PAN Lab example

M-Shwari & Kenya's Digital Credit Market

Scored from a phone's records, listed where the market reads

A bank scores people who have no credit file at all, from six months of the records their phone company holds: airtime top-ups, low-balance days, money in and out. Modeled on Kenya's mobile-money credit market and its named anchor deployment. The score is set when the account opens, before anyone asks to borrow, and the customer sees only a limit — zero below a cutoff. Repayments land back in the same trace that scores the next limit. At 120 days of arrears, a pipeline writes a durable entry into a register every lender in the market reads, and it took three state instruments — a listing floor with a mass delisting, a licensing regime with prior approval over pricing, and data-collection guidance — to put a governed surface under it. The same evaluation that shows the score does not separate repayment at its own margin also shows real access gains for the households it reached. Before you pick a target level: with this record's own budget and levers, this board is winnable in Explore and under Service Targets Only; under Service and Safety Targets and All Governance Targets the measured result is that no legal stack of the offered levers closes it, at any spend — the one pathway nothing on offer reaches is the product team's own hand on the rules, which in the real record only the regulator's approval gate touches, from outside.

Stylized model of a documented deploymentLending & credit collections AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Thin-file-scoring-class writing into a register it does not own network: 11 components and 24 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 10 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

assumed
This models the thin-file alternative-data lending pattern documented in the case file — the named anchor's scoring, limit and listing mechanics plus the documented market layer — not a reconstruction of any one lender's model.
baseline
Attribution split, binding on every value here: the scoring frame, the hidden score set at account opening, the zero limit below the cutoff, the 30-day term, the 7.5% facilitation fee, the automatic rollover and the 120-day bureau report are the named anchor's documented mechanics (M-Shwari, evaluated by regression discontinuity). The phone-data extraction, the mean effective APR of 280.5%, the comprehension figures and the market default and listing rates are class-level findings about the Kenyan digital-credit market and app-based lenders, and are never attributed to the named anchor, which runs on the SIM toolkit from operator-held records.
baseline
Demand reads 3 from documented load: 7.5 million loans worth KSh 133.5 billion granted by the licensed segment alone as of February 2026, 227 licensed providers from 800-plus applications, 77% of mobile loan users unable to repay at least once, and roughly 90% of all bureau blacklistings attributed to digital credit. Capacity reads 1 because no per-application human exists anywhere in the pipeline — scoring precedes any application, escalation is time-driven, the product team works at portfolio level, and the redress machinery is documented as little used.
baseline
The scoring frame is drawn as an input source the deploying organisation does not own, because that is the record's central structural fact: the trace is the mobile network operator's, six months of it is the eligibility condition, and the lender governs a decision made from a store it cannot inspect record by record. The population outside the frame is not a set of rejected applicants — it is people who appear in no log this network reads, and no coverage figure exists anywhere in the record, so none is implied here.
assumed
The arrears pipeline is drawn as an enforcement component rather than a second model because the record describes a deterministic time-driven rule, not a learned system: 30 days, an automatic rollover with a further 7.5% fee, a bureau report at 120 days of non-payment. The PAN org models the same element as a second rule entity; the Lab draws it in the catalogue's enforcement idiom — automatic replication from the record with a reconciliation check that statute populated only in 2020 and 2022.
baseline
The closed loop is documented, not stylistic: disbursals and repayments are themselves mobile-money transactions landing back in the trace the score is computed from, and the limit grows on timely repayment — the deployment writes its own future inputs. The register-to-model read is drawn at the low rung as a documented blindness: non-bank digital lenders were barred from the credit information system from April 2020 until the 2022 regime made two-way sharing mandatory, and the market inquiry found lenders did not know an applicant's full borrowing.
baseline
The model check is drawn empty on a documented absence: the record documents no organisation-run re-read of the score against outcomes. The one published measurement at the margin — a first-loan default coefficient of 0.007 against a control mean of 0.066 just above versus just below the cutoff — is an outside academic evaluation's check on its own regression-discontinuity design, local to the cutoff, and is treated here as a bound on what the score separates where it decides, never as a global accuracy figure.
baseline
