Data as Collateral: Unit Economics for Lenders

ClariFi

Financial intelligence and decision support for Kenyan MSMEs.

Educational commentary
MSME Finance
Information Capital
Kenya

Direct answer

Is data really collateral?

No. Permissioned MSME operating data is information capital: it can improve screening, pricing, monitoring and operating efficiency, and can make cash flows and encumbered assets more visible. It is not loss-absorbing collateral unless a separately enforceable security interest exists in a recognised asset—and even then the personal records usually remain restricted.

Educational commentary

Opening scene: three ledgers, no title deed

A Kenyan merchant can show an M-Pesa trail, POS sales, eTIMS invoices, a supplier notebook, and repeat customers—yet the credit conversation may still stop at “Where is the title deed?”

Composite illustration based on common Kenyan MSME record-keeping and credit-application workflows; not a verified ClariFi customer case.

When does permissioned MSME operating data improve a financial institution’s risk-adjusted lending economics enough to justify its full acquisition, integration, governance, monitoring, conduct and cyber cost—and when does the “data as collateral” metaphor become financially or ethically dangerous? Data is not loss-absorbing collateral. It is information capital. The unit of value is not the byte; it is the improved lending decision.

Follow the same fragmented evidence to the credit analyst reconciling incomplete PDFs and exports, the chief financial officer asking whether the data investment clears its hurdle, and the applicant asking what benefit or recourse the new visibility creates.

A title deed tells a lender what may be recovered after failure. A live data trail can reveal whether failure is approaching. Those are different economic jobs. Confusing them produces bad capital allocation, weak recoveries, and extractive data practices dressed up as inclusion.

Kenya’s collateral paradox

High physical cover, expensive recovery, incomplete visibility

CBK’s 2024 MSME bank-credit survey (as at 31 December 2024) shows a large portfolio still paired with heavy collateral demands and costly recovery—evidence that physical security alone does not guarantee healthy economics.

As at 31 December 2024, the banking industry reported about 0.89 million active MSME loan accounts valued at KES 784.4 billion. Of those accounts, 252,502 valued at KES 149.8 billion were non-performing—19.1% of outstanding MSME loan value (and 28.4% of MSME accounts). These are survey aggregates across commercial banks, mortgage finance companies and microfinance banks; denominators and clean-up effects matter when comparing years.

Reported average collateral stood at about 93% of loan amount for micro enterprises, 99% for small, and 112% for medium. Average 2024 recovery cost rose to 16% for commercial banks and 15% for microfinance banks. The survey itself notes collateral as a continuing barrier and discusses reforms that include movable-asset registration and “information capital” through credit-reference scoring models—policy language for better evidence, not a declaration that raw borrower data is prudential collateral.

These figures do not prove that alternative data reduces loss. The narrower conclusion: high physical collateral requirements and expensive recovery coexist. Better visibility and earlier action may improve economics; they do not guarantee them.

As-at date: 31 December 2024; release via CBK in 2025.
Supply-side banking survey aggregates; not a census of all MSME credit outside banks.
Collateral percentages are reported averages, not statutory LTV floors.
Recovery costs are institution-reported averages for non-performing MSME loans.

Three meanings that must not be confused

Data, inference, enforceable asset, and control

Boards lose money when they treat a transaction export as if it were a perfected security interest—or treat a model score as if it were cash.

Conceptual layers for credit committees. Not a legal opinion on any facility.
Layer	What it is	Economic role	What it is not
Raw business data	Transaction and operating evidence	Underwriting and monitoring signal	Enforceable collateral by default
Analytic score or inference	A decision-support output	Selection, pricing, limit, monitoring	Cash or a recovery asset
Identifiable receivable, contract, licence, inventory, deposit, IP right, or proceeds evidenced by data	Potential encumbered asset	Possible recovery/control value if validly secured	Automatically perfected security
Escrow, payment sweep, direct debit, or controlled account	Collection/control arrangement	May affect exposure at default (EAD) or loss given default (LGD) when enforceable	Ownership of the customer’s personal data

Conceptual layers for credit committees. Not a legal opinion on any facility.

