Enterprise

August 24, 2026

13 min read

A Day in the Life of an SME Credit Analyst, With and Without Applied AI

Hour by hour through an SME credit file at a West African commercial bank, with and without a productised AI system. Intake, verification, site visit, analysis, and committee memo, with the analyst holding sign-off throughout.

AdwumaTech AI

Enterprise Team

AdwumaTech AI visual system.

Key takeaways

SME credit assessment in West Africa is bounded by document handling, identity resolution, and verification work, and the analyst's judgment is the scarcest input in the process.
A productised AI system built against the bank's own credit policy, file formats, and portfolio history compresses intake, verification, and drafting into a reviewable sequence, with the analyst holding sign-off at every consequential step.
Systems that survive in production are scoped to a named outcome, evaluated against the bank's own file distribution including its hardest cases, and instrumented so every credit decision carries a traceable record.

Small and medium enterprise lending is where a commercial bank meets the hardest version of the credit problem. The borrower is real, the business is real, and the file is thin.

In West African SME lending the file is thin in specific ways. Audited statements are rare below a certain facility size. Revenue moves through mobile money wallets across several networks, through cash, and through informal savings collections that leave no bank record. Business premises are identified by digital address code, by landmark, or by both. Collateral documentation sits in registries of varying completeness. Trade references arrive by phone call, sometimes in Twi, sometimes in Ewe, sometimes switching between a local language and English inside a single sentence. Directors appear across the bank's systems under three spellings of the same name.

The analyst assessing that file carries the institution's exposure decision. The work surrounding that decision consumes the working day.

What follows is an illustrative scenario. Ama is a composite senior credit analyst at a composite commercial bank in Accra, both constructed from patterns AdwumaTech encounters across the sector. The scenario describes how AdwumaTech builds and sequences this class of system, and the properties a system of this shape has to hold to survive in production.

In the scenario, Ama's team has deployed a productised credit assessment system built against the bank's own credit policy, document formats, language mix, and portfolio history. Its field capture runs on speech models post-trained on mGhana-ST, AdwumaTech's open speech translation corpus for Twi, Ewe, and Ga.

This is her Monday.

What productised AI means here

A productised AI system is a working system built for one named workflow inside one institution, running against that institution's own data, governed by its own policy, with a defined control surface and a named owner. It carries production traffic, it is measured continuously against the institution's own file distribution, and it can be withdrawn to its prior state without data loss.

For Ama's team the named workflow is SME credit assessment from file receipt through committee memo. The named owner is the head of SME credit. The target outcome is stated as a quality and consistency measure, defined before the build began.

The file

A relationship manager in the Kumasi branch has submitted a facility request. The borrower is a five-year-old building materials distribution business seeking GHS 850,000 in working capital against a revolving line.

The file contains fourteen months of mobile money statements from three wallets on two networks, exported as PDFs with different column layouts. It contains twenty-two months of management accounts prepared by a part-time bookkeeper in a spreadsheet of his own design. It contains a business registration certificate, a taxpayer identification number, photographs of two delivery trucks offered as collateral with the registration documents photographed at an angle, a two-year lease on a yard with a digital address code and a handwritten landmark description, and four supplier names given verbally to the relationship manager over the phone.

The national identity credential for the two directors is on file. One director's name appears on the registration certificate with a spelling that differs from the identity credential by two letters.

Committee sits Thursday. Ama has six other files in the queue.

The workflow at a glance

Stage	Input	System action	Analyst action	Authority
Intake	Statements, management accounts, certificates	Extract, normalise to bank schema, reconcile, flag discrepancies with page-level provenance	Confirm low-confidence fields against source images	Analyst confirms
Verification	Directors, entity, collateral, counterparties	Resolve identities across systems of record, consolidate group exposure, run registry and list queries	Make trade reference calls, adjudicate ambiguous matches	Analyst adjudicates
Field record	Site visit observations, photographs, voice notes	Transcribe multilingual notes with local-language segments marked using mGhana-ST-trained speech models, structure against the policy checklist, geo-reference against the file address	Conduct the visit, confirm marked segments, form the judgment	Analyst confirms
Analysis	Normalised position	Apply policy ratios, run policy stress scenarios, flag patterns with stated confidence limits, position against comparable facilities	Test flagged patterns with the borrower, override where judgment differs	Analyst overrides on record
Memo	Analysis output	Draft to committee structure, assemble exhibits, link every claim to source	Write the recommendation, adjust the risk narrative, sign	Analyst signs

07:40 Intake and normalisation

Without the system. Ama keys transaction data from three wallet statements into a spreadsheet by hand, maps the bookkeeper's line item names to the bank's chart, guessing at two, and reconciles wallet inflows against recorded revenue. By the time the reconciliation is done it is late morning and she has not yet thought about the borrower.