The service side of the same evaluation is real and carried in the case file, not asserted away: 34% of eligible households took a loan, eligible households were 6.3 percentage points less likely to forego expenses after a negative shock, and the loans did not substitute for other credit. Nothing on this diagram computes that benefit; the network models institutional propagation only.
baseline
Three reviewers are drawn because the record documents three distinct review functions with their own reads and their own instruments: a statutory redress desk that can dispute a register entry but can never reach the undisclosed score; a central bank whose 2020 instrument put a numeric floor under the listing write and whose 2022 regime made pricing parameters subject to prior written approval; and a market inquiry that audited provider account-level data and holds publish-and-recommend authority only. Each has a documented inbound pathway; none dangles.
baseline
The egress pathway and the recovery-side write are class-level: the data protection authority's December 2023 guidance documents app-based lenders taking phone contacts, call logs, pictures, location and messages and holding them as collateral, and providers sharing customer data with marketing companies without consent. These flows are drawn for the class this org models and are not attributed to the named anchor.
baseline
Era stamps travel with the numbers: the minimum loan of one hundred shillings is an evaluation-era and 2015-era figure — the named anchor's minimum was five hundred shillings by 2020 and has been two thousand since August 2020, with smaller borrowing redirected to an overdraft product. The 7.5% per-term facilitation fee is documented unchanged from 2012 through 2026. The 2022 licensing regime is in force and broadening: a 2024 statute recategorised digital credit as non-deposit-taking credit business, and draft successor regulations were exposed for comment in August 2025 — a draft, not a replacement.
assumed
Borrowers are boundary-only. No repayment, default, over-indebtedness, exclusion or household outcome is computed from anything drawn here; a credit limit, an arrears status and a bureau entry are institutional signals. The market's default and listing rates, the APR and comprehension figures, and the access rates by gender are recorded external observations living in the case file, and the digital-borrower cohort is documented as lower-income and more vulnerable than the formal cohort, so no cross-market comparison here licenses a causal reading.
baseline
Three things this network once drew as pathways of their own are carried in the description of a neighbouring element, and none of their facts has left this page. The first is the single rule set: one set of rules scores every eligible customer at account opening, so a mis-weighted proxy repeats across the whole book, which is what the one scoring element stands for and is stated on it. The second is recovery reading the register's listing status back when working an account; the record documents that read-back and no consequence of it, so it rides on the recovery-submission pathway it mirrors. The third is the scoring frame reaching the product team as modelling aggregates from a store the lender does not hold; it arrives at the same portfolio-level team, in the same aggregate form, as the model's portfolio figures, so it rides on that pathway. The source model this network is derived from draws the second and third as narrow reads of their own and has no counterpart to the first.

What this example does not show

No borrower outcome is modeled. The Lab reads institutional propagation only; borrowers are boundary-only, and the market's default and listing rates, the access rates by gender, the APR and comprehension figures, and the evaluation's resilience results live in the case file, never computed on this diagram. The digital-borrower cohort is documented as lower-income and more vulnerable than the formal-market cohort, so the cross-market comparisons license no causal reading.
The attribution split is load-bearing: the scoring, limit and listing mechanics and the regression-discontinuity results belong to the named anchor; the phone-data extraction, the APR and comprehension figures and the market default and listing rates are class-level findings about the Kenyan digital-credit market, and nothing here attributes them to the named anchor.
The cutoff result — a first-loan default coefficient of 0.007 against a control mean of 0.066 just above versus just below the score cutoff — is the evaluation authors' check on their own study design, local to the cutoff. It is the only published measurement in this family of what the score separates at the margin where it decides, and it is not a global accuracy figure; the diagram draws it as the latent model check, not as a computed error rate.
Era stamps: the one-hundred-shilling minimum loan is an evaluation-era figure — the anchor's minimum has been two thousand shillings since August 2020. The 2022 licensing regime is in force and broadening: a 2024 statute recategorised digital credit as non-deposit-taking credit business and draft successor regulations were exposed for comment in August 2025; they are a draft, not a replacement. Licensed-provider counts are the central bank's own (227 as of April 2026).
The share of the eligible population outside the scoring frame is measured nowhere in the record, and phone-ownership rates are not a substitute for the six-month active-use condition; this scenario implies no coverage figure.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

