The critical gap between ‘good enough’ and truly safe in autonomous finance.
In general life, the difference between 95% and 98% is negligible. It’s one more bite into a loaf of bread, or a few extra pennies in your piggy bank. But as the financial industry moves cautiously towards Agentic AI, a 3% difference in data accuracy represents an exponential gap in risk, revenue and regulatory compliance.
An AI agent that autonomously sweeps savings based on a 95% accurate understanding of a user’s liquidity will make a mistake once every 20 transactions. For a bank with one million customers, that is not a margin of error; it’s a systemic failure.
Join me in investigating: what’s in 3% as we look at what this accuracy gap really means, and your critical minimum threshold for viable automated decisioning.
Why 95% is a failing grade
Banks: let’s do the maths
The severity of this 3% gap hinges on the volume of data that modern institutions process on a daily basis.
Our work with Lloyds Banking Group and Nationwide gives us a solid insight into what this volume of data is, so let’s imagine you’re seeing a transaction volume of 500 million each month.
Operating at 95% accuracy: a Tier 1 Bank with this volume of data would experience a 5% error rate, which is equivalent to 25 million misclassified transactions per month.
Transition this to a 98% accuracy: and the error count drops to 10 million.
Just a 3% improvement eliminates 15 million errors, monthly, 180 million per year.
Moneyhub estimates that banking app transaction disputes cost UK banks at least £23.5 million annually, just to investigate. That figure refers to the operational costs of handling disputes, including administrative, investigative, and customer service time. It does not include the wider cost of chargebacks, payment reversals, compensation, or customer churn, so it’s really the tip of the iceberg.
All this to say: requesting 98 or even 99% accuracy is not pedantic – it’s reflective of the true cost of remediation. At 95%, the data is noisy enough to require constant human supervision. But at 98-99%, it becomes reliable enough for Straight-Through-Processing (STP). Here is where financial institutions can truly unlock the value of Agentic AI and automated workflows.
Lenders: the gap becomes a critical risk
While the 3% gap causes unnecessary costs in the banking world, it can be a hard barrier-to-revenue in lending scenarios.
In fact, this accuracy gap can be the difference between writing a good loan, and rejecting a profitable customer.
This accuracy gap is due to the error rate in transaction processing. Applicant’s income is misclassified, or expenses are exaggerated because 5% of all transactions are erroneous. Even though the data on these applicants suggest they are prime borrowers:
- Affording consistent repayments
- Maintaining a health debt-to-income ratio
- Demonstrating financial stability
While 98% is not perfect, it’s a significant increase in accuracy and reduction in lost opportunities. Operating at 95% accuracy on 1,000 daily applications means 50 applicants per day are processed with erroneous data. If we assume half of those errors result in an unfair rejection (a false decline), that is 25 lost customers every single day.
Transition this to 98% accuracy: the error count drops to 20, and the false declines drop to just 10.
Admiral: a strong case for 98%
When Admiral Money evaluated data enrichment providers, accuracy was a key benchmark. By integrating high-precision categorisation and enrichment by Moneyhub into their underwriting, they successfully automated the approval of an additional 18% of loan applications, applicants that would have typically been referred or declined through traditional Credit Reference Agency (CRA) data alone.
In high-volume lending cases like this one, the difference between 95% and 98% can be just as extreme:
| 95% accuracy | 98% accuracy |
|
|
When accuracy levels climb from 95% to 98%, the noise in the data subsides enough for the lender to trust the machine.
At 95%, it’s too unreliable for full automation, every borderline case requires a human underwriter to manually verify the bank statements. But at 98%, the lender can trust their Agentic systems. For lenders like Admiral, it enabled the 18% uplift in approvals without increasing the portfolio’s risk profile.
The regulatory risk
Under the FCA’s Consumer Duty, firms are mandated to act to deliver good outcomes and avoid foreseeable harm. In the world of automated finance, inaccurate data can become the primary driver of that harm.
When an AI agent or risk model relies on low-fidelity engines, often those dependent solely on Merchant Category Codes (MCC), the granular reality of a customer’s life is lost in translation.
