Does Channel Sequencing Matter?
What calls, texts and emails do, and in which order. Contact cadences are widely sold, but the evidence behind them is rarely published. This preliminary research note follows every call, text and email on the accounts of the firms that opted in, and asks which order and mix of channels goes with more engagement and more recovery.
What a text or an email does depends on the event that triggers it.
A link texted the moment a call is answered. An email sent after every call that goes unanswered. An inbound line ready for the callback. Each works because of when it fires, and each is only possible when calls, texts and emails run on one timeline.
Every collections team has to decide how to order its channels. Open with a text or a call? When to add an email? How many attempts before moving on? And does a payment link sent while the consumer is still on the phone help the conversation or distract from it?
This note answers those questions with data. It is a longitudinal study, meaning it follows accounts over time, of real engagements run by financial services firms on Krew’s platform. Only the customers who opted into the study are included. For each of their accounts, every call the voice agent placed, every text delivered and every email sent is lined up into one timeline. Each firm chose its own sequences on the platform’s workflow layer; Krew set none of them. The question is what those choices had to do with engagement and recovery.
The timeline was captured and scored with Observer, which records every call, text and email and grades each one against the same scorecards. The comparisons were run with Progenitor, which reports what is actually moving outcomes and turns the winning sequence into workflow rules. Every figure below comes from that one timeline.
Results are given as lift: a lift of 3× means three times the rate of the comparison group. Where a figure has been adjusted to hold other things equal, the note says so.
First, texting the payment link the moment the voice agent’s call is answered roughly doubles promises to pay on that call. The raw lift is 3.2×. Holding firm, month, script, connect number and balance equal, the adjusted figure is 1.6 to 1.9 depending on the model, and that is the figure this note relies on. Replies by text within a week rise fifteen-fold. More consumers end up on a plan. The call runs no longer, and callbacks do not rise. Because the text fires before anyone has spoken, the result cannot depend on how the conversation went.
Second, an email sent after every unanswered call is the sequence that reaches consumers who never speak to anyone. Accounts emailed after an unanswered call go onto a payment plan at 5.5 times the rate of accounts worked by calls alone, and at two to three times the rate within every balance band. The gap holds when the comparison is limited to consumers who had an email address on file. The email works without a conversation.
Third, the channels add up. Accounts that received calls, a text and an email in their first 30 days engaged at nearly three times the rate of calls alone and went onto a plan at more than six times the rate. The mix, not any one channel, produced the largest recovery lift in the study.
Fourth, promises come early. Four promises to pay in five arrive within the first two dials, and each dial after the first yields less than half as many promises. Effort moved from the sixth dial to the link, the email and the callback produces promises rather than just more attempts.
Reminders work, friction decides, and every extra touch has a cost
1. A reminder moves repayment, for the price of a message
Cadena and Schoar (2011) randomly gave small-loan borrowers one of three things: a text reminder, a cash reward, or a lower future interest rate. All three raised on-time payment by about the same amount. The text worked as well as a reward worth roughly a quarter of the interest cost. Campbell, Grant and Thorp (2022) found that a single digital reminder to overdue card accounts lifted repayment by roughly 4% compared with accounts that got none, and the effect lasted more than a year. Karlan, Morten and Zinman (2015) found that a reminder worked when it mentioned a relationship the borrower already had, cutting late payments by a quarter. Which day it arrived before the due date made no difference at all.
A text sent right after a call the consumer just answered is the closest a workflow can get to that reminder.
2. Friction at the point of action decides whether a decision becomes a payment
Bhargava and Manoli (2015), in a field study with the IRS, found that making the notice simpler raised take-up by roughly two-thirds compared with the complex version. That was a bigger effect than the reminder itself. Sunstein (2019) gives the general problem a name: sludge, the forms, waiting and navigation that stop willing people from finishing an action.
The distance between “agreed to pay on the phone” and “paid” is friction. A link in the consumer’s hand removes most of it.
3. Each additional message has a cost, so every touch has to earn its place
Every message uses up some of the consumer’s willingness to be contacted, and the telephone channel has a regulatory ceiling. Under Regulation F, more than seven call attempts in seven days on an account is presumed to be harassment (12 CFR § 1006.14(b)).