Under Kenyan secured-transactions thinking, enforceable collateral is typically a security right in movable or immovable property perfected and enforceable under applicable law—for example under the Movable Property Security Rights Act, 2017 for many movables—not a metaphorical claim over “the data”.

A controller is not automatically the “owner” of personal data. Consent is not an assignment of ownership. Pseudonymisation is not necessarily anonymisation. Data rhetoric does not create accounting recognition under IAS 38 or prudential eligibility under Basel credit-risk mitigation rules for recognised financial collateral.

Follow the data

From sale to learning—and what disappears at each hand-off

Every hand-off converts lived commerce into a thinner variable. Follow the data. Follow the money. Follow the harm.

1. Business event
Cash, credit sale, return, or pass-through may look identical at the till.
Mislabelled turnover becomes false repayment capacity.
2. Record
M-Pesa, POS, eTIMS, bank and notebooks disagree on timing and identity.
Document chasing becomes decision cost.
3. Permission
Consent may be lawful on paper yet economically coerced.
Consent abandonment reduces scoreability and inclusion.
4. Cleaning
Shared phones, Tills and household wallets mix actors.
Reconciliation labour and model latency rise.
5. Inference
Features strip seasonality, margins and contractual context.
False positives and false negatives drive loss and exclusion.
6. Decision
Slow or opaque decisions kill take-up of good offers.
Approval without take-up wastes acquisition spend.
7. Monitoring
Distress often coincides with disappearing digital trails.
Early-warning lead time affects EAD and collections cost.
8. Outcome & learning
Overrides and appeals may correct truth or reintroduce bias.
Conduct cost and model drift if learning is unmanaged.

The unit-economics model

Value the improved decision, not the dataset

The management metric is risk-adjusted contribution per application (RACPA)—not records ingested, API calls, or model features.

RACPA

RACPA = approval rate × take-up rate × risk-adjusted booked-facility contribution − acquisition cost per application − decision cost per application − allocated fixed data cost per application

Booked-facility contribution ≈ interest + attributable fees + attributable cross-sell margin − funding cost − expected credit loss − fraud − servicing − collections − ongoing data/model cost − compliance/vendor cost − economic-capital charge

Expected credit loss (ECL): In a simple one-period illustration, often sketched as probability of default (PD) × loss given default (LGD) × exposure at default (EAD). Full IFRS 9 ECL is scenario-weighted, time-aware and staging-sensitive—do not treat the sketch as accounting policy.
Data-acquisition cost: Consent design, APIs, bureau queries, parsing, storage, cyber controls and vendor fees to obtain and keep evidence usable.
Customer-acquisition cost: Marketing, origination and relationship effort to bring an applicant to decision—distinct from the data stack.
Counterfactual: Compare against incumbent underwriting, a no-new-data model, or a matched champion–challenger cohort with equal definitions, mature vintages and out-of-time tests.

Avoid ambiguous “LTV”: write loan-to-value or customer lifetime value in full. Do not insert borrower data into collateral coverage, loan-to-value, LGD, ECL or regulatory-capital calculations without a separately enforceable and supportable basis.

KES 500,000 illustrative facility

A simplified twelve-month management view

Illustrative only: same income and funding, PD falls from 8% to 6%, LGD stays 65%, manual cost falls, and a data lifecycle cost appears. Net incremental contribution is KES 8,500 against KES 6,000 incremental data cost (~142% as net incremental contribution ÷ incremental data lifecycle cost—not a universal ROI).

Figures in KES thousands. Not a ClariFi, bank, or market forecast. Recalculate with your own vintages.
KES ’000	Conventional	Data-enabled
All-in income	100.0	100.0
Cost of funds	(50.0)	(50.0)
Expected credit loss	(26.0)	(19.5)
Manual operating cost	(20.0)	(12.0)
Data lifecycle cost	—	(6.0)
Risk-adjusted contribution	4.0	12.5

Figures in KES thousands. Not a ClariFi, bank, or market forecast. Recalculate with your own vintages.

