AI in Collections, Getting Started, Part 2: An Intake and Triage Workflow Your Compliance Team Can Review

In Part 1 of this series we sorted AI opportunities in a collection agency by one question: how close does the work get to the consumer? Zone 1, internal operations on non-consumer data, is where to start. Zone 3, consumer communications and higher-risk workflows, is not a getting-started project at all.

Part 2 picks one Zone 1 workflow and builds it out: intake and triage. What arrives, how it gets sorted, who sees it, and how you prove the whole thing is working. By the end you will have a design you can hand to your compliance lead on a single page.

This is general marketing information, not legal or compliance advice. Consult your attorney or compliance officer before acting.

Why Intake and Triage Comes First

Every agency has a pile. Client correspondence, placement paperwork, vendor invoices, regulatory bulletins, RFPs, internal requests. Someone opens each item, figures out what it is, decides whether anything is missing, and routes it to the right person. The work is not hard. It is constant, and it interrupts everything else.

This is not a collections problem so much as a modern work problem. Microsoft's 2025 Work Trend Index, published April 23, 2025, reported that 80 percent of the global workforce says they lack the time or energy to do their job, while 53 percent of leaders say productivity must increase. The same report describes employees being interrupted every two minutes by a meeting, email, or ping. Those figures cover knowledge workers broadly, not agencies specifically, but the shape will be familiar.

Intake and triage is a good first AI project for three reasons. It is repetitive, so a written rule set can describe most of it. It produces a draft, not a decision, so a person can check every output quickly. And a large share of it involves no consumer account data at all, which keeps your first pilot far from the regulatory line.

Draw the Boundary Before You Draw the Workflow

The single most important design decision is which lanes the tool is allowed to touch.

Start with lanes that contain no consumer information: client-relationship correspondence, vendor and billing documents, regulatory and industry news, RFPs and proposal requests, and internal requests from your own staff. That is a lot of volume, and none of it requires a conversation about consumer data handling.

Placement files and anything tied to individual accounts are a separate, later decision. Before any of that goes near an AI tool, your compliance, security, and legal teams need to confirm what is permitted, and your client contracts may have their own terms on vendors and data. Part 1 covered those guardrails, and nothing in this piece changes them.

Why so careful inside your own walls? The CFPB offered a useful reminder in June 2023 with an issue spotlight on chatbots in banking. That report was about consumer-facing tools, which is exactly why they sit in Zone 3, but its core observation applies to any AI deployment in a regulated business: a poorly deployed tool can lead to frustration, lost trust, and, in the Bureau's words, even violations of the law. The way to avoid that in Zone 1 is to keep the tool's job narrow and keep a named person on every output.

The Workflow, in Five Parts

Here is a design that fits most agencies. Adapt the categories to your operation.

1. Pick one lane. One shared inbox, or one document type. Not everything that arrives. A client-correspondence inbox is a common choice: busy, tied to revenue, and typically free of consumer account data.

2. Write the sorting brief in plain English. List your categories (for example: new business inquiry, existing client question, document request, invoice, industry news, something else). Define what counts as urgent. Define what "complete" looks like for each category, so the tool can flag what is missing. This brief is the heart of the project. It is also the document your compliance lead will read, so write it as if they are the audience.

3. Produce drafts, not actions. For each item, the tool outputs four things: a suggested category, a three-line summary, a list of anything missing, and a suggested queue or owner. It does not reply, forward, file, or move anything. A person reads the four lines and confirms, edits, or overrides. In the first weeks that person checks every item. If the phrase "AI agent" keeps coming up in vendor pitches, our plain-English explainer on what an AI agent actually is covers the difference between a tool that drafts and one that acts.

4. Build the exception path. Two rules. First, anything the tool is unsure about goes to a "needs a human" pile rather than a best guess. Second, if an item appears to contain consumer account information in a lane that is not supposed to have any, the tool stops, tags it, and routes it to the person your compliance lead names. That second rule is the one that will reassure compliance most.

5. Log everything. Every item, the tool's four outputs, the reviewer's decision, and the timestamp. This log is where your metrics come from and where your compliance documentation points.

Our guide to using AI tools for marketing work, and when not to covers the same drafting discipline from a different angle.

Four Numbers to Watch

Part 1 said to write down your baseline before you start. Here are the four numbers that tell you whether the workflow is earning its place.