M-Shwari, launched November 2012 by the Commercial Bank of Africa (now NCBA) in partnership with Safaricom over the M-PESA rails, scores customers with fully automated credit approval rules run on six months of the mobile network operator's transaction and airtime record: a score assigned at account opening irrespective of when the customer borrows, never disclosed to the customer, expressed only as a credit limit that is zero below the cutoff and grows on timely repayment. Loans in the evaluated era started at KSh 100 (the minimum has been KSh 2,000 since August 2020, with smaller borrowing redirected to the Fuliza overdraft), run 30 days at a 7.5 percent facilitation fee unchanged from 2012 through 2026, roll over automatically with a further 7.5 percent, and are reported to a credit reference bureau after 120 days of non-payment.
empirical
- Academic Suri, T., Bharadwaj, P., & Jack, W. (2021). Fintech and household resilience to shocks: Evidence from digital loans in Kenya. Journal of Development Economics, 153, 102697 (read in the NBER Working Paper 25604 version, February 2019) https://doi.org/10.1016/j.jdeveco.2021.102697
- Reference Cook, T., & McKay, C. (2015, April). How M-Shwari Works: The Story So Far (Forum No. 10). CGAP and FSD Kenya https://www.cgap.org/sites/default/files/Forum-How-M-Shwari-Works-Apr-2015.pdf
- Trade press Cytonn Investments (2020, August 12). Safaricom Raises M-Shwari's Minimum Loan Limit to Ksh 2000 (news digest) https://cytonnreport.com/news/safaricom-raises-m-shwari-s-minimum-loan-limit-to-ksh-2000
A peer-reviewed regression-discontinuity evaluation of M-Shwari found that 34 percent of eligible households took a loan, eligible households were 6.3 percentage points less likely to forego expenses in response to a negative shock, and digital loans did not substitute for other credit; the same evaluation's administrative data show no significant difference in first-loan default between borrowers just above and just below the score cutoff — a coefficient of 0.007 against a control mean of 0.066 — reported by the authors as a check on their study design, local to the cutoff, and the only published measurement in this deployment family of what the score separates at the margin where it decides.
empirical
- Academic Suri, T., Bharadwaj, P., & Jack, W. (2021). Fintech and household resilience to shocks: Evidence from digital loans in Kenya. Journal of Development Economics, 153, 102697 (read in the NBER Working Paper 25604 version, February 2019) https://doi.org/10.1016/j.jdeveco.2021.102697
At market level — findings about the Kenyan digital-credit market and its app-based lender class, not about M-Shwari — nationally representative survey evidence (FinAccess 2019, borrower n = 2,194) puts digital-credit default at 13.8 percent against 6.4 percent formal, negative bureau listing at 5.3 percent of digital borrowers against 1.4 percent formal, 38.4 percent of digital defaults reported to a bureau against 22.7 percent of formal ones, and roughly 90 percent of all blacklistings attributed to digital credit, on a digital cohort the source documents as lower-income and more vulnerable than the formal cohort; the Competition Authority of Kenya's market inquiry with Innovations for Poverty Action (a 793-user phone survey plus an audit of provider transaction and account-level data) found a mean effective APR of 280.5 percent and median 96.5 percent, 77 percent of mobile loan users unable to repay at least once, 40 percent able to state their last loan's cost within five percent, and 27 percent aware of other providers' charges.
empirical
- Academic Johnen, C., Parlasca, M., & Musshoff, O. (2021). Promises and pitfalls of digital credit: Empirical evidence from Kenya. PLOS ONE, 16(7), e0255215 https://doi.org/10.1371/journal.pone.0255215
