Worried About an AI Bubble? You’re Asking the Wrong Question.

What the Bubble Debate Misses

There is growing nervousness about an AI bubble. Valuations are stretched, promises are extravagant, and sceptics are waiting for the crash. We have seen this before. The dot-com bubble burst and took countless businesses with it, yet the internet still transformed everything.

AI appears to be following a similar path, and in the workplace the shift is being driven by agentic AI. These systems can interpret context, plan actions, make decisions and carry out work across multiple tools and environments. They are not passive models that wait for prompts. They behave more like digital teammates that can navigate real-world complexity. The nervousness about a bubble is justified. The mistake is assuming that any correction would mean the underlying transformation was not real. Agentic AI will deliver major change and growth, but only for organisations that learn how to use the technology in the right places.

So, the starting point is not whether to adopt agentic AI. It is knowing which problems are well served by this technology and which are not.

Not Everything Needs an Agent

Most organisational processes don’t need agentic AI. Rules-based workflows, scripts and automation tools are easier to deliver, easier to maintain and reliably effective. They often don't need anything more.

If a process has predictable inputs, fixed decision points and a clear sequence of steps, it belongs in traditional software. Pulling data from known sources, transforming it in known ways and formatting a report is not agent territory. The same is true for simple segmentation tasks or systems that behave like a flowchart.

Agents only become valuable when the world becomes messy. When information is unstructured. When data arrives from multiple inconsistent places. When the correct response depends on judgement rather than rules. Think of a purchase manager at a jeweller preparing for a supplier negotiation. The information they need lives across sales reports, margin analyses, historical discount terms and external market trends. No single system holds the answer. Agents can gather and synthesise all of this; traditional software cannot.

If you can sketch your entire workflow on a single flowchart, agents are unnecessary. You are looking for traditional software, not a thinking system.

Where Agents Shine

The next question is where agents can deliver value. Three signals matter most.

The first is real pain. Instead of brainstorming AI ideas, begin by searching for friction. Look for teams bogged down by repetitive work, experts spending more time gathering data than analysing it, and analysts split across disconnected systems. Information silos, delays and duplicated effort are strong indicators that agents may help.

The second is choosing where your agent sits on the autonomy spectrum. At one end are advisory agents that prepare humans with the right information at the right time. At the other are execution agents that handle entire workflows with flexible judgement, surfacing only genuine exceptions for human review.

Consider our earlier example of a purchase manager at a jeweller preparing for a supplier call. Instead of pulling reports from multiple systems, an advisory agent gathers current sales by category, margin data, historical discount terms, and relevant market trends like gold prices and seasonal demand. The purchase manager walks into the call knowing what to push back on, what to stock up on, and where there is room to negotiate. The decision remains theirs, but the data gathering is done.

Further along the spectrum sits an accounts payable workflow that receives invoices in varied formats, extracts the relevant details, matches them against purchase orders, flags discrepancies for human review, and routes approvals based on value thresholds. The judgement comes in handling exceptions like partial shipments, pricing mismatches, or suppliers who format invoices inconsistently. What once required manual review of every invoice now only surfaces the ones that genuinely need human attention.

Most successful implementations start closer to the advisory end and expand autonomy as trust builds. This is not timidity. It is how organisations learn what their agents can handle and where the guardrails need to be.

The third is technical reality. Some projects sound great until you check whether the basics exist. Ask whether the necessary data is digitally accessible, whether your team can access the required systems, and whether the process itself is documented. If these foundations are missing, the work is not an agentic project. It is a data and process project first and foremost.

Be honest about this. Teams routinely underestimate the effort required to surface, clean and connect the data an agent needs. What looks like a three-month agent build becomes a twelve-month data infrastructure programme with an agent bolted on at the end. The good news is that surfacing data from legacy business applications is a solved problem for organisations willing to treat it as a proper workstream. The mistake is assuming agents can paper over data gaps they cannot see.

Success metrics matter too. If you cannot describe what success looks like, you will not recognise it when it arrives. Solutions that fail to deliver tangible benefits get abandoned.

Where to Start

Once you identify strong agent opportunities, you need to prioritise them.

Return on investment matters, whether measured in time saved, cost reduced, risk lowered or quality improved. Implementation complexity matters just as much, including data accessibility, integration effort and technical constraints. User readiness is often the hidden factor, including comfort with AI, willingness to adopt new workflows and leadership support.

A high impact idea with low readiness and extreme complexity is far worse than a smaller idea that is easy to deliver and widely supported. Momentum matters, particularly when introducing novel and transformative technology.

One Step at a Time

Launching multiple agent initiatives at once rarely works. It feels bold, but it reliably leads to slow progress and fragmented learning.

There is a real tension here. Moving too slowly means ongoing inefficiency and lost competitive ground. Moving too quickly means wasted investment, stakeholder frustration and technical debt. One focused project is usually the best choice.

Agents are powerful yet unpredictable. They behave differently from traditional automation and require patience, experimentation and careful monitoring. Starting with one project is not timid. It is strategic. It allows your team to learn how agents behave inside your organisation, which issues they encounter, and which guardrails they require.

Choose one problem. Define success before writing a line of code. Deliver it with a small cross-functional team. Measure everything. Apply the lessons to the next project.

Beyond the Bubble

When people worry about an AI bubble, it is easy to get swept up in the noise. A correction may come, just as it did after the dot-com boom. But the internet still transformed everything, and the organisations that thrived were not the loudest or the fastest. They were the ones that built with care on solid foundations. Agentic AI will be no different.