Kofi Ndaikate has spent years at the intersection of fintech innovation and bank-grade rigor, advising on how to blend frontier AI with the controls that keep money safe. He’s worked hands-on with teams re‑engineering lending, deposits, and payments—often in high‑stakes environments where trillions move through platforms and regulators scrutinize every log entry. In this conversation, he unpacks why a multiyear, bespoke approach beats off‑the‑shelf shortcuts, how AI is already live from document intake to portfolio monitoring, and what it means to become an AI‑native bank by 2026. Expect candid trade‑offs, concrete operating practices, and a view into how performance awards and an NPS of 81 turn into durable moats, not novelty.
You chose a multiyear path to build bespoke AI rather than adopt off‑the‑shelf tools. What strategic outcomes justified that choice, and how did you rank lending, deposits, and payments for impact versus risk? What trade-offs did you accept to move quickly without breaking controls?
We optimized for three outcomes: measurable client impact, control integrity, and reuse across the enterprise. Lending came first for impact, deposits next for scale and onboarding speed, and payments last to protect a platform moving roughly $2 trillion annually. We accepted trade‑offs like narrower initial scope and staged feature flags so we could ship safely inside our secure infrastructure. The prize was embedding our institutional knowledge—built over five straight top‑10 performance years from 2021 to 2025—directly into models, not someone else’s template.
In lending, how does AI now create credit memoranda and legal documents end‑to‑end? Please share metrics on turn‑time, quality, and exception rates, plus a concrete example. How did you align relationship managers, credit, and legal so workflows improved rather than just sped up?
We orchestrate a chain: document intake, data extraction, covenant mapping, then draft memos and legal schedules with redline suggestions. We track turn‑time, exception density, and post‑close defect rates; I won’t quote figures beyond what’s public, but we’ve seen consistent gains since going live in 2023. A recent C&I deal flowed from intake to near‑final memo in one pass, with legal receiving AI‑tagged clauses that matched house style and reduced back‑and‑forth. Alignment came from shared prompts, joint playbooks, and gating approvals so speed never outruns credit judgment.
Document collection often means messy PDFs and emails. How are you extracting data from unstructured sources, validating it, and surfacing gaps? Where are humans in the loop, and how do you evidence auditability for regulators and counterparties?
We route unstructured files through OCR, layout parsing, and entity extraction tuned on our forms, then bind outputs to a schema used across lending. Validation chains cross‑check totals, maturity dates, and covenant math; any mismatch triggers gap surfacing to bankers and clients. Humans review edge cases and sign off on exceptions before anything influences risk ratings. Every step writes lineage and prompts to our enterprise logs, giving regulators a replayable trail from prompt to decision.
For post‑closing portfolio monitoring, what early‑warning signals matter most, and how are thresholds set? Walk through a case where AI flagged a risk that changed exposure decisions. How do you tune for fewer false positives without missing rare events?
We focus on financial drift versus covenants, payment behavior shifts, and external signals mapped to sectors. Thresholds are policy‑based with overrides, then refined by monitored outcomes and human adjudication. In one case, the system flagged supply‑chain stress across a borrower’s counterparties; we tightened exposure and sequenced outreach, which stabilized terms before renewals. We reduce noise with feedback loops from credit committees while preserving sensitivity to rare but material events.
On digital onboarding and account setup, what steps were re‑designed, and which frictions were removed? Please share abandonment, KYC pass‑through, and average time‑to‑fund improvements. How do you keep fraud controls tight while reducing clicks for good customers?
We collapsed duplicative data entry, front‑loaded document checks, and embedded AI help inside applications. We track abandonment, KYC pass‑through, and time‑to‑fund closely; the improvements are real, even as we avoid quoting new figures beyond our public metrics. Fraud controls run parallel risk scoring and step‑up verification only when signals warrant, keeping good clients on a low‑click path. Everything sits within enterprise access controls, so speed never compromises policy.
Your payments platform processes trillions annually. How are agent‑ready APIs and AI‑driven risk tools embedded without adding latency? Please detail throughput targets, failover design, and how policy changes propagate safely across services.
We built agent‑ready APIs that pre‑compute risk features so inference occurs off the hot path. Throughput targets align with a platform processing about $2 trillion, with active‑active failover and policy snapshots versioned across services. Changes ship behind feature flags and canary routes, with rollback that takes minutes, not hours. Risk tools enrich but don’t block the core rail unless thresholds trip, minimizing added latency.
You deploy AI within enterprise infrastructure. How are data governance, access controls, and lineage enforced from prompt to decision? Describe model risk management, red‑teaming, and monitoring. What independent checks satisfy internal audit and regulators?
We enforce least‑privilege access, encrypted storage, and token‑level redaction before prompts leave any boundary. Model risk management mirrors credit model governance—inventory, documentation, validation, and performance drift watchlists. We red‑team prompts and outputs, stress test jailbreaks, and monitor for hallucinations with business‑rule backstops. Internal audit reviews lineage and approvals, while regulators see end‑to‑end artifacts tied to each production use case.
AI is live in production. Which use cases shipped first, and what was the rollout playbook from pilot to scale? Share the KPIs, guardrails, and stop conditions you used, and an example of something you rolled back and why.
