Why agentic AI is about leverage, not automation
Agentic AI changes what a small team can credibly take on
Many conversations about AI in small and medium sized businesses start with automation: making current processes faster or cheaper. It’s an understandable framing, and a limiting one. It treats AI as another efficiency tool and measures it by how much existing work it can replace.
Agentic AI deserves a different measure: how much responsibility it can take on. At its simplest, an agentic system is a language model given a goal, access to tools, and the ability to check its own progress. It can break a goal into steps, act, observe what happens, and adjust. That loop is what separates it from a script, and it’s why an agentic system can share the decision-making burden. Used well, it’s a capacity booster: a way for a small team to take on more than headcount alone would allow.
Where traditional automation stops
Traditional automation is built on certainty. A rule is defined, a trigger occurs, and a predetermined action follows. That works well while processes stay stable, but it’s brittle. Every exception has to be anticipated in advance and encoded, and a small change in input or context can stop the whole workflow.
Agentic systems tolerate ambiguity. When something unexpected happens they can reassess, try an alternative, or escalate to a person. The trade-offs run the other way: they’re slower, they cost more per task, and they don’t behave identically twice, so they need oversight in a way a script never does.
Traditional
Designed to execute predefined workflows as efficiently as possible.
- ■ Predictable and repeatable
- ■ Cheap to run at high volume
- ■ Works best when processes are stable
- ■ Every exception must be anticipated
- ■ Breaks when conditions change
Agentic AI
Designed to pursue goals and adapt, within clear boundaries.
- ■ Goal-driven, plans its own steps
- ■ Adapts when conditions change
- ■ Escalates when it's unsure
- ■ Slower and costs more per task
- ■ Non-deterministic, so needs oversight
Neither replaces the other. Stable, high-volume processes still belong in traditional automation. Agentic AI justifies its cost where the work is variable, needs some judgement, and would otherwise land on a person.
The talented new hire
The most useful mental model is a talented new hire: capable from day one, but with no knowledge of the business until it’s provided. An agentic system performs best when it’s onboarded properly: context about the business, the tools it may use, what success looks like, and clear guidance on what to avoid and when to escalate. Anthropic’s guidance on building effective agents points the same way: the design of the goal and its constraints matters more than the sophistication of the model.
Prompts are onboarding, not commands
Software development is the domain where this is furthest along, with most developers now reporting AI tools as part of their workflow (Stack Overflow Developer Survey), and it makes a good worked example. A coding agent is onboarded through a document describing how the system works and the house style to follow. Automated tests act as guardrails that catch it when it drifts. It’s expected to stop and ask before doing anything destructive or ambiguous, and everything it produces goes through review, with corrections folded back into its guidance. On a recent engagement, described in modernising legacy SaaS without a big-bang rewrite, that setup let agents explore a two-decade-old codebase in days rather than weeks, and build working prototypes while the options were still being weighed.
Onboarding, guardrails, escalation, review. None of that is specific to code, which is what makes the example transferable. The same loop can triage support tickets, chase data quality issues, or draft first responses for a person to approve. The data-quality example comes from real work: on an insurance engagement, an agentic research system took a company name and returned structured data reliable enough for live quoting, work that would once have meant a brittle rule-based crawler.
Like any new starter, an agentic system also fails in familiar ways. Vague goals produce actions that are technically reasonable but commercially unhelpful. Missing constraints lead to overstepping or inconsistent behaviour. When domain knowledge is assumed rather than provided, misjudgements follow. The quality of outcomes follows the quality of the guidance the system receives.
Keeping humans in the loop
Agentic AI works best when people step in at a few well-chosen moments. Intervention is valuable at three points:
-
At the start, when goals, constraints, and context are defined.
-
During exceptions, when something unexpected occurs or confidence is low.
-
In review, where outcomes are assessed and the guidance is refined.
In the coding example those are the task brief, the agent stopping to ask, and code review. Feedback is what makes the loop compound: corrections fed back into future behaviour gradually reduce the need for oversight, though they never remove the need for accountability. And however capable the system is, it carries no accountability, so high-stakes judgement isn’t the place to withdraw supervision. Usefulness degrades quickly when goals keep shifting or the underlying data is unreliable.
Leverage as the real opportunity
The value of agentic AI is leverage. A small team can pursue opportunities that would otherwise have to wait, hand the coordination to the system, and keep its attention for the work that genuinely needs human judgement: decisions, relationships, and direction.
Getting there doesn’t need a transformation programme. It starts the way any good hire starts: one well-briefed goal, sensible boundaries, and someone reviewing the output. The leverage compounds from there.
For teams wondering where agentic AI could take on real work, a short exploratory conversation is a sensible first step.
References
- Building effective agents, Anthropic
- Developer Survey, AI section, Stack Overflow