So a sequence should be judged on what each extra touch adds, not on how many people it reaches in total. That is the standard applied below.
Reminders work, friction decides, and every extra touch has a cost
The published results this note compares its own findings with.
An SMS reminder raised on-time payment as much as a cash reward worth about a quarter of the interest cost.
Cadena and Schoar (2011)
One digital reminder to overdue card accounts lifted repayment by roughly 4%, and the effect persisted for more than a year.
Campbell, Grant and Thorp (2022)
Simplifying the notice raised take-up by roughly two-thirds, more than the reminder itself. Friction at the point of action decides.
Bhargava and Manoli (2015)
Regulation F presumes no more than seven call attempts in seven days, which caps the telephone channel outright.
12 CFR § 1006.14(b)
Texting the link as the call is answered roughly doubles promises to pay
Across the participating firms, almost every account starts with a call from the voice agent. Texts and emails come after calls, not before them. The follow-up that matters most is the one that arrives while the call is still open. It happens in two ways. On-answer: for consumers the firm is permitted to text, the link is texted within seconds of the call being answered, before anyone has spoken. This is the large majority of cases. Consumers marked do-not-text receive no text. Mid-conversation: the voice agent asks “in case this call disconnects, can I call or text you back at this number?” and, if the consumer says yes, texts the account details and link about halfway through the call.
The comparison covers calls where a person spoke, February to August 2026, in three groups: calls with an on-answer text, calls with a mid-conversation text, and calls with no text. To keep the comparison fair, calls are matched on firm, month, script, connect number and do-not-text status before the groups are compared (the Mantel–Haenszel method), so consumers who cannot be texted are never compared with those who can, and a regression model is run as a second check. Because the on-answer text fires before the conversation, its effect cannot depend on how the conversation went. The mid-conversation text is reported separately, because the consent question that triggers it reaches a different group of consumers. Its lift in promises is 10.2× raw and 5.22 adjusted.
The second check agrees with the first. Holding firm, month, script, connect number, do-not-text flag and balance band equal, the on-answer text carries an odds ratio of 1.94 (95% CI [1.37, 2.76]) on promise to pay, 44 (95% CI [33, 60]) on a text reply, and 1.28 (95% CI [1.10, 1.50]) on being on a plan. (An odds ratio of 1.94 means the odds of a promise are about twice as high; the 95% CI is the range the true value is likely to fall in.) Inbound calls do not rise on either model (odds ratio 0.80, 95% CI [0.70, 0.93]): the link takes the place of the callback rather than adding to it. The adjusted figures are the ones to quote.
The same pattern shows up within a single consumer. Among consumers who had spoken calls both with and without a text, the promise rate on texted calls was 2.3 times the rate on untexted ones. More than twice as many consumers did better with the text as did worse (sign test p = 0.03), and replies ran more than twenty to one.
The text does not change the conversation. Spoken calls run the same length with or without the on-answer text, so the conversation is neither cut short nor dragged out. The consumer is simply given somewhere to go afterwards. Limited to calls that reached the account holder, the two kinds of text together carry 2.1 times the promise rate, 2.1 times the plan rate and 15 times the reply rate of untexted calls. Promises made on calls with a text went on to a plan 1.3 times as often as promises made without one, so the link does not weaken the commitment it comes with.
For a firm weighing this, the evidence points to the answer event as the trigger. A link that fires when the call is picked up, rather than after a question the agent asks, reaches every consumer who answers, costs nothing in call length, and in this study went with roughly twice the promises on that call.
Texting the link as the call is answered lifts every outcome but the callback
On-answer text against spoken calls with no text. The adjusted figure holds firm, month, script, connect number and do-not-text flag equal.
Promise to pay on the call
1.6 to 1.9
adjusted odds ratio, the figure this note relies on
Replied by SMS within 7 days
15.5×
beyond the axis; adjusted odds ratio 15.75
Call then email is the sequence that reaches consumers who never speak to anyone
Call then email is the most common two-channel sequence among the participating firms. The workflow queues the email the same day as the first dial, and about one newly worked account in six receives one in its first 30 days. Because the email is what happens after a call goes unanswered, the calls-plus-email group is mostly made up of accounts that never connected. Its engagement rate is lower than calls alone for that reason. Its recovery rate is not.
