Using AI to enable automations

Companies embarking on their first investments in Artificial Intelligence-led projects aim to use the new technology to automate tasks and processes. Boards seek to measure new AI investments in terms of realising productivity gains. Heads of operations, similarly, are hoping to use AI to reduce friction involved in executing everyday work – freeing up expert capacity to get on with new work which adds more value to customers and ensures the smooth running of the business.   

In this piece, we uncover that the difference between success and failure of an AI pilot designed to automate tasks lies not necessarily in the sophistication of the model selected for the job, but more in detailed analysis of the nature of the work that you are looking to automate and the data controls put in place at outset.  

It is important to understand and map the tasks which sit within processes-  the activities which need to be linked and cross checked. In other words, through careful analysis of the workflows within a process upfront, it is possible to reliably deliver automations using AI.  

What type of work should you AI automate? 

When we look at these sorts of projects for clients, we search for workflows where variability, unstructured inputs, judgement or exception handling dominate. If the work requires interpretation rather than deterministic ‘yes/no’ type rules, or if exception rates are high, AI is usually a better fit than predecessor automation technologies such as Robotic Process Automation (RPA).  

It’s important to measure and map the ‘cognitive load’ required to handle the manual processes you are hoping to automate. The most valuable opportunities for AI to be used to unlock productivity gains sits within tasks that are both time-consuming and mentally taxing. 

It is also important to map real workflows, not idealised ones. Many processes are messy when you begin picking them apart. Workflows must be broken down, task by task. Each task must be understood and described, whether it be about discovery, triage, classification, summarisation, data retrieval, recommendation or decision.  

Again, ideally the workflows we are looking to automate with AI are characterised by judgement-heavy tasks, involving ambiguity and accountability. We analyse decision frequencies, outcome variance, reversibility (ability to return to the original state if something goes wrong), and measure risk associated with automated decision-making.  

Task mining 

To understand how work really happens, several approaches are used together. System logs can be analysed to spot patterns in how tasks move through systems. This is often called ‘process and task mining’. Usage data shows where work sits idle or builds up in queues.  

Reviewing documents, tickets, and files reveals where information is copied, reworked, or passed between people. Finally, observing teams as they work uncovers informal hand-offs, workarounds, and escalation routes that are rarely captured in official process maps. 

Hidden manual work often emerges at system boundaries where ownership is unclear or tools fail to integrate cleanly. These tend to be the areas where productivity is silently lost and AI can have a disproportionately high positive impact. 

Prepare for exceptions 

It’s also important to plan for ‘exception paths’ which take over when a standard, automated process encounters an error, ambiguity or data anomaly.  

While a ‘straight-through’ path handles routine repeatable tasks working with clean data, the exception path might manage up to 20 per cent of cases that do not conform to expected rules, ensuring the process does not halt completely while handling these exceptions in rework loops. With AI projects, exception paths have evolved from simple ‘try-catch’ error handling into intelligent, adaptive and potentially fully autonomous processes. 

Building trust 

Isolating AI-driven productivity improvements from parallel gains such as new training regimes, tooling upgrades, process improvements or policy changes requires careful planning. It’s important to recognise that well designed projects should show productivity gains in a few weeks. However, sustained gains depend on adoption levels and confidence building. The people interacting and overseeing the new system must feel any improvements from its use from their perspective. Early productivity loss is common when users are forced into new workflows before trust is established.  

To mitigate this, AI assistance is typically run in parallel with existing processes during early phases. Behaviour change is encouraged (not mandated) until confidence in the new system builds. Clear communication of positive intent, role-specific training and visible company leadership sponsorship, are all essential. Humans must remain firmly in the loop, while decision boundaries, approval thresholds, and escalation triggers must be explicitly defined and acted on. Human review of new system’s outputs must be a designed-in component of the workflow.  

AI assistance must be designed to help reduce cognitive load for workers. To this end, the system must be able to surface recommendations with context, confidence signals and clear next actions. Raw model output alone is never good enough in terms of building confidence in new AI systems. GaiaLens prioritises controllability first, then explainability, and finally trust. Users only trust systems they can override and understand in terms of decisions and recommendations they are producing. 

Further guard rails are built into our AI systems by focusing on event-driven architectures and use of orchestration logic. By keeping orchestration explicit and deterministic, AI can be used safely inside workflow steps, rather than as the controller of an entire process.  

Measuring productivity gains 

Productivity gains must be real, measurable and attributable. The metrics for measurement need to be agreed by AI project sponsors upfront. Baselines need to be established by measuring how long work takes now, how much is completed, on average, and how often it needs correcting over a specific period of say one month or one quarter.  

Metrics will vary by function. In customer service, typical measures include average handling time, first contact resolution, deflection rates, backlog reduction, and agent utilisation levels. In software development you are likely to be tracking cycle times, throughput, defect escape rate, and quantity of rework needed.  

Summary 

Any use of AI to automate processes needs to begin with a thorough understanding of the nature and type of work you are looking to automate. Some processes are much better suited to AI automation than others.  

It is important to break processes down into the constituent parts and understand properly how long each process takes pre-automation, how many systems the people involved in the tasks have to draw on and cross check. Without all this work it is impossible to accurately calculate the productivity gains from automation. Building AI systems which provide controllability, explainability together engenders vital trust which drives up adoption.