Consider this frequently-seen error in lower-fidelity engines:
- An AI agent misclassifies a vulnerable customer’s gambling transactions as entertainment
- The lender grants further credit to the customer, who is now potentially in financial distress
- The customer defaults, spiralling into debt
- The lender faces regulatory censure for failing to identify the vulnerability, despite having the data
The reason why the accuracy gap is considered exponential is because it’s not spread evenly. 95% accuracy fails precisely where the data is most complex, which is usually where the highest-risk or most vulnerable customers sit.
Moneyhub goes beyond the 3% gap and achieves a 99% accuracy rate when merchants are identified, significantly decreasing the risk of mistaking ‘Ladbrokes’ for ‘Lidl’.
The way that the data engineers have built visibility into the decisioning process also lends well to regulatory pressure. We separate the decisioning logic from the action to approve, refer or deny the applicant, and give underwriters access to both the data and decision path.
Instead of excusing the black box nature of AI, this provides traceability, outcome repeatability and oversight. And in a regulatory environment that demands evidence of good outcomes, the blind spot gets a whole lot smaller.
The ‘garbage in, disaster out’ equation
As we all know, the tolerance for error in finance is near zero as it is. But when you add Agentic AI systems into the equation, decisions must be validated to the nth degree in order to stand up to scrutiny.
In traditional finance, ‘garbage in’ simply leads to ‘garbage out’. It’s like a misclassified transaction in your banking app: annoying and chips away at the user experience, but not cataclysmic.
But ‘garbage in’ becomes ‘disaster out’ when autonomous agents join the system. When a £500 rent payment is automatically swept as surplus into a 90-day savings account, the user can’t pay their rent without taking out a new loan, and their credit score takes a dive.
Cleaning up the trash
It’s up to financial institutions to turn ‘garbage’ into clean data before it even enters the system, and then validate that it is in fact accurate. At Moneyhub, we do this through deterministic filtering.
In the Categorisation and Enrichment Engine, we limit data retrieval to only the customer’s aggregated financial data. Because this is verified (or ‘hard data’ instead of probabilistic predictions based on the behaviour of others in similar cohorts), it acts as a safety rail.
Why 98% is the Agentic sweet spot
Why do we cite 98% and not 100%? Because in the sea of raw banking data, 100% automated accuracy is simply a myth. Cash withdrawals, cryptic ‘in jokes’ as references and new merchants appearing daily make perfection impossible without human context.
At 95% accuracy, users notice the errors constantly. After all, they’re 1 in 20. The trust erodes, they stop linking accounts and algorithms barely get enough feedback to improve.
98% is the tipping point. It’s the threshold at which the institutions and the users go from correcting the machine to trusting it. Errors are rarer, users trust the insights and engagement soars.
Zooming into the Moneyhub model in particular, we have 15 years (and counting) of training data and continuous feedback loops. Our model has seen more variations of Starbucks than any competitor, and every user correction feeds back into the global model. So if a new coffee shop opens in Bristol today, the Engine learns it by tomorrow.
Stop writing off the cost of the 3%
When suppliers ask, “is 95% good enough?”, the answer lies in the cost of the 3%.
It’s the 18% of lost opportunities in loans. Or the multi-million pound regulatory fines for failing to prove good outcomes. And it’s also the operational cost: the £7.40 to investigate each and every transaction that a customer doesn’t recognise in their banking app.
Ultimately, I don’t want to scare you or delay you in adopting true AI automation in your financial institution, but to ensure your foundation is structural enough to support it. Prioritising data accuracy isn’t pedantic; it is the fundamental factor that determines whether your Agentic systems create value or create liability.
Don’t let a 3% gap be the barrier between your institution and the future of autonomous finance.
About John D.
John D is currently a Solutions Architect at Moneyhub, having previously served as the company’s Lead of the Machine Learning and Data Engineering team. John values the opportunity to use ‘tech for good’, and is particularly energised by projects that protect vulnerable users, such as his collaborative work exploring how financial data can be used to identify and support individuals affected by gambling harm.
Outside the world of Open Finance, John is a Director and organiser at Herofest, a major Live Action Role-Playing (LARP) event.
share