KES 500,000 exposure held constant for explanation.
Twelve-month management view.
Conventional ECL sketch: 8% × 65% × 500 = KES 26,000.
Data-enabled ECL sketch: 6% × 65% × 500 = KES 19,500.
Same all-in income and funding cost in both columns.
Net incremental contribution 12.5 − 4.0 = 8.5; 8.5 ÷ 6.0 ≈ 142%.

If PD does not improve, manual work does not fall, take-up weakens, or governance costs are omitted, the data layer may destroy value.

Total cost of data

Fixed and variable costs the board must load

Break-even volume is fixed data-governance and integration cost divided by net contribution uplift per booked facility after variable data costs—not a slogan about “more data”.

Fixed / programme costs

Consent and lawful-basis design
Identity resolution and know-your-business (KYB)
LOS / core-banking integration
Feature engineering and model development
Independent validation and fairness testing
Vendor concentration, lock-in and outage contingency

Variable / per-decision costs

APIs, bureau queries, ingestion and parsing
Cleansing, classification and reconciliation
Compute, storage, encryption and cyber controls
Drift monitoring and retraining
Human review, explanations, appeals and complaints
Expected incident, remediation and regulatory cost

No invented market threshold is offered here. Your break-even depends on segment mix, take-up, override rates and realised losses.

When more data produces a worse credit decision

Counter-examples that must sit on the credit-committee table

More signals can worsen economics when the model confuses turnover with cash, identity with behaviour, or correlation with recoverable loss.

Illustrative failure modes. Controls are design prompts, not product claims.
What the model might infer	What is actually happening	Metric distorted	Control required
High GMV proves capacity	Negative margins; cash exits faster than it enters	PD / limit too optimistic	Margin and working-capital views; human review on high-turnover thin-margin files
Aggregator receipts are income	Pass-through farmer payments inflate turnover	Income and PD	Entity mapping; net-of-pass-through features; agribusiness playbooks
eTIMS invoice = paid receivable	Customer has not paid; invoice is evidence of sale, not cash	Cash-flow and PD	Reconcile invoices to receipts; age receivables
Volatility = distress	Seasonal or climate-exposed enterprise with normal cycles	PD / pricing	Seasonal cohorts; climate overlays; relationship context
One phone or Till = one risk subject	Shared household or agent wallet mixes several people	Identity and behaviour score	Identity resolution; shared-device flags; contest routes
No digital trail = high risk	Viable cash-heavy merchant remains digitally invisible	Approval / exclusion	Alternative evidence packs; branch observation; avoid proxy discrimination
Cold-start = weak management	Capable new enterprise lacks history	Approval / limit	Graduated limits; mentor or supplier data; champion–challenger grey band
Dense transfers = strong trade	Round-tripping or fabricated orders game the model	Fraud / PD	Graph checks; supplier confirmation; velocity rules
Platform account = applicant entity	Till belongs to a different person or legal entity	Identity / recovery	KYB; beneficial-ownership checks; account-control evidence
Missed repayment = enterprise failure	Government or corporate payment delay	Delinquency / ECL	Receivable ageing; obligor concentration; override with evidence
Stable feed forever	API revocation, outage or vendor lock-in interrupts monitoring	Monitoring / EAD	Multi-source resilience; kill-switch; manual fallback
Device or language predicts risk	Proxy for gender, poverty, ethnicity or informality	Fairness / conduct	Fairness testing; prohibited-proxy reviews; documented recourse
History is truth	Training data reproduce past exclusions	Approval equity	Out-of-time tests; inclusion metrics; policy override governance
Better PD means better recovery	Data improves PD but LGD unchanged without enforceable recovery	ECL / capital narrative	Separate PD and LGD evidence; do not invent collateral value from scores
Monitoring will catch distress	Transaction data disappears precisely when distress begins	Early warning	Stress playbooks; relationship contact; alternative sensors
Override always helps	Relationship-manager override corrects context—or reintroduces bias	Calibration / conduct	Logged reasons; dual control on large overrides; learning loops
Consent = free data asset	Declining consent means no credit; consent is coerced	Conduct / lawful basis	Proportionate alternatives; clear value exchange; revoke without silent punishment where law requires
Data licence is liquid collateral	Transferable receivable may be securable; personal records remain restricted after default	Recovery / LGD fantasy	Legal review of security package; never treat restricted personal data as freely transferable recovery property

Illustrative failure modes. Controls are design prompts, not product claims.

