The word ‘agent’ has become meaningless. Gartner predicts that nearly half of enterprise applications will include integrated task specific agents by the end of the year, and by 2028, a third of users will have moved from native applications to agentic front ends. Despite this, most of the AI agents on the market right now are automation with a rebrand: trigger-based workflows, scripted processes, API-connected chatbots.
Yet the same analysts expect more than four in ten of those agentic AI projects to be cancelled by the end of 2027, and estimate that of the thousands of vendors claiming agentic capabilities, only around 130 genuinely deliver them. Adoption is not the problem; attrition is. Returns will only come from agents that autonomously observe, act, reason, and evaluate within interactions. The expectations and reality of what constitutes a “real” AI agent are quickly diverging, and businesses are paying for the gap.
Automation vs autonomy
In agentic AI, two concepts are frequently conflated. The first is event-driven autonomy which refers to the specific conditions or “triggers” — such as a user prompt — that initiate an AI’s actions.
The second is reasoning autonomy, or how a system makes decisions. A system with genuine reasoning autonomy decides which tools to use rather than following a script. It reacts to what it finds and plans multi-step approaches to novel problems. It recovers from errors and dead ends.
The tech industry has become very good at trigger autonomy but, without reasoning autonomy, it’s just faster busywork. Part of the confusion is a misunderstanding of what large language models are for. An LLM is a language model, not a maths model. Its strength is interpretation: it can read the systems a business already runs, understand context, and reason through ambiguity. What it can’t do reliably is computation at scale.
An agent asked to underpin a multi-million pound decision needs its language and reasoning layer paired with deterministic calculation, where the same inputs always produce the same auditable results, so the answer is exact rather than plausible. This is where most agentic projects break down: a probabilistic reasoning layer bolted onto systems with no deterministic execution produces answers that nobody will sign off a major decision against.
A language model on its own gives you a plausible answer; a language model on top of a deterministic engine gives you an exact one, traceable to source, with the workings attached. In an enterprise, the second is the only kind you can take to the board.
The real AI agents
There’s a straightforward framework for evaluating whether an AI is worthy of being called an agent. Real agents:
- Observe. They don’t just pull data; they actually understand how your business operates. They read the live state of the business: the current plan, the driver logic, and the model that connects them, not a stale export.
- Act within existing workflows. They operate within core business systems, making changes from within, rather than working beside them. An agent that can only generate text is more of an expensive autocomplete function. Generating text alone does not make something an agent. Acting means computing precise financial and operational outcomes and writing them back into the systems of record.
- Reason. Effective agents do more than follow instructions; they apply structured logic to ambiguity. By identifying trade-offs and dynamic multi-step paths, high-value agents transition from passive executors to strategic partners—challenging user assumptions, asking clarifying questions, and forcing a deeper reflection on the task at hand.
- Evaluate. Real agents assess whether actions achieved intended outcomes, identifying gaps, and refining accordingly. This is what turns a one-shot response into genuine problem-solving.
Not all agents do the same job
Agents that fit this framework still operate at different levels of value.
The first tier is workflow automation: accounts payable, standard operating processes, high-volume repetitive tasks. This is where most of the market resides currently.
The second tier is augmentation. Every enterprise has gaps between its core systems, historically filled by people manually gathering and reconciling data, a spreadsheet army assembled to make one platform talk to another. Augmentation agents close those gaps directly, fetching and connecting data across systems so a plan that once took weeks of collation can be iterated dozens of times in a fraction of the time.
The third tier is advisory: supporting decisions with no existing precedent, like M&A or new business cases, where data must be assembled in ways it never has been before. This is where agents stop saving time and start building on what an organisation is capable of. For example, an agent that assembles the case for an acquisition: connecting two organisations’ plans, stress-testing the assumptions behind the synergies, and producing a decision-ready view in hours rather than weeks.
Why architecture determines outcomes
Enterprise software has spent decades at the centre of the organisation, forcing people to learn its logic, fill its gaps and work around its processes. Genuine agents reverse that relationship. The person is at the centre, and the agent adapts the systems to them. Users no longer need to know where a number lives or which system produced it; they ask the question and get a verifiable answer.
Tech leaders can’t expect agentic AI to hit peak performance out-of-the-box. As predictive, generative, and agentic AI become more embedded in enterprise operations, a new function is emerging: creating, training, and governing agents. Businesses can’t afford to skip this step. Success depends on governed, auditable access to enterprise data, with controls on cost and security, alongside evaluation frameworks that keep agents aligned with business intent.
Success will not be measured by who has the most AI features, but by integration. Adoption often fails when AI shows up as a separate tool; a dashboard you have to check, another system demanding your attention. This is an architecture problem, not a tech problem, and it’s keeping businesses from realising their AI potential.
The most effective companies have made AI invisible by embedding it into existing workflows, transforming static legacy systems into a real-time infrastructure that connects operational decisions to measurable financial results. The test for any vendor selling an AI agent is simple. Ask what happens when the agent is wrong, and whether the same inputs produce the same answer twice. If they cannot answer both, you are buying automation with an agentic price tag.