Recovery is measured three ways, each within a fixed window: a promise to pay on a call within 30 days, the account moving onto a payment plan within 45 days, and the balance falling by more than a set floor within 45 days. Plan and balance data are reported by the firm, so they are read as direction and expressed as lift.
The email lift holds in every balance band. Below $250, calls plus email moves accounts onto a plan at 2.9 times the calls-only rate; between $250 and $999, at 2.0 times; at $1,000 and above, at 2.2 times. The overall 5.5× is bigger than any single band because emailed accounts are concentrated in the small-balance band, where plan rates are highest. The within-band figures of two to three times are the safer estimate. Balance drops run roughly six to nearly thirty times higher on the emailed accounts, depending on the band.
It is not just about who gets emailed. Emails only go to consumers with an address on file. Among consumers who had an address, the ones the workflow emailed went onto a plan at 3.6 times the rate of the ones it did not (95% CI [3.2, 4.0]), showed balance drops at 5 times the rate (95% CI [3.9, 6.5]), and spoke with the voice agent less than half as often. As with the balance bands, this is the safer reading of the overall figure.
The email works without a conversation, which is the point. It is the touch that recovers the consumer the telephone never reaches. An email address on file that the workflow never uses is the largest missed opportunity in this study.
Mixes with SMS engage most; mixes with email recover most
Channel mix in the first 30 days. Lift relative to accounts worked by calls only; bars show the 95% CI where reported.
Each channel does something the others cannot, and the text works best at the first connect
Accounts that received all three channels in their first 30 days engaged at nearly three times the rate of calls alone, had conversations at three times the rate, and went onto a plan at more than six times the rate, with the largest balance-drop lift in the study. The email lift holds in every balance-band and address-on-file cut. The text lift holds after adjustment and within the same consumer.
Each channel does something the others cannot. The call produces the promise. The link makes the promise easy to act on. The email reaches the consumer who never answered. The text is the channel a consumer replies to.
When the text lands matters as much as whether it lands. This is not the calendar timing Karlan et al. (2015) found made no difference, a day earlier or later before a due date. It is timing relative to a live contact. A text delivered at the first connect goes with seven times the conversation rate and 3.6 times the plan rate of accounts with no text. A text delivered before any connect, usually the on-answer text landing on voicemail or a text sent after a no-answer, goes with 2.4 times the plan rate and plenty of replies and callbacks, but not with later conversations.
A text at the first connect goes with the most conversations and the most plans
SMS timing relative to the first connect, first 30 days. Lift relative to accounts with no SMS. Before any connect is mostly the on-answer text landing on a mailbox.
Callbacks follow every touch; the text is the one that brings a promise with it
Consumers call back within seven days of any outbound touch at broadly similar rates, mostly the same day and mostly from caller ID. The callback comes from the missed call, not from the message. The SMS row in the chart pools every delivered text, most of which follow a missed call. That is why it shows more callbacks than a connected call, even though the on-answer text on a spoken call adds none.
The one channel that changes what the callback leads to is SMS. A callback after a text arrives with a promise nearly three times as often as a callback after a connected call, and seven times as often as one after a no-answer.
The difference is what the consumer is holding when they dial. After a missed call, they are returning a call without knowing what it was about. After a text, they have the account details and the payment link on their screen, so they call back knowing who it is, what it is about, and what they can do about it, and the call more often ends with a promise.
Callbacks follow every touch; the text is the one that arrives with a promise
Lift relative to a connected call. A callback is an inbound call within 7 days of the touch.
Four promises in five arrive within two dials
This looks at outbound dials from February to August 2026 across the participating firms, counting the dial number per account. The share of dials that connect, and that turn into a conversation, barely changes with dial number. The share that produce a promise falls by more than half at the second dial, and by more than two-thirds from the third dial on.
The promise rate per conversation is highest on the first conversation with an account, lower on the second, and lower again on the third or later. A second conversation rarely converts what the first did not. Sixth-or-later dials are about a sixth of the volume of first dials and together produce 3% of promises.
The point is not to make fewer calls for their own sake. It is that most of a book’s promises come in its first two dials, and the touches that add most after that are the ones this note describes: the link on the answered call, the email behind the unanswered one, and the callback a text brings in.