Relationship banking still matters

Codifiable data and human context are complements

Models compress what can be measured. Relationship context often holds what cannot—yet humans must be able to disagree, not rubber-stamp.

BIS Working Paper 1244 (Artificial intelligence and relationship lending) is research, not binding guidance. It explores how AI and soft information interact in lending relationships. Treat it as a prompt for design: keep soft information where it still predicts, and keep a qualified reviewer who can understand, correct and document.

Kenya’s Data Protection Act, 2019—section 35—already addresses solely automated decisions with legal or similarly significant effects, with notification and reconsideration routes, elaborated in the Data Protection (General) Regulations, 2021. For a fuller mapping of policy → system → consequence, see Of Company Policy and Resulting Consequences rather than repeating that analysis here.

What the board should measure

A dashboard that prefers contribution to AUC

AUC or Gini is a diagnostic, not the board’s commercial success measure. Ask whether incremental income, loss avoided, manual cost avoided, fully loaded data cost, and borrower benefit clear the hurdle together.

Access

Scoreability, consent conversion, approval, take-up, first-time formal borrowers, time-to-cash.

Economics

RACPA, contribution per booked facility, data cost per decision, pricing-floor change, break-even volume, risk-adjusted customer lifetime value versus customer-acquisition cost.

Risk

Vintage default, ECL, realised net loss, fraud, roll rates, recovery time, collection cost, stress loss.

Model

Calibration, approval at constant bad rate, bad rate at constant approval, drift, overrides, missing-data rate.

Conduct

Outcomes by gender, geography, business type, informality and shared-device status; explanations, reviews, disputes, complaints, consent withdrawals and breaches.

A governed 90-day pilot

Implementation window, not full loss maturation

Ninety days can stand up governance, champion–challenger mechanics and early operational metrics. It is usually not enough to observe all defaults. Separate pilot duration from performance windows.

1.Select one segment and one repayment source.
2.Establish incumbent cohort economics and definitions.
3.Choose the smallest relevant dataset.
4.Run champion–challenger or risk-neutral grey-band testing.
5.Keep meaningful human review and documented recourse.
6.Measure mature-vintage economics, inclusion and harm.
7.Stop, redesign or scale based on pre-agreed hurdles.

ClariFi relevance and boundaries

Evidence assembly for MSMEs—credit decisions stay with the institution

ClariFi helps MSMEs assemble clearer operating evidence and decision history so lenders review more consistent packs. It does not underwrite, score as a bureau, or guarantee approval.

Shipped capabilities (public position)

Connect or assemble operating evidence across sales, expenses, M-Pesa, POS, stock and bank signals where available.
Surface cash, margin and readiness gaps.
Build decision history that shows how the business acts.
Export lender-oriented packs, checklists and memos.

Boundaries

ClariFi is not a bank, credit bureau, or licensed lender.
ClariFi does not approve or underwrite loans.
Readiness signals are decision-support inputs, not credit guarantees.
Credit decisions remain with the financial institution.
Institutional integrations require an appropriate commercial and governance agreement.

If the institution gains visibility while the MSME gains nothing on access, price, limit, speed or resilience, the model risks becoming extraction rather than inclusion.

The future belongs to the institution that uses the smallest, most relevant, least intrusive dataset to make the fairest risk-adjusted decision.

Related pathways: lender readiness, institutions, privacy, consents, and security. For an agribusiness evidence trail, see Omwami’s eTIMS commentary.

