As AI becomes an ever-present in many businesses, most will now be looking to determine not just when AI will genuinely make a difference – as opposed to when it might just add cost and complexity.
So does every business process really need AI, or should companies be focusing on specific needs which will help them the most?
We spoke to Gregg Aldana, Senior Vice President, Head of Global Solutions Consulting at Appian, to find out more.
- There’s a growing assumption that if AI can be applied to a business process, it should be. Is that the wrong starting point? How should businesses decide where AI genuinely adds value – and where it’s simply adding complexity?
The wrong starting point is asking, ‘Where can we use AI?’ because that puts the technology before the business problem. Organisations should start by asking what they are actually trying to solve and where the biggest bottlenecks or inefficiencies sit.
That means looking closely at which decisions they are trying to improve, what data those decisions rely on and where human oversight needs to sit. That discovery should involve the people who understand the process and its challenges, from IT and business teams through to executive leadership.
Only then can businesses judge whether AI is genuinely going to make a difference. The test isn’t whether AI can do something. It’s whether it can make the process measurably better.
- How do you distinguish between a process that actually needs AI and one that can be handled more effectively with traditional automation, rules or workflow?
The key distinction is whether the task follows predictable rules and has a clearly defined outcome. If it does, traditional automation or business rules can often get you there faster and more cheaply. In those cases, adding AI may not improve the outcome and can introduce complexity that simply isn’t needed.
For example, an insurance company would use business rules to review and classify incoming online claims based on what the claimant selected from a drop-down list in a form. Automotive claims go to one department and home insurance claims go to another.
AI agents become more valuable when the work requires adaptive reasoning, complex context or handling variable inputs. In this case, an AI agent could analyse claims from web and mobile forms with structured data, as well as incoming emails with unstructured content. Based on keywords such as ‘motorway,’ ‘clash,’ and ‘passenger,’ a trained AI agent would deduce that the claim is probably about an automotive accident and route it to the appropriate department.
The mistake is assuming that because an AI agent can perform a task, it should. Businesses need to look at what each part of a process actually requires and use the technology that best fits that need, rather than replacing conventional automation with AI for the sake of it. A balanced portfolio approach is essential – use rules for high-volume, repeatable logic and agents for dynamic triage or investigation.
- What’s the hidden cost of putting AI into every process? Beyond the cost of the model itself, are businesses underestimating the costs around data, integration, monitoring, governance, security and managing exceptions?
One of the highest hidden costs lies in the operational infrastructure required around the AI model to make it work effectively and safely. The cost isn’t just the model. Agents need access to the right data and systems, but businesses also need monitoring, security and governance around what they can access and what actions they can take.
There is then the question of accountability. Organisations need to understand why an action was taken, trace what happened if something goes wrong and have a clear way of dealing with situations an agent cannot handle confidently. Without a unified platform to orchestrate data and enforce policy, the overhead of managing exceptions and audit trails will outpace the AI’s productivity gains.
That overhead can absolutely be worthwhile when AI is delivering meaningful value. But every additional agent introduces complexity, so businesses need to be confident the improvement to the process justifies it.
- Where do you think human judgement remains irreplaceable in enterprise processes – particularly when decisions affect customers, employees, finances or regulatory outcomes? And how should organisations decide when a human needs to remain “in the loop” rather than simply being notified after an AI agent has acted?
Human judgement remains critical where the cost or consequence of getting a decision wrong is high, particularly in regulated industries such as financial services, insurance, life sciences and the public sector. AI can accelerate areas such as client onboarding, insurance, underwriting or clinical trials, but there will still be decisions where the consequences mean a person needs to be involved.
That doesn’t mean a human has to approve every action an AI agent takes. The important thing is understanding the level of risk attached to the decision. Routine activities may proceed autonomously, while sensitive, unusual or more consequential tasks and decisions should be escalated for human review.
Those risk assessments and boundaries need to be designed into the process from the start. Organisations should know which decisions an agent is authorised to make on its own and which require human judgement before an action is taken.
- Is the more realistic future one where people, AI and conventional automation work together, rather than one where AI agents eventually take over entire end-to-end processes?
End-to-end automation isn’t about giving an entire process to a single AI agent. The future is orchestration. A more realistic future is one where people, AI agents and conventional automation work together, because tasks vary within a process that require different skills.
Complex processes require a combination of predictable business rules, adaptive AI agents and human expertise. The goal is to build an integrated process where each component handles what it does best. The key is to understand which parts are best handled by an agent, where a simple rule or workflow will work better and where a person needs to be involved.
