How AI and ML work
People tend to consider AI (artificial intelligence) and ML (machine learning) as synonyms. More accurately, AI is the broader term, while ML refers to innovative self-programming. Typically, computers rely on strict data processing algorithms developed in advance by humans. In the case of ML, computer systems, including account reconciliation software, must program themselves independently to identify specific data patterns.
For this, machine learning builds upon three pillars – processing power, large volumes of digital data and learning algorithms. Human operators feed an ML system with training sets of labeled data for analysis. In effect, the ML system ‘receives’ a ‘baseline’ to adhere to. From this, it then adjusts its algorithms continually to improve the target outcome in different scenarios.
Machine learning, therefore, goes through multiple iterations of trial and error. Each corrected mistake in the applied logic becomes an integral part of the AI’s training program. The result is machines that rewrite their own software with humans complementing the process with feedback and validation.
There are two common implementation approaches for business process optimisation through AI/ML:
- unsupervised learning: the algorithm searches for patterns and correlations in the data set without prior training
- supervised learning: training the algorithm with large sets of labelled data to predict outcomes in new sets of unlabeled data (this is the more economical approach in computing power terms)
Although computers exceed human capacity for pattern recognition in data, particularly in large volumes of structured data, machine interpretation of data correlations often fails what ‘common sense’ tests. Not all identified patterns are relevant or fit into the specific context of data comparison. This means AI demands multiple training cycles, rigorous correction and heavy customisation if it is to produce the desired results.
AI & ML in account reconciliation – use cases
The issue then becomes – how does an organisation harness emerging AI technology for application into account reconciliation? The following uses cases will offer examples.
Automated data matching
Complex large volume transaction matching is a ReconArt specialty. We support customers in payments and remittances, retail & e-commerce and travel industries where automated matching is critical. They handle enormous transaction volumes and create intricate sets of rules within the system to cover every matching scenario.
Continuous rules adjustment improves match results. AI can assist reconcilers in analysing large data sets for patterns that are relevant but not apparent.
With automated data matching using AI/ML, an ML algorithm trains on a dataset of known reconciled items. Following such training, a refined algorithm automates the matching and reconciliation of subsequent data sets. This can be very effective in reducing the time and effort required compared to manual reconciliation.
Transaction matching by probability
Another use case scenario exploits AI capabilities to identify patterns in data sets without prior knowledge of the matching logic. With this approach, the algorithm will classify transactions based on the probability that a given transaction is equivalent. It uses ‘hard’ or ‘soft’ decision criteria:
- with ‘hard’ criteria, columns need to be equivalent to match
- with ‘soft’ criteria, columns may contain similar data, evaluated by proximity related algorithms.
The output will be a list of possible match groups. These allocate ‘high’, ‘medium’ and ‘low’ probabilities of matching – which are then assessable by human intervention.
Thus, through ML algorithms, a system explores alternatives to increase successful match rates. These then undergo user scrutiny and correction (though many will prove highly probable and valid matches). The significance is that subtle similarities, particularly in large sets of data, which tend to go unnoticed by human operators, are picked up by the system.
Rule creation based on manual transaction matching
Creating rules for transaction matching involves translating the logic of comparison into mathematical algorithms. Experienced reconciliation professionals design and test various combinations of match rules regularly. Although the underlying concept is straightforward, the execution requires manual entry of formulas. No coding or ‘touchless’ rule creation, however, ranks high in enhancing the user experience. ML can minimise human intervention. This makes possible the retirement of writing rules from scratch.
The training cycle here sets the goal for the system to define or suggest match rules for a manually matched set of data. Learning through observation and discovery of patterns in human behaviour modifies the algorithms as an immediate response to change. This type of human-machine cooperation can add flexibility to the reconciliation process. It addresses exceptions as well as new events.
Soft string criteria matching
In every financial reconciliation process, reconcilers inspect unmatched transactions to understand why a specific line has no counterpart. The investigation of outstanding items is crucial from a business perspective. A careful look into unmatched exceptions often reveals a myriad of root causes that need attention. Such root causes might include system and/or human errors, delays, fraud, incurred fees or taxes, foreign exchange differences, etc.
One of the most frequent challenges in data reconciliation seems to be the poor quality of the input data. Typos, abbreviations, missing fields in columns and misplaced characters within a string each can ruin an otherwise relevant identifier. In turn, this prevents the desired match. Reconciliation specialists frequently can pinpoint such instances almost intuitively as they work through the process of manual records comparison. AI has the potential to automate this on a large scale.
When recurring data flaws become part of the rule creation agenda, an AI-enabled reconciliation system can scan for similarities with an acceptable probability level and present them for validation. Improved visibility on such obscured patterns can compensate for data deficiencies. But this is only applicable to a limited extent. If, say, the unique transaction identifiers (in the input data) are missing altogether, a reliable match result cannot be possible.
Identification of account risk levels
One of the objectives of robust account reconciliation processes is the mitigation of financial and operational risk. Accountants habitually monitor account movements and balances for suspicious or unusual behaviour.
Reconciliation software with ML capabilities adds a further layer of controls because they can recognise or flag up underlying risks. An example is fraud detection in credit cards that exceed transaction thresholds.
Anomaly detection
Finally, AI & ML can play a key role in anomaly detection. An ML algorithm can ‘tune in’ to detect any items in the data that do not match expected patterns. This has real applicability for identifying and correcting errors in data. It also has relevance for period-end processes like account certification and variance analysis.
The future of AI & ML in business solutions
R&D in the field of AI has intensified in the past decade. There is increased enthusiasm from business software users fuelled by the marketing slogans. However, it is also necessary to manage and qualify expectations.
The proliferation of AI and ML-enabled solutions is accelerating. In reality, they have yet to demonstrate their full potential. AI still, for example, struggles to comprehend context. Real-life often disagrees with rigid mathematical logic. Closing the gap between the cost of ownership and benefits is also ongoing.
At ReconArt, we regard one of our responsibilities (as technology/reconciliation consultants) as being to implement novel approaches that bring material, practical benefit in the processes we work with. To us, AI and ML are parts of the larger task of monitoring, selecting and introducing all emerging technologies that might provide better, faster and more precise reconciliation processes.
Given the current state of technology, it is inappropriate to expect AI fully to replace humans. Nevertheless, AI and ML in conjunction can enhance human capabilities for problem-solving and introduce a higher level of automation.
In time, AI will likely dominate almost every business function. What is clear to us is that financial data reconciliation will be in the vanguard of adoption.
Since 2011, ReconArt has driven innovation in reconciliation software technology and helps a global client base establish robust, flexible and scalable data reconciliation processes. ReconArt delivers an enterprise class, entirely web-based reconciliation platform leveraged by a varied portfolio of customers from every industry, size, and geography.
ReconArt integrates all transactional matching, account reconciliation, and financial close management processes in a secure and auditable environment with a strong focus on best practices, regulatory standards, and compliance mechanisms. The ReconArt SaaS solution ushers in dramatic efficiencies into clients’ existing manual processes generating time savings and reinforcing precision.
