Trending Articles
Trending Topics
Trending
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
AI is great, until it isn’t. Then what? The way it’s sold often veers toward the fantastical. We’re led to believe that AI is a knight in shining armour that’s arrived on earth to singlehandedly save the productivity crisis. We’re told that if we just open our arms to AI, we can live in a world where no humans are required, costs are tiny, and results are huge. Sure, as a business owner, it sounds tempting until the AI model in question hits on an ambiguous request, forgets what you asked, hallucinates, or routes something urgent to the wrong team.
The truth is, if you want to make AI work at scale, you need to get real first. Assume things will go wrong, and plan for this eventuality upfront. That means building a world where people and AI can work together, based on what each party is best at. Keeping humans in the loop isn’t just nice-to-have, it’s essential.
AI gets a lot right, particularly in the realm of language. However, when it goes wrong, it does so quickly and confidently, scuppering your NPS score in the process. For example, a customer email lands in your service team’s inbox with vague language, and the model misclassifies it and sends it to the wrong team. Or, a document looks clean but has missing fields, so the extraction engine skips critical details without flagging the gap. Another example is a chatbot that handles routine queries but fails the moment a customer adds context that doesn’t fit the script.
These aren’t anomalies; they’re what can really happen when AI is dropped into a workflow without proper guardrails around it. B2B service work is full of edge cases, missing data, and customer nuance. When a model doesn’t know what to do, it doesn’t ask; it guesses. That’s when errors creep in. The problem isn’t that AI makes mistakes – humans are far from perfect after all – it’s that too often the workflow has no Plan B for what happens next, and teams waste time fixing work that could’ve been done right the first time.
Instead of chasing perfect automation, design your system to handle what happens when things go off-script. What happens when the AI doesn’t know? Or when the confidence score dips below your threshold? The system needs to know what to do next, not guess or fail quietly.
This is where orchestration matters. Decent operations are built with clear rules for when to escalate issues and who to route those issues to. That could be a human reviewing a low-confidence classification or stepping in on a high-value contract with missing data. Use tooling that connects the dots. Eg. The context of what the AI tried to do, why it paused, and what’s needed to keep things moving.
These kinds of handoffs should be wired into the workflow from the very start, so issues can get resolved early, not kicked down the line. Done right, keeping a human in the loop doesn’t slow you down, but it prevents small problems from turning into big ones. Most importantly, it builds trust in the system, proving that AI doesn’t have to be perfect; it just needs a safety net.
There’s a growing gap between AI ambition and how it plays out on the ground. According to PwC’s Global Pulse Survey, 49% of business leaders say they’ve fully integrated AI into their operations, but 41% cite workforce issues like training and culture as top challenges to using it. That disconnect matters. When people don’t understand how a system works or don’t trust what it’s doing, they often end up either bypassing it or quietly redoing the task. Usage drops, shadow processes creep in, and before long, AI adoption stalls and the whole investment underdelivers.
You don’t fix that by adding more AI. The answer is better integration between people and systems. That means transparency, clarity on when humans should step in, and feedback loops that help both sides improve. When the model and the team work together, performance sharpens and confidence builds across the board.
In mature service setups, AI can handle a lot of the grunt work. High-volume, highly repetitive tasks such as categorising emails, extracting data from forms, and routing standard requests. However, it should never be left to its own devices; thresholds should always be in place. When the model isn’t confident or the task needs deeper context, it needs to be passed to a human who can take it forward.
That loop turns every handoff into a learning moment, and so the model sharpens. It also means the team speeds up, and trust in the system grows. Over time, you don’t just get automation, you get a system that adapts. It handles the bulk, flexes for the exceptions, and keeps getting better with use. That’s how modern service delivery should work.
Even if AI did get to the stage where it could handle your business operations alone, we have to ask ourselves what kind of world it would be with human job displacement. Not just from an economic perspective, but on an emotional level too. People buy from people, and in a world where markets are shrinking, productivity is down, and competition is fierce, customer success is the one true differentiator between good and bad service. Your people are still your greatest asset.
Yes, AI is competent at many tasks. But it will always fall down when it comes to empathy and forging a true connection with your customers. After all, an AI model has never ruminated on what they said at a party, or consoled a friend after bad news, or tasted amazing pasta, or touched grass. Human intelligence is still a superpower. AI is designed to take away the mundane, repeatable tasks nobody really wants to do, and free up your teams to move away from customer service to success. That’s the true magic of this technology.
Don’t expect AI to run the whole show if you want it to deliver value. The future of service isn’t robots versus humans, but AI built to work with your people. That means clear design, smart escalation paths, and full visibility into what’s happening.
Automation done well shouldn’t cut people out; it should give them the tools to step in at the right time and keep things moving when the tech hits a wall. Because, however good the model gets, actual intelligence still matters.