- Government Competition Authority of Kenya and Innovations for Poverty Action (Putman, D., Mazer, R., & Blackmon, W.) (2021, May). Report on the Competition Authority of Kenya Digital Credit Market Inquiry https://poverty-action.org/sites/default/files/2023-05/Digital_Credit_Market_Inquiry_Report_2021.pdf
The Central Bank of Kenya's Credit Reference Bureau Regulations of 8 April 2020 barred negative listings at or below one thousand shillings, forced a one-off mass delisting below that floor, and required the credit score in every credit report, while non-bank digital lenders were barred from the credit information system until licensing; the Digital Credit Providers Regulations of 18 March 2022 brought digital lenders under CBK licensing, made two-way credit information sharing mandatory with the KSh 1,000 floor carried forward, required pre- and post-listing notice, capped recovery on non-performing loans, required ability-to-repay assessment and a complaints mechanism with 30-day resolution, prohibited phone-book access and shaming in collection, and made product and pricing-parameter changes subject to prior written approval; the Office of the Data Protection Commissioner's December 2023 guidance documents app-class providers taking 'phone contacts, call logs, pictures, location, messages, all mobile phone applications and everything in the phone' stored 'as a collateral' and sharing customer data with marketing companies without consent, directing collection of transaction detail only. By 14 April 2026 the CBK had licensed 227 digital credit providers from more than 800 applications, with licensed providers having granted 7.5 million loans worth KSh 133.5 billion as of February 2026; the Business Laws (Amendment) Act 2024 recategorised digital credit as non-deposit-taking credit business, and draft successor regulations were exposed for comment in August 2025 — a draft, not a replacement.
empirical
- Government Central Bank of Kenya (2020, April 8). The Banking (Credit Reference Bureau) Regulations, 2020 (Legal Notice No. 55, Kenya Gazette Supplement No. 42) https://www.centralbank.go.ke/wp-content/uploads/2021/05/Credit-Reference-Bureau-Regulations-2020.pdf
- Government Central Bank of Kenya (2022, March 18). The Central Bank of Kenya (Digital Credit Providers) Regulations, 2022 (Legal Notice No. 46, Kenya Gazette Supplement No. 45) https://www.centralbank.go.ke/wp-content/uploads/2022/03/L-.N.-No.-46-Central-Bank-of-Kenya-Digital-Credit-Providers-Regulations-2022.pdf
- Government Competition Authority of Kenya and Innovations for Poverty Action (Putman, D., Mazer, R., & Blackmon, W.) (2021, May). Report on the Competition Authority of Kenya Digital Credit Market Inquiry https://poverty-action.org/sites/default/files/2023-05/Digital_Credit_Market_Inquiry_Report_2021.pdf
- Government Office of the Data Protection Commissioner, Kenya (2023, December). Guidance Note for Digital Credit Providers https://www.odpc.go.ke/wp-content/uploads/2024/02/ODPC-Guidance-Note-for-Digital-Credit-Providers.pdf
- Government Central Bank of Kenya (2026, April 14). Press Release: Licensing of Digital Credit Providers https://www.centralbank.go.ke/uploads/press_releases/263342722_Press%20Release%20-%20Licensing%20of%20%20Digital%20Credit%20Providers.pdf
- Government Central Bank of Kenya (2025, August 7). Draft Non-Deposit Taking Credit Providers (NDTCPs) Regulations, 2025 https://www.centralbank.go.ke/2025/08/07/draft-non-deposit-taking-credit-providers-ndtcps-regulations-2025/

Where this connects

Institutional pressures in this domain

Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Lending & credit collections AI domain page.

Levers available here and the patterns behind them

Upgrade model — Improve the model
Review on schedule — Oversight cadence & retrospectives
Gate record entries — Human-in-the-loop write gating
Check copied records — Reconcile copied records
Mark AI-written records — Provenance labeling
Store less data — Data minimization
Vet connections — Connection authorization
Understand the system — Understand the system