not clicks. The skill that matters is knowing when to override the model. Run sessions where collectors review the model's top-risk accounts and explain why they agree or disagree. That is how trust gets built, and how errors get caught.

Change incentives. If collectors are measured on number of calls, AI will not help. Measure them on cash collected per hour and on roll rates.

AI Agents in Receivables: What Is Real Today

Every AR vendor now talks about agents. It helps to separate what works from what is marketing.

What works today. Agents that perform a bounded task end to end with a person approving the result: gathering documents for a dispute, preparing a customer statement, drafting a reminder sequence for a segment, or researching why a payment was short. These tasks have clear inputs, clear outputs and an easy check.

What is still early. Agents that negotiate payment plans on their own, change credit limits without approval or decide whether to write off a balance. These touch money, customer relationships and accounting judgments. The technology can attempt them. Your controls should not allow them yet.

My test for any agent in finance is simple: could you explain to your auditor, in one sentence, what the agent is allowed to do and who checks it? If the answer takes a paragraph, narrow the scope. I wrote more about how to set those boundaries in the AI agent governance guide.

Lessons from Outside Finance

I have not run an AR transformation for a named client that I can share here, and I will not invent one. What I can share is what transferred from projects in other functions.

When a sports distribution company I worked with used AI to sharpen its marketing, sales grew by 30%. The lesson for AR is not about marketing. It is that the model only worked once the customer data was clean and the team trusted the output enough to act on it. The same is true for a collections score.

At a hotel I worked with, revenue grew from 9 million to 10 million. Better forecasting and faster decisions on pricing and capacity drove that growth. Receivables is a forecasting problem too: when will this customer pay, and what should we do today because of it?

The common thread: AI did not create the result alone. It made the right decisions faster and more consistent.

Common Mistakes to Avoid

• Automating a broken process. If your credit policy is inconsistent, AI will apply the inconsistency faster.
• Skipping the baseline. Without before numbers, you cannot defend the investment in year two.
• Chasing 100% automation. The last 10% of matches cost more than the first 90%. Accept that people handle the hard cases.
• Ignoring the customer experience. Aggressive automated dunning to a strategic account can cost you a relationship worth far more than the invoice.
• No owner after go-live. Models drift as customers, products and terms change. Someone has to watch.
• Treating AR in isolation. Receivables connects to billing, credit, sales and treasury. If you are also modernizing payables, read the accounts payable automation guide so both sides of the cash cycle use consistent data.

How AI in AR Fits Your Broader Finance Strategy

Receivables is rarely the only finance process on the AI agenda. It sits inside a bigger picture that includes close, planning, payables and reporting. The AI for accounting guide covers that wider landscape, and the AI in FP&A guide shows how better receivables data improves planning.

My recommendation for sequencing:

1. Start with receivables or payables, whichever has higher volume and clearer pain. Both have fast feedback and measurable cash impact.
2. Use the data to improve cash forecasting. Invoice-level payment predictions are the best input a forecast can have.
3. Move to planning and analysis once the transactional data is trustworthy.

This sequence builds credibility. A finance team that delivers a measurable DSO improvement in one quarter gets the budget for the next project.

If you are planning that sequence now and want a structured roadmap built around your systems and your team, you can request a consultation through the consulting request page. We start from your numbers, not from a vendor deck.

FAQ

What are the main AI in accounts receivable use cases?

The main AI in accounts receivable use cases are cash application, collections prioritization, payment date prediction for cash forecasting, dispute and deduction management, continuous credit risk monitoring, drafting collection communications and fraud detection. Most companies get the fastest payback from cash application and collections prioritization, because they cut manual work and focus effort on the invoices most likely to be paid late.

How much can AI reduce DSO?

There is no reliable universal figure, and you should be skeptical of anyone who quotes one without seeing your data. The reduction depends on your starting DSO, customer mix, credit policy and how consistently your team acts on the model's output. The right approach is to measure your baseline, run a pilot on a segment for 60 days and compare. Each day of DSO removed releases one day of credit sales in cash.

How long does it take to implement AI in accounts receivable?

A focused pilot on one use case typically fits within 90 days: about 30 days to measure the baseline and clean master data, 30 days to run the pilot side by side with your team, and 30 days to evaluate and scale. Full rollout across several use cases and business units takes longer, mostly because of ERP integration and data cleanup rather than the AI itself.

Will AI replace accounts receivable staff?

In most companies AI changes AR roles rather than eliminating them. Cash application specialists shift to resolving exceptions, collectors focus on accounts that need a real conversation, and credit managers move to continuous monitoring. Research on AI assistants in customer support found the biggest productivity gains among less experienced staff, which suggests AI raises the floor of team performance rather than removing the team.

Is AI-generated collection email legal?

Generally yes, as long as the content complies with the rules that apply to the debt. In the United States, consumer debt collection is governed by the Fair Debt Collection Practices Act and Regulation F, while pure business-to-business trade receivables generally fall outside that law. A person should approve every message template, and disputed or high-value accounts should always get human review before sending.

Do I need a specialist AR platform or can my ERP do it?

It depends on volume and complexity. Companies on a modern cloud ERP with moderate invoice volume can often start with the ERP's built-in AI features. High-volume businesses with many remittance channels, frequent deductions or complex credit decisions usually get more from a specialist platform. Ask both options for a pilot on your own historical data before deciding.

What data do I need to start with AI in AR?

You need clean customer master data with correct payment terms, at least 24 months of invoice and payment history, and dispute records with consistent reason codes. Remittance data from emails and portals helps cash application. Data quality matters more than data volume: one customer recorded under several IDs will degrade every model you deploy.