Every CFO knows the receivables age. Most CFOs cannot do much about it. The collections team is small. The customers know exactly when to delay. The bank reconciliation lags by days. By the time the team is chasing the right invoices, the next month's billing is already on the way.
Collections is the rare AI pattern where the payback is fast, the data is right there, and the team change is small. We have built this many times in Indian distributor businesses, SaaS companies, consumer brands, and services firms. The pattern is the same. The wins are the same. The places it stalls are the same.
What collections looks like today, in most businesses
Monday morning, an ageing report is exported from the accounting system. The accounts receivable team looks at the top 30 to 50 names. They call or message a handful. The customer says "by Friday". Friday comes, some pay, most do not. The team chases the same names the next week. The reconciliation happens once the bank statement is downloaded, usually weekly. Disputes get logged in a separate spreadsheet. The recovery rate is roughly what it was last year.
This is not a failure of the team. It is the unavoidable rhythm of small AR teams with large customer bases. The work is sequential, the volumes are too high for individual attention, and the team is doing the best a human team can do.
What the agent does instead
The change is qualitative, not just quantitative.
The agent reads ageing daily. It segments the receivable into who is likely to pay if nudged, who is genuinely stuck, who is signalling a dispute, and who is a quiet write-off risk. For the first bucket, it sends personalised nudges on the channel each customer prefers, at the time of day each customer is most likely to respond. It then watches the bank statement parse daily and reconciles payments to invoices, including the messy one-to-many and many-to-one cases.
The human team sees the disputes, the genuinely stuck cases, and the write-off risks. The team does not see the eighty percent of the AR they used to have to chase manually.
The agent's loop, step by step
Each step matters.
Read the ageing daily. Most teams pull it weekly. The cost of weekly is that two days of payment behaviour is invisible. With daily ageing, the agent sees a partial payment land at noon and adjusts the same day's nudge plan.
Decide who to nudge. The model learns from history which customers respond to which channel, which tone, which day of the week. Some customers always pay on Thursday. Some respond to WhatsApp but never reply to email. The agent learns and exploits.
Send the nudge. WhatsApp messages are personalised: the name, the invoice number, the amount, the option to pay through a link. Email goes to customers who use email. Phone calls escalate to the team for high-value or stuck cases.
Reconcile. The bank statement is parsed daily. Each credit is matched to invoices, including the cases where the customer paid three invoices at once or paid one invoice in two parts. Where the matching is ambiguous, the agent flags it for a human to confirm. Once confirmed, the rule gets learned for next time.
Typical impact, first six months
The DSO improvement is the headline. The team time freed is the unglamorous bigger win, because that team can now do the work that needs human judgement: large-account credit calls, dispute resolution, exceptions. The reconciliation speed-up is the third win that compounds, because finance closes the books faster, the forecast updates faster, and the working capital position is real-time rather than weekly.
Where most builds stall
Three reasons we have seen in failed or stalled builds. All three are fixable. They have to be designed in upfront.
The accounting system integration is messy. Tally, Zoho Books, SAP, NetSuite, all expose data differently. The team underestimates the integration effort and the agent ends up working off stale data. Fix: spend week one on the integration, do not move on until the agent is reading live ageing reliably.
The customer base is not message-receptive. Some industries respond to WhatsApp nudges. Some industries (large pharma, large institutional buyers) need a human voice. The agent does not improve those cases. It is fine to scope the agent to the customers it will help, and leave the rest to the team.
The reconciliation rules are unique. Every business has weird payment conventions. Bharat Electronics deducts a specific amount for some reason. A particular distributor always rounds down. The agent has to be trained on these conventions, and the training has to keep happening as new exceptions show up. This is not a one-week project, it is an ongoing one. Treat the rules as a living asset.
How to start, smallest viable version
- Pick one customer segment, ideally one with high volume and standard payment behaviour. SMB distributors are usually a good place to start. Large enterprise is usually the wrong place to start.
- Build the integration to your accounting system. Get the ageing flowing daily.
- Start with WhatsApp nudges only. Email and voice later.
- Reconcile bank statements daily, but keep a human review for ambiguous matches for the first month. Use that month to train the rules.
- Track DSO weekly. The number should start moving by week three. If it has not moved by week eight, something is wrong with the build.
What to do this week
Look at your last twelve months of DSO. If it is over fifty days and your team is small, you have a real candidate for this build. If it is under thirty days, your team is already doing brilliant work and the gains will be smaller.
If you want help scoping the smallest viable version and shipping in six weeks, talk to us. We have done this many times. The mistakes are foreseeable and the wins are real.