Under Regulation F’s seven-in-seven presumption, fewer dials per account also keeps every account further from the line.
Four promises in five arrive within two dials
Dial number counted per account, across the participating firms.
Judge every touch on what it adds
The sequence the evidence supports is short: a call from the voice agent, a link texted the moment the call is answered, an email sent after every call that goes unanswered, and an inbound line ready for the callback. Each step is triggered by an event, which means calls, texts and emails have to run on one timeline.
On Krew, Progenitor turns these rules into workflow triggers, tests each version against live traffic under the firm’s policy, and keeps what moves the outcome. Observer scores every resulting call, text and email, so the sequence is measured on the book it actually runs on.
- 01
Text the link the moment the call is answered
After adjustment, promises roughly double, replies rise fifteen-fold, and the call runs no longer. The trigger is the answer itself, so the link reaches every consumer who picks up. This is a workflow rule, not a script change: the call and the text have to share a timeline so that one event can fire the other.
- 02
Send an email after every unanswered call
Call then email is the one sequence that reaches consumers the voice agent never speaks to: two to three times the plan rate of calls alone in every balance band, 5.5 times overall, and 3.6 times among consumers who had an address on file. The email should be automatic, same-day, and depend on how the call ended.
- 03
Run the channels together, not side by side
The largest recovery lift in the study belongs to accounts that received calls, a text and an email in their first 30 days. The channels add up because each does something the others cannot, and only when the workflow knows what the other channels did: a text timed to the first connect, an email that fires because the call did not connect, a callback that lands on an account the consumer already has in hand.
- 04
Move effort from the sixth dial to the triggers
With four promises in five arriving within two dials, each extra dial adds a connection on a shrinking base. The same contact budget spent on the link, the email and the inbound line adds promises, and keeps every account well inside the seven-in-seven presumption.
- 05
Measure the callback, and staff for it
Callbacks follow every touch at a similar rate. The text is the one that makes them arrive with a promise. An inbound line that answers with the account already identified, and a workflow that credits the callback to the touch that caused it, turns the missed call into a recovery channel.
- 06
Confirm with a link, not another reminder
Each further message uses up some of the consumer’s willingness to be contacted. In this study a link on the call does not weaken the promise it comes with. The link is the confirmation step; a second reminder is not.
- 07
Analyze within firm and balance band, and expect the answer to move
Participating firms differ in their books and workflows, and pooled figures carry those differences. The balance-band and address-on-file cuts are the fairer comparisons here. Any deployment should redo them on its own book, which only takes having the touch timeline in one place.
What the study covers, and how to read the numbers
This is a study of real engagements, followed over time, on the subset of customers who opted in. The results describe that subset, not every firm on the platform. Each firm chose its own sequences; none was assigned, and Krew set none of them. So every comparison describes what happened to the accounts that received each sequence. The balance-band and address-on-file cuts remove the most obvious differences between those groups. A firm that wants to know the effect of assigning a sequence should run a randomized rollout, which is the natural next step.
The on-answer result is adjusted and cross-checked. Firm, month, script, connect number and do-not-text flag are held equal in both the matched comparison and the regression, and the within-consumer comparison points the same way. An odds ratio of 1.9 (95% CI [1.4, 2.8]) is the firmest number in this note.
The sequence groups partly define themselves. Calls plus SMS is mostly the accounts that connected; calls plus email is mostly the accounts that did not. Read the charts with that in mind. The balance-band and address-on-file cuts are the fairer comparisons.
Recovery is read as direction, expressed as lift. Plan flags and balance changes are reported by the firm, and some accounts go onto a plan through the firm’s own channels. A promise to pay on the call is the one outcome the platform records at the moment it happens. Sequences that start with a text, or use texts alone, were not run by the participating firms and are not covered here.
Source. Krew production database as of 31 August 2026. The timeline was captured and scored by Observer and the comparisons run with Progenitor. It is built per account from calls (dial time from the status audit log, or else the end-of-call time minus the call length), delivered texts and sent emails. Internal test accounts are excluded. All figures pool the firms that opted in, and each firm set its own sequences. The data is aggregated and anonymised, identifies no individual consumer or firm, and is used with the participating firms’ agreement.