We started with lending document intake and memo drafting, then moved to onboarding and payments risk enrichment. KPIs included turnaround time, exception rates, NPS deltas, and operational loss trends, all with stop conditions for drift or error spikes. One feature that auto‑drafted complex clauses was rolled back when redlines clustered in a narrow legal segment. We retrained on our own clause library and reinstated with tighter guardrails.
With most employees using AI tools, how did you onboard teams at scale? What training, prompts, and workflows drove real productivity gains, and how did you measure them? Share one role that changed materially and how you addressed resistance.
We rolled out guided prompts, role‑based playbooks, and office hours, building from our 2023 enterprise deployment. Adoption reached about 75% of the workforce, supported by curated prompt libraries and embedded QA checks. Credit analysts shifted from first‑draft writing to higher‑order analysis and scenario testing. We measured gains through cycle‑time studies and quality audits, and eased resistance by keeping humans as final approvers.
By 2026, you aim for bankers to spend more time client‑facing. How will you measure time reallocation and client impact, not just efficiency? What processes are you retiring or re‑sequencing, and how are incentives changing to reinforce the shift?
We’ll track calendar telemetry, client meeting counts, and changes in share of time on proactive outreach by the end of 2026. Retiring duplicative data entry and re‑sequencing legal reviews closer to decision points frees hours without risking controls. Incentives tilt toward relationship depth, measured alongside an NPS of 81 versus an industry 41. The idea is simple: more meaningful conversations, fewer back‑office loops.
Frontier models are powerful but sensitive with PII. How do you select, fine‑tune, and evaluate models on bank data while meeting privacy constraints? Describe anonymization, encryption, and retention policies, plus your vendor‑risk and fallback plans.
We fine‑tune inside our enterprise perimeter, with anonymization layers and encryption in transit and at rest. Retention windows follow bank policy, and prompts plus outputs are logged with lineage but scrubbed of direct identifiers. Vendor risk management evaluates frontier providers and enforces fallback models to maintain availability. We choose models that balance capability with predictable behavior under red‑teaming.
You operate a single point‑of‑contact model. How does a banker’s AI copilot gather context, propose actions, and hand off to specialized teams? Where do you deliberately cap automation to preserve relationship quality and judgment?
The copilot assembles context from CRM, recent transactions, and credit posture, then drafts next best actions with rationales. It can generate emails, checklist tasks, and pre‑fill forms before handing to underwriting, legal, or payments ops. We cap automation for pricing discretion, covenant negotiations, and sensitive outreach where judgment matters most. The single point‑of‑contact remains accountable, with AI as a tireless assistant.
Performance awards and a high NPS suggest strong client outcomes. How are you tying AI features to NPS lifts, win rates, or fee income? Share an experiment design, the metrics you tracked, and how you guarded against novelty effects.
We run A/B tests by segment, gating AI features to treatment groups while holding seasonality constant. Metrics include cycle time, win rate, and NPS movement against our 81 baseline. To control for novelty, we extend tests over multiple quarters and look for sustained effects, not day‑one spikes. We also monitor fee income mix to ensure value creation, not just discounts or faster quotes.
Regulators expect explainability and fairness. How are you documenting models for credit, payments risk, and onboarding, and testing for bias? Walk through your independent validation process and how exam feedback has changed your roadmap.
Each model has a full documentation pack: purpose, data lineage, feature logic, limitations, and human‑in‑the‑loop points. We test for disparate impact by segment and maintain override review logs to catch drift. Independent validators replicate results and challenge assumptions before anything is promoted. Exam feedback sharpened our transparency and pushed earlier involvement of compliance in prompt design.
Your journey began with enterprise generative tools in 2023. What surprised you—technically or culturally—when scaling to production systems? Share a hard lesson, a fix that stuck, and how budgeting evolved from experiments to multi‑year investments.
Technically, context fragmentation created subtle errors that only surfaced under load; culturally, teams underestimated how quickly habits would change. A hard lesson was that perfect prompts can still fail without rock‑solid retrieval and lineage. The fix was standardizing on shared schemas and guardrail policies across the three core domains. Budgeting shifted from pilots to multi‑year funding tied to platform milestones and clear stop‑go gates.
Among regional banks, where will lasting moats form—proprietary data, process IP, distribution, or platform partnerships? How do you balance collaboration with vendors and building internal capabilities so you don’t get commoditized?
Moats will come from proprietary data signals and the process IP that operationalizes them across lending, deposits, and payments. Distribution helps, but without embedded workflows it can be copied. We partner for frontier models and keep core orchestration, data governance, and policy logic in‑house. That balance lets us benefit from innovation while avoiding a race to the bottom.
What is your forecast for AI in regional banking?
By the end of 2026, the banks that win will have AI wired into daily work, not just pilots, with bankers spending more time with clients and less wrestling with forms. Single point‑of‑contact models will be amplified by copilots that know context and policy, while platforms processing trillions, like cubiX, run safer and faster. Expect regulators to reward strong governance, pushing explainability, lineage, and fairness from day one. And clients will feel it directly—in fewer steps to onboard, faster answers on credit, and service that earns an NPS closer to 81 than 41.