1. →Time from arrival to routed. Measured in minutes or hours, compared with the baseline you recorded before the pilot.
2. →Sort accuracy. Each week, pull a sample of items and check whether the suggested category was right. Track the percentage. A simple spreadsheet is fine.
3. →Missing-item catch rate. Of the items that were actually incomplete, how many did the tool flag? Compare with how often your team caught the same gaps before.
4. →Override rate. How often the reviewer changes the tool's suggestion. A high override rate early is normal and tells you the brief needs sharpening. A rate that never comes down tells you the lane may be a poor fit.

Hold a fifteen-minute review each week to look at the four numbers and the "needs a human" pile. Most pilot improvement comes from rewriting the brief after those reviews, not from changing tools.

The One-Page Record Compliance Will Actually Read

Your compliance team already works from official material, including the CFPB's debt collection compliance resources, which hosts the small entity compliance guide for Regulation F, the rule FAQs, and examination materials. How Regulation F and UDAAP apply to your AI use is their call to make, not ours. What you can do is make that call easy by handing them one page that answers the questions they will ask anyway.

A useful structure comes from the National Institute of Standards and Technology. Its AI Risk Management Framework 1.0, released January 26, 2023, is voluntary and organizes AI risk work into four functions: Govern, Map, Measure, and Manage. You do not need to adopt the full framework to borrow its headings.

• →Govern. Who owns this workflow by name. Which tool is approved and on what business terms. Enterprise-tier plans, such as Claude Enterprise or another vendor's equivalent, generally give an agency the strongest controls here: administrator-managed users, written data-handling terms, and audit logs. Who reviews outputs. Who can shut it off.
• →Map. Which lane the tool touches. Which data it is never allowed to see. What happens when it sees that data anyway (your exception path).
• →Measure. The four numbers above, the sample size, and who records them.
• →Manage. What you do when the tool is wrong. How an item gets pulled back. How you retire the workflow if it stops earning its place.

Four headings, a dozen bullet points, one page. Walk your compliance lead through it before the pilot starts, not after.

What It Looks Like in Practice

Here is a hypothetical, invented for illustration and not drawn from any real agency.

A fourteen-person agency picks its client-correspondence inbox. The operations manager writes a one-page sorting brief with six categories. For four weeks, an approved enterprise-tier AI tool reads each incoming message and drafts the four outputs. The office coordinator reviews every draft before anything moves.

In week one the override rate is high, mostly because "existing client question" and "document request" overlap. The manager rewrites those two definitions. By week three the "needs a human" pile has shrunk, and the coordinator spends mornings on client follow-up instead of sorting. Twice during the month, a message arrives that references an individual account. Both times the tool stops, tags it, and routes it to the compliance lead, who confirms the exception path works as written.

At the week-four review, the team compares the four numbers against the baseline and decides whether to keep the lane, adjust it, or try a different one. Notice what never happened: the tool never sent anything, never touched a placement file, and never made a decision a person did not confirm.

What's Next in This Series

Part 3 moves to Zone 2 and covers using AI to market an agency: proposal libraries, thought leadership for creditor audiences, and clearer client reporting. Later parts cover writing an AI use policy people will follow and the questions to ask any AI vendor.

The Bottom Line

An intake and triage workflow is the right first AI project for most agencies because it is narrow, reviewable, and can be kept entirely away from consumer data. Pick one lane, write the brief, produce drafts only, build the exception path, log everything, and watch four numbers. Then put it on one page your compliance lead can read in a sitting. If you would rather not scope it alone, our explainer on what an AI consultancy actually does describes what outside help looks like at this stage.

Want help scoping a first workflow or the marketing that comes after it? Start a conversation with Yayah Creative Co.

Sources

• →Microsoft: The 2025 Annual Work Trend Index: The Frontier Firm is born (April 23, 2025)
• →Consumer Financial Protection Bureau: CFPB Issue Spotlight Analyzes "Artificial Intelligence" Chatbots in Banking (June 6, 2023)
• →Consumer Financial Protection Bureau: Debt Collection (FDCPA) compliance resources (page last modified March 12, 2025; accessed October 2026)
• →National Institute of Standards and Technology: AI Risk Management Framework (AI RMF 1.0 released January 26, 2023)

---

This is general marketing information, not legal or compliance advice. Consult your attorney or compliance officer before acting.

Statistics are cited from their original publishers as of October 2026. The agency example is hypothetical.