AI tools for collections: a teardown of what actually works
An opinionated review of AI-driven collections tools and workflows for small-to-mid lenders — what genuinely helps, what's vendor theatre, and where DIY prompts beat the SaaS.
This post is part of our series on tool reviews and teardowns. Unlike most tool-review content on the internet, this one isn’t sponsored, isn’t an affiliate farm, and isn’t optimised for search volume. It’s what I’d tell an operator-friend honestly.
We’ll cover: the three categories of AI collections tools, honest teardowns of the category leaders, when DIY prompts beat the SaaS, and the one collections workflow that’s worth the $30 and the hour of setup.
The three categories
Every AI-collections tool sits in one of three buckets.
Category A — AI-augmented communications. Tools that draft emails, SMS, WhatsApp, and phone-call scripts, with tone calibration and cadence management. These are the most useful category for small-to-mid lenders and the easiest to DIY. Examples: CollectIt AI, DebtAI, various Zendesk/Intercom integrations.
Category B — AI agents that make outbound contact. Voice agents that call delinquent borrowers, negotiate, and accept payment arrangements. Examples: Skit, Gnani, Bland AI in its collections-specific deployment. Promising category. Regulatory land mines.
Category C — AI-driven portfolio intelligence. Tools that predict which delinquent accounts will self-cure, which need intervention, and which should escalate to legal. Examples: EkCount, Collectly’s forecasting module, most of the enterprise analytics stacks’ collections modules. Useful at scale; overkill for small lenders.
Let me walk each.
Category A — AI-augmented communications
What they do. Draft the email, pick the channel (email / SMS / WhatsApp / call), sequence the touchpoints, and adapt tone based on how many days overdue the borrower is and how they’ve responded to prior outreach.
The honest assessment. This category is genuinely useful, and also the category where you can replicate 80% of the value with $30 of prompts and a simple spreadsheet.
The reason: the core of what these tools do is (a) a prompt library for email/SMS drafts, and (b) a cadence rule like “day 3 email, day 7 SMS, day 14 email + call, day 30 final notice.” Both are replicable.
When the SaaS justifies its price:
- You’re running over 1,000 active loans and the ops overhead of manual cadence management matters.
- You need audit trails integrated with your LMS that show exactly what went to whom when.
- You want channel-mix optimisation (the SaaS will tell you “your SMS performs 20% better than your email for 15-30 DPD accounts; shift mix”).
- You need multi-language templating (regional Indian languages, Spanish, etc.) at scale.
When DIY wins:
- Under 1,000 active loans.
- You want to customise tone beyond what templates allow.
- You want the flexibility to not pay $500-$2,000/month for what’s essentially a structured prompt + a spreadsheet.
For a small lender or P2P pool, do it yourself with the prompts from the Lending Prompt Library (two-prompt combo: the early-stage collections email + the late-stage collections email) plus a simple cadence tracker. Category A SaaS is overkill until you’re comfortably above 1,000 active loans.
Category B — AI voice agents
What they do. Place outbound calls to delinquent borrowers. Conduct the conversation. Negotiate payment dates. Handle objections. Route escalations to humans.
The honest assessment. Technologically impressive, operationally risky, regulatorily fraught.
The technology is genuinely good — voice naturalness has crossed the threshold where most borrowers can’t immediately tell they’re speaking to AI. Conversion rates on payment-promise extraction are comparable to human agents for simple cases.
The problems:
-
Regulatory. India’s RBI Fair Practices Code requires collections contact to be within business hours, professional, and non-harassing. AI agents can technically comply, but the auditing of edge cases (the call where the agent said something problematic when the borrower became emotional) is much harder than with human agents. One viral WhatsApp clip of an AI collections call gone wrong and your brand is in trouble for a quarter.
-
Reputation. Borrowers, particularly in India’s retail lending market, increasingly recognise AI voice agents and increasingly resent them. A human collector — even a firm one — reads as respect. An AI agent reads as the lender not bothering to assign a human. This damages brand in ways that cost more than the efficiency gains in mid-tier markets.
-
Compliance review burden. You still need humans listening to call samples for compliance. At scale, this cancels a meaningful portion of the labour savings.
Who should use Category B: large lenders with a mature compliance-review pipeline, dedicated AI-voice vendor relationships, and comfort with the reputational risk. Not most readers of this post.
Who should skip: small lenders, P2P pools, anyone without a head of compliance. Use human collectors with AI-drafted scripts (Category A) instead.