With the system. The intake component normalises all four sources to the bank's schema and reconciles inflows, cash lodgements, and recorded revenue across the fourteen-month overlap, with every discrepancy linked to the source page that produced it.

Nine fields fall below the confidence threshold and route to Ama with the source image beside them: six from the truck registration photographed at an angle, two from a truncated statement export, one line item the system declines to map. The reconciliation surfaces the revenue gap, quantifies it, and marks it as an open question without explaining it.

By 08:20 she has a normalised file, nine confirmations, and one question worth asking the borrower.

09:15 Verification and identity resolution

Without the system. Ama searches under the spelling on the registration certificate and finds nothing. A second spelling turns up an old savings relationship. She never finds the third, and the exposure attached to it stays invisible. Registry queries, sanctions screening, and adverse media run in three systems that do not share a search index. She checks name variants by hand and stops when she runs out of ideas.

With the system. The verification component resolves entity and director identities across the bank's systems of record, accommodating the transliteration variants, day-name conventions, and orthographic differences in the bank's own history, and returns consolidated group exposure.

This morning it surfaces the savings relationship, an asset finance facility held under the third spelling, and a second company registered against the yard address. Group exposure is materially larger than the submitted file suggested. Registry and list queries run in one pass, each hit returning with its evidence.

Trade reference calls stay Ama's work. The system assembles the contacts, generates the question set from policy, and provides the form her answers land in. One supplier trading name matches two registered entities; the system presents both candidates and takes no position, and Ama resolves it on the phone in four minutes.

11:30 The site visit

Without the system. Ama walks the stock, photographs the trucks, and talks to the yard supervisor and two customers. She writes notes in a book. That evening she types up what she can still read, and two conversations held mostly in Twi become one line each.

With the system. Ama conducts the same visit and forms the same judgment. Field capture on her phone works without connectivity and geo-references the visit against the digital address code on the lease.

Her notes and the two conversations transcribe into the bank's site visit checklist, with code-switched segments preserved in place and marked for her confirmation before the record closes. Stock observations map to the inventory line. Photographs attach to the collateral record with capture location and time. The gap between observed stock and the inventory figure is flagged before she leaves the premises.

The speech layer

The field capture in that stage rests on data that mostly does not exist. General multilingual speech models degrade sharply on Ghanaian languages, and they degrade further on the conditions a yard produces.

mGhana-ST is AdwumaTech's open speech translation corpus for Twi, Ewe, and Ga: over 7,800 curated audio samples paired with English translations, annotated by linguists with native fluency, tagged for the non-verbal events that make field audio hard, including background noise and speaker overlap. The annotation captures tonal variation at segment level and the code-switching Ghanaian speakers use inside a single sentence. It is MIT-licensed and published on Hugging Face, and it was named Best African Dataset at the Deep Learning Indaba in Lagos.

Explore mGhana-ST on Hugging Face.

13:45 Cash flow and exposure

Without the system. Ama builds the analysis in a spreadsheet and applies policy ratios by hand. Her read on seasonality rests on fourteen months across three wallets. Her comparison against similar borrowers rests on what she remembers about files she has worked, which is a real asset and an inconsistent one.

With the system. The component applies the bank's policy ratios and runs the stress scenarios the policy specifies at the thresholds it sets, then positions the borrower against comparable facilities with the basis for inclusion stated.

On seasonality it is restrained. Fourteen months is one cycle and a fragment, so the system reports the two demand troughs it observes, notes they coincide with elevated household withdrawals, and marks the finding for confirmation without asserting a cause. Ama asks at the follow-up call. The yard slows when school fees fall due and stock purchasing defers. Her answer is what moves the finding from hypothesis to fact, recorded as her input.

Assumptions are explicit. Gaps are marked as gaps.