The aim isn’t to maximise the amount of AI being used. It’s to build the most effective process around the business outcome you want to achieve.
- Could the pursuit of autonomous AI actually make some business processes less efficient? For instance, if you introduce an AI agent into a process that was previously handled by a simple rule, are you potentially adding latency, uncertainty and governance overhead without creating much additional value?
Absolutely. Putting AI into the wrong part of a process can make it less efficient, not more. Replacing a deterministic rule with a probabilistic AI model introduces unnecessary latency, cost and unpredictability. If a task can already be completed quickly and reliably through a simple rule, introducing an AI agent may not materially improve the outcome.
That doesn’t mean businesses should be overly cautious about AI. It means being selective. Agents are incredibly valuable where a task requires unstructured reasoning, extensive research or adapting to changing information. But they shouldn’t replace conventional automation simply because AI is newer or more sophisticated.”
Sometimes the smartest AI decision is deciding you don’t need AI at all. Using them where a simple process rule already works reliably creates technical debt.
- What should a business ask itself before introducing AI into a process? Is it the business outcomes, the nature of the decision, the cost of failure and the availability of human expertise, rather than simply asking where AI can be inserted?
The most important question is very simple: ‘What problem are we trying to solve?’ Too often, organisations start by asking where they can use AI and then look for processes to apply it to. That puts the technology before the business need.
Instead, businesses should look at where their biggest challenges and bottlenecks really are. Which processes are taking too long? Where is there too much manual intervention? Which decisions could be improved, what data are they based on, and where does human oversight need to sit?
That discovery needs to happen upfront and involve the people who understand the process, from IT to business users to executive leadership. Identify where processes stall or manual handoffs create errors. Map out the required data and processes, regulatory requirements and risk parameters.
Ultimately, AI should be driven by business pull rather than technology push. Once you understand the problem and the outcome you want, you can decide whether AI is genuinely the best way to achieve it – or whether rules, automation or human judgement supported by AI would work better.
- How should companies think about risk when an AI system moves from making recommendations to actually taking action?
The move from recommending an action to actually taking one is where bounded autonomy becomes critical. The goal should be to give AI agents freedom to reason and act, but within clearly defined process guardrails and risk parameters.
That means being explicit about what an agent is allowed to do, what information it can access and when it needs to escalate to a person. An agent shouldn’t operate in isolation; it should sit within governed operational processes with relevant business rules and data security. All AI agent actions should be traceable and auditable within the governance framework.
Human oversight remains particularly important for exceptions, outliers and higher-risk decisions. Ultimately, businesses need to know what an agent did, why it did it and who is accountable for the outcome.
- Do you think businesses are at risk of measuring AI success by how many processes they’ve “AI-enabled”, rather than by whether those processes actually became better? What metrics should executives use instead?
The number of processes using AI tells you very little about whether an organisation is actually getting value from it. If anything, focusing on AI adoption as the metric risks encouraging exactly the wrong behaviour, putting AI into processes simply to say they have been AI-enabled.
Businesses should measure AI in the same way they would any other operational and technology investment: by the outcomes it delivers. Has it reduced cycle times? Lowered the cost of running a process? Improved the quality of decisions?
We’ve seen how significant those gains can be. Processes such as client onboarding for a bank that once took up to two weeks being reduced to just a few minutes. (That bank also automated 96% of onboarding processes whilst growing their new customers by 900%.) The cost of running an end-to-end process can fall tenfold or even more. That’s the test that matters. The goal isn’t to have AI everywhere, it’s to make the business work better.
- If you were advising a CEO who had been told they need an “AI strategy” for every part of their organisation, what would you tell them to stop doing, and what would you tell them to prioritise instead?
Stop mandating AI everywhere for the sake of adoption. I’d tell them to stop asking, ‘Where can we use AI?’ and start asking, ‘What problems are we trying to solve?’ Pressure to have an AI strategy everywhere can quickly become technology for technology’s sake.
Start with the biggest bottlenecks in the organisation. Which processes take too long? Where is there too much manual intervention? Which decisions could be improved? Identify your three highest-value operational bottlenecks and orchestrate the right mix of rules, agents and human judgement to resolve them.
Importantly, CEOs should prioritise business value over AI volume. A successful AI strategy isn’t one that puts an agent in every process; it’s one that makes those processes measurably better.
