Cutting 40% of AR admin with AI-assisted workflows

Days sales outstanding is a core metric for any finance team. The longer it takes to collect on an invoice, the worse the company's cash flow. The collections team's job is to close that gap: contact the right accounts, at the right time, and follow up consistently until payment is received.

The business problem was simple to state. The team was not keeping up. DSO was higher than it should have been, and the root cause was not that analysts were working slowly. It was that the tools they were working with were adding overhead at every step of the process.

AR analysts are not casual software users. They work through a structured daily worklist, handling dozens of accounts per day. Every minute spent on admin is a minute not spent on a collection call. They are measured on outcomes: how much debt was recovered, how many follow-ups were completed on schedule, how quickly DSO came down.

They were not asking for new features. They were asking for less friction between the thing they needed to do and being able to do it.

The existing CRM and accounting system integrations were fixed scope. The redesign had to work with the data those systems provided, not require a new data pipeline.

Analysts were resistant to tools that felt like monitoring rather than support. Any AI-generated content, whether a suggested call script or an automated summary, had to be editable and clearly presented as a starting point, not a directive.

The workflow had to support both high-volume low-value accounts, where speed matters most, and high-value strategic accounts, where context and relationship history matter more. The same interface had to serve both without being optimised only for one.

Call recording and automated transcription raised compliance considerations that determined what could be automated and what required explicit analyst review before saving.

I spent time observing analysts working through their daily worklist before designing anything. What I saw was not a tool problem in isolation. It was a workflow problem expressed through tools: a browser tab for the CRM, a separate dialer, a spreadsheet for notes, and a calendar for follow-up reminders, all open at once.

I ran structured interviews with analysts across different account tiers to understand where the overhead was concentrated. Two themes surfaced consistently. First, the period between ending a call and logging the outcome was where context was most frequently lost, because analysts often moved to the next account before completing the write-up. Second, follow-up reminders set outside the workflow tool were regularly missed, because they had no connection to the account context when they fired.

I also spoke with team managers to understand how they monitored analyst performance and where they saw the most variability in outcomes. Inconsistent call logging was a recurring concern: different analysts recorded different levels of detail, which made account history unreliable for anyone picking up an account later.

I went into the project expecting the core problem to be the calling experience: analysts were using a separate dialer, which felt like the obvious friction point to fix.

The research showed the call itself was not where things went wrong. Analysts were comfortable making calls. The breakdown happened in the ten minutes after: writing up what was discussed, deciding what the next step was, and setting a reminder that would actually surface at the right time in the right context.

That shifted the design priority from improving the call experience to improving everything that surrounds it. The call is the event. The prep before and the logging after are where the workflow was actually failing.

I assumed at the start that the calling experience was the core problem. Analysts were using a separate dialer, which seemed like the obvious friction point. I expected the solution to be primarily about integrating telephony.

Observing analysts working showed something different. The call itself was the part of the workflow they were most comfortable with. The breakdown happened before and after: the scramble to pull up account context before dialling, and the manual write-up and reminder-setting that happened after hanging up.

That shifted my focus from the call to everything surrounding it. Integrating telephony was still important, but it was one piece of a larger problem about context continuity across a three-stage workflow.

The pattern that came out of this project, a continuous workflow that carries context through preparation, action, and follow-up, is not specific to AR collections. It applies to any role where someone works through a structured list of accounts or cases, takes an action, and needs to record and schedule a next step.

The AI summary approach is also reusable. The same principle of generating a structured first draft and letting the user confirm rather than create applies to any workflow where note-taking is currently a manual bottleneck.

The most important decision in this project was to observe before designing. I could have assumed the calling friction was the problem, built a solution around it, and shipped something that improved one step without addressing the workflow as a whole. The observation sessions prevented that.

The hardest part of the design was making AI-generated summaries feel trustworthy without feeling authoritative. Analysts needed to see them as a draft to edit, not a record to accept. Getting that framing right in the interface, in the language, in the editability of the content, took more iteration than I expected.

If I were doing this again, I would spend more time with team managers at the start of the project. Their perspective on where variability was costing the business most would have helped me prioritise the logging consistency problem earlier, rather than discovering it through the analyst interviews.