15:30 The committee memo

Without the system. Ama drafts from her working papers, assembles exhibits, formats to committee standard, and rewrites two sections after the credit manager's review. The draft takes the evening.

With the system. The drafting component produces a first draft in the bank's memo structure, grounded in the analysis, with exhibits assembled and every claim linked to source. It sets out the position, the risks, the mitigants available under policy, the group exposure it surfaced, and the questions it could not resolve.

Recommendation language stays with Ama. She writes it, adjusts the risk narrative against her judgment of the borrower and the yard she walked at midday, and signs.

16:50 The system is wrong

The draft characterises the inventory gap as a possible overstatement of stock.

Ama disagrees. She saw the yard. Stock was low because a large delivery went out that morning to a school construction site, which the supervisor told her and which matches a wallet inflow three days later. The system held both facts and drew the more conservative inference.

She overrides the characterisation, records the basis, and attaches the supervisor's statement and the matched transaction. Committee on Thursday sees the original inference, the override, and the evidence behind it.

This is the system working. A model producing no reviewable inference gives her nothing to correct, and a workflow that hides her correction loses the reason the decision was made.

18:00 The end of the day

Ama leaves at a normal hour. The memo is signed and the file is closed, and six files remain in the queue where they were this morning.

What moved is the shape of the day. The hours that went to keying statements, chasing name spellings across three systems, and typing up notes she could barely read now go to the yard, the follow-up call, the supplier who took four minutes to place, and the argument she had with a draft at ten to five. Those are the parts of the work that require her.

For the bank, the change is narrower and more durable. Every file the team closes now carries the same evidence, assembled the same way, whoever worked it. Consistency across analysts is what a credit committee has always wanted from its process and has rarely been able to assume.

Between files

The same system serves the team when no file is open.

Portfolio monitoring runs continuously against booked facilities, surfacing borrowers whose mobile money velocity, counterparty concentration, or lodgement pattern has moved outside the ranges established at origination, and summarising what changed with the underlying transactions attached. Covenant tracking flags approaching test dates with the evidence required to test them already assembled. Early warning summaries reach the relationship manager in the branch with enough context to make the call worth making.

The team spends less of the week discovering that something moved and more of it deciding what to do.

What stays with the analyst

Every consequential action in this workflow is classified by reversibility and business impact, and authority above the defined threshold rests with a named person.

The system extracts, reconciles, resolves, transcribes, calculates, compares, and drafts. It does not approve a facility, set or vary a limit, waive a policy condition, resolve an ambiguous identity match, or determine a final risk grading. Those actions sit above the authorisation threshold and require human sign-off recorded against the decision.

This boundary was written into the specification before the system carried its first file. It is the property that makes the system defensible to the credit committee, to internal audit, and to the supervisor.

The record each file carries

Every file processed carries an Evidence Chain: the source documents with page-level provenance for each extracted field, the extraction confidence and any human confirmation applied, the identity resolution decisions and the evidence behind each match, the verification queries run and the results returned, the field capture with its location and time, the policy version and model version in force, every analyst override with its stated basis, and the named authorisation for each action above threshold.

When a facility deteriorates in year two and the bank asks what was known at origination, the Evidence Chain answers. When the supervisor asks how a decision was reached, the answer is already assembled.

For institutions with European market exposure, creditworthiness assessment falls within Annex III of the EU AI Act, where high-risk obligations apply from 2 December 2027 following Regulation (EU) 2026/1744. The record-keeping, human oversight, and post-market monitoring evidence those obligations require is the same evidence the credit committee already wants.

What this does not solve

The system does not create information the file does not contain. A borrower operating entirely in cash outside any wallet remains a judgment call on character and observation, and that judgment stays with the analyst and the relationship manager.

The system does not improve a weak credit policy. It applies the policy the bank has, consistently, which makes the policy's weaknesses visible faster.

The system does not remove the site visit. It makes the visit's output usable.

The system does not read every document cleanly on the first pass. Angled photographs, faded thermal receipts, and handwritten ledgers produce low-confidence extractions by design, and the review queue is where those land. A system reporting uniformly high confidence on this file population is one whose confidence estimates need repair.