Fresh cohort. Accounts whose first outbound touch fell between 15 June and 17 July 2026, so that every account has a full 30 days of outreach and 45 days of outcomes inside the period when balance and plan changes are time-stamped. Mid-call text. A delivered outbound text to the same consumer, created between the start and end of the call. “On-answer” means a text delivered in the first seconds after the call starts. Calls excluded, as in the earlier report: voicemail, call screeners, wrong parties, transfers, non-English calls, and calls too short for a conversation.
- Engaged means a conversation with a person (the consumer spoke, not a mailbox or screener), an inbound call, or a reply by text or email.
- Promise to pay means a payment agreement record, a payment-commitment tag in the call memory, or a card or bank detail captured on the call.
- On plan means the firm’s payment-plan flag switched on within the window. Balance drop means a fall past a floor set to ignore cents-level import adjustments, within the window. Dial number is counted per account.
- Lifts are rate ratios against the named comparison group. Their approximate 95% CIs come from the 95% Wilson intervals of the underlying rates.
- Odds ratios are Mantel–Haenszel over firm, month, script, connect number and do-not-text flag, or logistic regression with HC1 errors and balance band added. Within-consumer comparisons use a sign test on consumers with both kinds of spoken call.
A fresh cohort with a complete window
Accounts whose first outbound touch fell inside the window, so every account has 30 days of outreach and 45 days of outcomes.
Prepared by
Michael Goh
Cofounder, Member of Technical Staff
Formerly led speech benchmarking at Artificial Analysis. Previously a FinTech VC and Bain consultant. MS Computer Science, University of Chicago; BA Economics, University of Oxford.
Rafael Khaykin
Cofounder, Member of Technical Staff
Formerly a research engineer at Figure Robotics, working on dexterity algorithms and latency research. Computer Science, University of British Columbia.
Alice Zhang
Head of Product Compliance, Member of Technical Staff
Formerly advised the United Nations on AI Ethics and Safety for defense industries. Previously In-House Counsel at a FinTech VC. BA (LLB) Jurisprudence, University of Oxford.
- Bhargava, S., & Manoli, D. (2015). Psychological frictions and the incomplete take-up of social benefits: Evidence from an IRS field experiment. American Economic Review, 105(11), 3489–3529. https://doi.org/10.1257/aer.20121493
- Cadena, X., & Schoar, A. (2011). Remembering to pay? Reminders vs. financial incentives for loan payments (NBER Working Paper No. 17020). National Bureau of Economic Research. https://doi.org/10.3386/w17020
- Campbell, D., Grant, A., & Thorp, S. (2022). Reducing credit card delinquency using repayment reminders. Journal of Banking & Finance, 142, 106549. https://doi.org/10.1016/j.jbankfin.2022.106549
- Karlan, D., Morten, M., & Zinman, J. (2015). A personal touch in text messaging can improve microloan repayment. Behavioral Science & Policy, 1(2), 25–31. https://doi.org/10.1177/237946151500100204
- Sunstein, C. R. (2019). Sludge and ordeals. Duke Law Journal, 68(8), 1843–1883.
This is a preliminary research note, not a peer-reviewed study. It reports observations from historical engagements on an opt-in subset of the platform, and its estimates carry the uncertainty described in the Scope section. Every comparison is observational; no sequence was assigned. The results are specific to the firms, portfolios, workflows, channels, and period studied; they are not a prediction of what any other deployment will show, and actual results will vary by portfolio, configuration, and context. Nothing here is a guarantee of performance, and the analysis may be revised as more data is collected. All data underlying it is aggregated and anonymised, identifies no individual or firm, and is used with the participating firms’ agreement.
This document is provided for informational purposes only and does not constitute legal or regulatory advice. Motivated by Krew’s Collaborative Assurance Framework’s principles of shared accountability and streamlined assurance, we partner with our customers, combining our AI-driven credit servicing platform’s built-in security safeguards with each customer’s own controls, system configurations, and user training, to help manage regulatory risk. It remains the responsibility of each customer to ensure compliance with all applicable federal, state, and local statutes. All rights, responsibilities, and liabilities of Krew in relation to its customers are governed exclusively by the terms of Krew’s customer agreements. This document neither forms part of, nor alters, any contractual agreement between Krew and its customers.