Conclusion

Value at the margin

Data becomes valuable at the margin, not in the abstract. It becomes collateral-like only when it makes genuine cash flows and enforceable assets visible, measurable and financeable.

The winning institution will not be the one that collects the most data, but the one that uses the smallest, most relevant, least intrusive dataset to make the fairest risk-adjusted decision.

Sources

Sources and method note

Method note: ethnographically informed commentary using public CBK, ODPC, Kenya Law, BIS, IFRS, IFC, World Bank and CGAP materials, plus public ClariFi product pages. No private customer records, private chats or unverified interviews were used. The opening scene is a labelled composite. AI-survey percentages cited elsewhere in ClariFi materials should be read as self-reported responses from CBK’s surveyed population (37 commercial banks, 1 mortgage finance institution, 14 MFBs, 3 CRBs and 70 DCPs as of 31 December 2024)—not audited ROI.

FAQ

Frequently asked questions

Is data legally collateral in Kenya?

Generally no. Collateral is a security right in property perfected and enforceable under applicable law—for many movables, under the Movable Property Security Rights Act, 2017. Permissioned operating data is usually an underwriting and monitoring signal. Data may evidence an encumbered receivable, contract, inventory or proceeds without itself becoming freely transferable recovery property.

What is the difference between data as collateral and alternative credit scoring?

“Data as collateral” is a metaphor about recovery and financeability. Alternative credit scoring is a screening and pricing technique that uses non-traditional features to estimate risk. Scoring can change PD estimates; it does not, by itself, create enforceable collateral or lower LGD unless a separate security or control arrangement exists.

Does transaction data reduce PD or LGD?

It may improve PD estimation when it reveals repayment capacity earlier and more accurately. LGD typically moves only when recoveries improve—through enforceable security, control of proceeds, faster collection, or better workout. Do not assume a PD gain automatically becomes an LGD or capital gain.

How should a bank calculate the ROI of alternative data?

Compare risk-adjusted contribution against a defined counterfactual after loading full data lifecycle cost. Use RACPA-style arithmetic: approval × take-up × booked contribution, minus acquisition, decision and allocated fixed data costs. Require mature vintages, equal definitions and out-of-time tests—not AUC alone.

Which MSME data is closest to repayment capacity?

Evidence closest to cash available for debt service after margins, obligations and seasonality—reconciled receipts, bank and mobile-money inflows net of pass-throughs, aged receivables, and verified obligations. Turnover, invoices and platform GMV without cash reconciliation are weaker proxies.

What does Kenya’s Data Protection Act require for automated credit decisions?

Section 35 addresses decisions based solely on automated processing that produce legal effects or significantly affect a person, with exceptions and duties to notify and allow reconsideration. The 2021 General Regulations elaborate explanation, human intervention and fairness expectations. Map your actual decision chain before assuming an exemption.

Can a financial institution recognise customer data as an intangible asset?

Economic usefulness does not automatically satisfy IAS 38 recognition. Internally generated customer data and predictive models often fail recognition tests. Do not treat “data as an asset” rhetoric as accounting policy without specialist advice on the specific facts.

Does ClariFi approve or score MSME loans?

No. ClariFi provides financial intelligence and decision support so MSMEs can assemble clearer evidence. Readiness signals are not credit scores for regulatory bureau purposes, not guarantees, and not underwriting decisions. Lending decisions remain with the financial institution.

CLEARER EVIDENCE, NOT A CREDIT DECISION

Help MSMEs assemble lender-oriented operating evidence—without claiming the underwriting.

ClariFi has a commercial interest in clearer MSME evidence. Digital visibility does not guarantee approval, pricing, or repayment.

ClariFi helps MSMEs connect or assemble operating evidence, surface cash, margin and readiness gaps, build decision history, and export lender-oriented packs or memos. Credit decisions remain with the financial institution.

Explore MSME intelligence for banks and lenders Contact ClariFi Run the Free Financial Diagnostic

Institutional pathway for credit teams and partners.