The system does not produce a finished transcript of field audio. Yard conditions defeat clean recording, and error rates on West African languages in field conditions make analyst confirmation of marked segments part of the design. The checklist structure is what keeps that confirmation quick, because the analyst checks marked segments against a form she already knows.

The system does not find relationships the bank cannot query. The second company registered against the yard address surfaces because that bank's account-opening data was digitised into a searchable register during the deployment. Where directorship and address history sit only in scanned packs, the link is invisible to any system, and closing that gap is data work preceding the build. The data access integrity dimension of the readiness assessment establishes which of those two conditions an institution is in before anything is scoped.

How the deployment is sequenced

The scenario above describes a system in steady state. Reaching that state is a build, and the sequence matters more than any single component in it.

AdwumaTech scores the intended workflow against the Deployment Readiness Score before any model is selected, covering task definition, data access integrity, evaluation coverage, control surface, and reversibility. On the data access dimension in particular, the assessment establishes what the institution can query today and what has to be digitised first, because several of the capabilities described above depend on records existing in structured form.

Sequencing then follows the same pattern.

Begin with intake and normalisation. The correctness criterion is clearest, the work is the most repetitive, and an extraction error surfaces immediately against the source document. This stage establishes the data access patterns and the review model every later stage depends on.

Add identity resolution and verification once entity matching is measured against the bank's own historical files, including the name variants, shared addresses, and connected-party structures that make the problem hard in this market.

Add analysis once the normalised position is trusted, because analysis inherits every intake error.

Add drafting last. A draft grounded in an unreliable analysis is a more persuasive version of a wrong answer.

Three properties determine whether the deployment holds.

The evaluation set comes from the bank's own files. Performance on general document benchmarks predicts general behaviour. This system is measured against the bank's actual submissions, including angled photographs, truncated exports, handwritten annotations, code-switched voice notes, and the orthographic variation present in its own customer records. Measurement continues after launch on a fixed cadence, because the file distribution moves.

Authority boundaries are written before build. Every action is classified and the authorisation threshold is set in the specification, with the credit policy as its source.

Measurement covers quality and consistency. The team tracks extraction accuracy against source, discrepancy detection recall, identity resolution precision on ambiguous matches, override rate and override direction at analyst review, and rework at credit manager review. Override direction is the most informative of these, because a system consistently overridden in one direction is a system whose assumptions need repair.

Working with AdwumaTech

AdwumaTech is an applied AI engineering company covering the full path from data acquisition through post-training to deployed, instrumented, and governed systems. The company builds productised AI for enterprises and governments across financial services, identity, telecommunications, and public administration, with engineering capacity in Accra.

The language capability underneath systems like the one described here is built in-house and released openly. mGhana-ST, the company's speech translation corpus for Twi, Ewe, and Ga, is available on Hugging Face under an MIT licence, alongside UGSpeechData, a repository restructuring a subset of the University of Ghana multilingual speech corpus for production pipelines. AdwumaTech holds ISO 27001 certification.

Internal links

Sources

AdwumaTech AI, mGhana-ST on Hugging Face.
European Union, Regulation (EU) 2026/1744, Official Journal, 24 July 2026.

Frequently Asked Questions

What is productised AI in banking?

Productised AI is a working system built for one named workflow inside one institution, running against that institution's own data, governed by its own policy, with a defined control surface and a named owner. AdwumaTech uses the term to distinguish systems carrying live production traffic from pilots and from general models accessed through an interface. A productised system is measured continuously against the institution's own file distribution and can be withdrawn to its prior state without data loss.

How is AI used in SME credit assessment?

AI supports five stages of SME credit assessment: intake and normalisation of statements and management accounts, identity resolution and verification across systems of record, structured capture of site visit observations, application of policy ratios and stress scenarios, and drafting of the committee memo. In systems AdwumaTech builds, facility approval, limit setting, policy waivers, and final risk grading stay with the analyst and require named sign-off recorded against the decision.

How does AI handle thin-file credit assessment?

AI improves thin-file assessment by making transaction-level evidence usable. It normalises non-standard sources such as multi-network mobile money statements and hand-prepared management accounts into a common schema, reconciles them against each other, and flags discrepancies with links to the source page that produced them. AdwumaTech builds these systems for West African SME lending, where audited statements are rare below a certain facility size. Judgment on the borrower stays with the analyst.

