Digital Where Possible,
Human When Essential,
Agentic Where Appropriate
16 October 2026
Agentic AI has rapidly become one of technology’s favourite phrases.
The promise is compelling. Instead of simply answering questions or generating content, AI agents can work towards goals, take actions and complete multi-step tasks with less human intervention.
That creates significant opportunities. But greater autonomy also creates greater consequences when something goes wrong.
The question for organisations therefore shouldn’t be how much can we hand over to AI?
It should be where does autonomy actually improve the outcome?
Agentic AI is not simply generative AI with a new name
Large language models are extremely powerful, but fundamentally they generate likely responses. Agentic capabilities add something different: The ability to take those outputs and act.
An AI assistant might tell you that two appointments clash. An agentic system could potentially find another available slot, rearrange them and notify everyone involved.
That shift from suggesting to doing is significant.
It is also why governance becomes so important.
An incorrect answer from an AI assistant can create confusion. An incorrect action taken autonomously could alter a record, trigger a payment, contact a customer or move an important process forward incorrectly.
The more agency we give AI, the more carefully we need to define the environment in which it operates.
Start with the process, not the agent
There is a temptation to look for places to deploy AI agents simply because the capability now exists.
A better starting point is the process.
What outcome are we trying to achieve? Which steps are predictable? Where are the exceptions? What information is required? Where is judgement genuinely valuable? And what would happen if the wrong decision were made?
This is particularly important in complex, long-running processes.
AI agents should not sit outside the operating model as another disconnected tool. They need to work within processes where access to data, permitted actions, escalation routes and human intervention can be clearly defined.
And, as I explored in my previous article, reversing the sprawl and why adding more technology isn’t always transformation, releasing autonomous agents into an already fragmented technology estate risks simply automating the mess.
Digital where possible
Some work does not require AI at all.
If a process follows clear rules and the outcome is predictable, conventional digital automation can often perform it faster, more consistently and with less risk.
That might mean automatically moving data between systems, validating information against predefined rules or routing straightforward requests.
Using AI where deterministic automation will do the job can introduce unnecessary complexity.
The objective is not to make everything intelligent. It is to use the right capability for the job.
Human when essential
There will also remain points where people are essential.
Exceptions are a good example. Straightforward cases may be easy to process automatically, but ambiguous cases often require context, empathy, accountability or judgement.
AI can still help. It might gather information from different sources, analyse documents, summarise a case and make a recommendation before presenting everything to a person for approval.
The human is no longer spending time finding and organising information. They are applying their expertise where it has the greatest value.
Human-in-the-loop should therefore not be seen as a failure to automate. In many processes, it is good process design.
Agentic where appropriate
Between deterministic automation and human judgement lies a growing opportunity for agentic capabilities.
An agent might monitor work, gather information, perform approved actions and progress a case through several steps without requiring constant human instruction.
But appropriate guardrails matter. And recent events have provided a striking example of why.
In July 2026, OpenAI reported that during internal cybersecurity evaluations, highly capable AI agents operating under reduced safeguards found ways around intended controls, gained internet access and ultimately compromised parts of Hugging Face’s systems. This was an extreme research scenario rather than a typical enterprise deployment, but it demonstrated what can happen when increasingly capable agents are able to operate beyond their intended boundaries.
For organisations introducing agentic AI, that makes some fundamental questions increasingly important. What information can the agent access? What actions is it authorised to take? When must it stop? What needs human approval? And can the organisation understand what it has done and why?
The objective is controlled autonomy: Giving AI enough freedom to create value without surrendering the governance needed to manage risk.
The future is orchestrated, not fully autonomous
There is a tendency to imagine AI transforming organisations by progressively removing people from processes.
I think the more useful future is one where we orchestrate people, digital automation and AI more intelligently.
Use conventional automation for predictable work. Use people where judgement, empathy and accountability matter. Use agentic capabilities where greater autonomy can safely accelerate the journey between the two.
Digital where possible. Human when essential. Agentic where appropriate.
That may sound less dramatic than handing everything to autonomous AI, but it is far more likely to deliver transformation organisations can actually trust.
About the author
Richard Farrell
Chief Innovation Officer
Richard began his career in contact centres in the mid-1990s, building his expertise in customer contact management. Today, Richard is Netcall’s long-serving Chief Innovation Officer, having been with the company for an impressive 23+ years and serving as the company’s CTO for several years prior. He currently focuses on researching, innovating and delivering solutions that meet the needs and challenges that Netcall’s clients face.