Category C — Portfolio intelligence
What they do. Predict which delinquent accounts are likely to self-cure, which need active collection effort, and which should escalate to legal or write-off faster.
The honest assessment. Useful at scale. Overkill for small lenders. And frankly replicable with decent SQL plus one good prompt.
The underlying logic of most of these tools is a statistical model (gradient-boosted tree, typically) trained on your historical data, which outputs a “cure probability” per delinquent account. That’s genuinely useful — it lets you allocate human collection effort to accounts where it matters.
But:
- If you don’t have 10,000+ historical loans with outcomes, the model doesn’t have enough data to train.
- If your loan product has changed recently, the historical patterns aren’t reliable anyway.
- The Q&A prompt pattern (“given this delinquent account’s recent payment history and communication pattern, assess cure likelihood”) gets you 70% of the value from an LLM with zero training required.
Verdict. Overkill for small lenders; Category C only pays off at institutional scale where the account volume justifies the model maintenance and data engineering.
The specific workflow I’d build at small scale
For a small-to-mid lender or P2P pool (say, 50-1,000 active loans), here’s the collections workflow I’d build for ~$30 of tooling and a weekend of setup:
Tools:
- ChatGPT Plus / Claude Pro subscription (~$20/month — already yours).
- The Lending Prompt Library ($29, one-time) for the collections-section prompts.
- Google Sheets for cadence tracking.
- Whatever email/SMS tool you already use (Zoho Mail, ClickSend, or similar).
The workflow:
- DPD 1-7 (early). Automated email drafted by AI (early-stage collections prompt), sent day 3. SMS reminder day 7. Cost: 30 seconds of AI time per borrower, 5 minutes of review.
- DPD 8-30 (mid). Personalised email + phone call. AI drafts the email and the call script. Human makes the call. Phone script prompt pulls borrower-specific context.
- DPD 31-60 (late). Formal notice (AI-drafted, human-reviewed for compliance language). Direct outreach attempt. If no response, begin escalation prep.
- DPD 61+. Human-led; AI assists with the post-mortem at the 90-day mark to inform future underwriting.
Add a weekly Sheets view that tracks where each delinquent loan is in the cadence. Add a weekly portfolio-level AI alert (“any patterns in this week’s delinquencies?”). That’s the system. It won’t hit the sophistication of a $2,000/month SaaS, but it’ll outperform the 85% of small lenders who are doing collections by ad-hoc email and occasional calls.
The one thing every AI collections setup needs
Compliance review on every templated message. Once at setup, and again every 6 months or any time your jurisdiction’s rules change. AI-drafted content is only as compliant as the template it’s drafting against. Get this reviewed by a real fair-practices-aware legal person once; after that, you’re fine.
Where to go from here
If you want the specific prompts for the collections workflow above — early-stage email, late-stage email, phone-call script, negotiation framework, hardship-case decision, post-mortem — they’re all in the AI Lending Prompt Library under the Collections section.
Next in the tool-review series: teardowns of AI credit-memo generation tools and bureau-plus-AI aggregators for small-lender use. Shipping over the coming weeks on the blog.
If you have specific tools you’d like torn down next, email me with your shortlist.
Frequently asked questions
Do AI collections tools actually recover more money?
For the first 60 days of delinquency, AI-augmented collections reliably outperform manual workflows — mostly through better message timing, tone calibration, and consistent follow-up cadence. Beyond 90 days, the gains from AI plateau and the recovery is mostly about legal process, field visits, and negotiation — tasks AI supports but doesn't drive.
Can I replace my collections team with AI?
No — and you shouldn't try. AI takes on the drafting, the cadence management, and the pattern recognition. Humans take on the hard negotiations, the exception handling, and the compliance judgment. The teams that have tried full-AI collections have mostly rolled it back after regulatory or reputational incidents.
What's the single highest-ROI AI move in collections for a small lender?
A tone-calibrated email-drafting prompt combined with a simple cadence (day 3, day 7, day 14, day 30). Most small lenders either send harsh or generic emails, and either under-contact or over-contact. The prompt + cadence combo fixes both problems for ~$30 of tools and an hour of setup.
Are there regulatory issues with AI in collections?
In India, the RBI Fair Practices Code and the Digital Lending Guidelines require collections communications to be professional, non-harassing, and within business hours. AI-drafted content is fine if reviewed for these rules. In the US, the FDCPA and state-specific collections rules apply — AI-drafted debt-collection content needs compliance review, particularly around mandatory disclosures.