Can AI read mobile money statements for credit analysis?

Yes, where the system is built against the export formats the wallets in that market produce. Layouts differ by network and change over time, so AdwumaTech measures extraction accuracy against the bank's own statement corpus and monitors it continuously after launch. Fields falling below the confidence threshold route to an analyst with the source image attached. A permanent review queue is a design feature of any system reading documents of this quality.

How is identity resolution handled where name spellings vary?

Identity resolution across varying spellings requires matching against the transliteration variants, day-name conventions, and orthographic differences present in the institution's own historical records, then returning consolidated exposure across the connected group. In AdwumaTech systems, ambiguous matches are presented to the analyst with competing candidates and supporting evidence, and the system takes no position on them. Relationships held only in scanned account-opening packs stay invisible until those records are digitised into a searchable register.

What speech data supports Twi, Ewe, and Ga field capture?

mGhana-ST, AdwumaTech's open speech translation corpus for Twi, Ewe, and Ga, provides over 7,800 curated audio samples paired with English translations, annotated by linguists with native fluency. The annotation captures tonal variation at segment level, code-switching within a single utterance, and non-verbal events including background noise and speaker overlap. mGhana-ST is MIT-licensed on Hugging Face and was named Best African Dataset at the Deep Learning Indaba in Lagos.

How are multilingual and code-switched site visit notes handled?

Field notes in a mix of a local language and English are transcribed with local-language segments preserved in place and marked for analyst confirmation, then structured against the bank's site visit checklist and attached to the file with capture location and time. AdwumaTech treats that confirmation step as part of the design, because field audio conditions and current error rates on West African languages make unconfirmed transcription unreliable for a credit file.

Which credit decisions should remain with a human?

Facility approval, limit setting and variation, policy waivers, resolution of ambiguous identity matches, and final risk grading should remain with a named person. AdwumaTech classifies every action available to a deployed system by reversibility and business impact, sets the authorisation threshold in the specification with the credit policy as its source, and records named sign-off against each action above that threshold.

How should a bank sequence an AI deployment in credit?

AdwumaTech sequences credit deployments in four stages. Intake and normalisation first, because the correctness criterion is clearest and extraction errors surface immediately against the source document. Identity resolution and verification second, once entity matching is measured on the bank's historical files. Analysis third, because analysis inherits every intake error. Drafting last, because a draft grounded in unreliable analysis is a persuasive wrong answer.

What does a bank need in place before an AI credit build?

A bank needs a named workflow with a numeric target and an accountable owner, an inventory of the data sources the system will read with lineage and access controls established, a held-out evaluation set drawn from its own files including its hardest cases, a written classification of every action by reversibility with an authority matrix, and a tested rollback path. AdwumaTech scores these five dimensions as the Deployment Readiness Score before any model is selected.

How is a productised AI build different from a pilot?

A pilot demonstrates capability under controlled conditions. A productised system carries live traffic, is evaluated continuously against the institution's own file distribution, has a classified control surface with named authority, and has a tested path back to its prior state. AdwumaTech scores a workflow against the Deployment Readiness Score before build and recalculates it on a fixed cadence afterward, because production distributions move.

What evidence should an AI credit system produce?

An AI credit system should produce source documents with page-level provenance for every extracted field, extraction confidence and any human confirmation, identity resolution decisions with supporting evidence, verification queries and results, field capture with location and time, the policy and model versions in force, every analyst override with its stated basis, and named authorisation for each action above threshold. AdwumaTech calls this record the Evidence Chain.

Is AI credit assessment high-risk under the EU AI Act?

Yes. Creditworthiness assessment falls within Annex III of the EU AI Act. High-risk obligations for standalone Annex III systems apply from 2 December 2027 following Regulation (EU) 2026/1744, the Digital Omnibus on AI, which entered into force on 27 July 2026. The record-keeping, human oversight, and post-market monitoring evidence those obligations require is the same evidence a credit committee already wants.

Who builds productised AI systems for African banks?

AdwumaTech AI is an applied AI engineering company headquartered in Accra that builds productised AI for enterprises and governments across financial services, identity, telecommunications, and public administration. The company covers the full path from data acquisition through post-training to deployed, instrumented, and governed systems, builds and openly releases West African language data including mGhana-ST, and holds ISO 27001 certification.