Jensen Huang says AI brokers are a ‘multi-trillion-dollar alternative’. Mark Benioff thinks brokers symbolize ‘what AI was meant to be’. And Satya Nadella thinks SaaS is lifeless. It’s 2025, and brokers are the one sport on the town (or so it might appear). The tech business adores its buzzwords, and ‘AI brokers’ is likely to be the buzziest of all of them! Whereas a couple of vendor platforms are genuinely constructing agentic options into their roadmaps, others are merely ‘agent washing’. I see plenty of confusion amongst Forrester shoppers – consumers of those applied sciences – who’re making an attempt to sift by all this frenzy to make sense of what brokers actually are, what they imply to the enterprise, and what their selections are.
I’ve ideas.
If it doesn’t have company, it isn’t an agent
We’re nonetheless early sufficient alongside the expertise maturity cycle that definitions and traits generally is a bit fluid, however it’s usually accepted that AI brokers are LLM-based constructs that show particular design patterns: planning, reflection, collaboration with different brokers, and power use. Underlying these patterns are two foundational constructing blocks of true ‘agentic’ functionality:
- Company: A defining attribute of an agentic AI system is the ‘company’ to regulate and direct its personal program movement, making unbiased choices in regards to the particular pathways, sequence and nature of actions it should execute to achieve its targets. After all, company might be slender or broad, however AI brokers are anticipated to have broad company throughout a wide range of targets inside a context-space.
- Autonomy: It is a product of an agent’s ‘company’ in addition to the generalized intelligence of at this time’s basis fashions. Autonomy refers back to the breadth of contexts (exceptions, externalities and edge circumstances) inside which the AI can function successfully and ship desired outcomes, with out requiring express directions or intervention from a human.
You may instantly see that company and autonomy feed off one another. Collectively, these traits distinguish true AI brokers from their lesser counterparts.
For those who look fastidiously at lots of the ‘agentic’ choices that SaaS merchandise supply, they arrive throughout as a blended bag. You’ll rapidly understand that these ‘brokers’ have restricted autonomy, or restricted company, or are restricted to such a slender context-space that you just may as nicely have simply used a deterministic workflow or an everyday LLM immediate to provide the identical consequence. Sadly, a number of of the purported agentic demos that I’ve seen from SaaS distributors are merely LLM prompts embedded right into a flowchart-y, deterministic course of movement, inside which they’re deployed to carry out slender duties. Mainly, these are LLM-wrappers round deterministic course of workflows.
These will not be ‘agentic’. As a rule, they’re merely ‘agent-ish’.
The autonomy spectrum
This isn’t to say that there’s little to no worth in these ‘agent-ish’ workflows. Agent-ish workflows have their place in an autonomous ecosystem, and the potential footprint of those ‘agent-ish’ workflows will get higher and higher over the following few months. However it’s nonetheless a stretch to name them AI brokers.
On this context, it’s useful to think about autonomy at totally different ranges. At Forrester we are inclined to map AI methods alongside a spectrum of various company and autonomy, throughout the distinct dimensions of management, execution and monitoring. That is analogous to the idea of ranges of autonomy in self-driving vehicles, however as a substitute, as utilized to enterprise processes. Let’s define the important thing ranges:
- Degree 0: Guide. People, satirically, embody the very best ranges of company and generalized functionality (or ‘widespread sense’). A human worker can normally be tasked into a job with no need detailed directions or step-by-step flowcharts to navigate their job. However the level of autonomy is to cut back this reliance on human labor, and so this degree varieties a baseline from which to measure higher-level autonomy.
- Degree 1: Software program-driven, or rules-based automation. This encompasses conventional software-driven automation, in addition to task-specific assistants that may be constructed utilizing conventional automation tech corresponding to Robotic Course of Automation (RPA) or workflow automation. These methods execute predefined duties alongside preconfigured pathways effectively however lack any significant decision-making skill past easy deterministic logical operations.
- Degree 2: Probabilistic automation. This contains methods that combine machine studying or massive language fashions (LLMs) to reinforce automation, but they continue to be tethered to static workflows. For instance, an RPA-like buyer outreach workflow might dip right into a machine studying (ML) mannequin to generate a listing of consumers who’re more likely to churn. We regularly hear distributors assert that their software program is ‘agentic’ as a result of it may well make non-deterministic choices… nicely, most machine studying fashions work with possibilities and are, subsequently, non-deterministic. That doesn’t make them agentic, as they haven’t any company and are solely centered on a selected job.
- Degree 3: AI operators, or agentic course of orchestration. These quasi-agents mimic company however function inside tightly outlined guardrails. Consider ‘LLM wrappers’ round deterministic workflows. A overwhelming majority of the present wave of so-called ‘brokers’ from SaaS distributors fall at this degree, as do instruments that Forrester phrases as ‘agentic course of automation’. These are ‘agent-ish’ as a result of they ship autonomy solely inside a narrowly outlined context house and have very restricted company inside these slender context-spaces. On this context, it is very important observe that for a lot of organizations, ‘agent-ish’ workflows and hybrid orchestration throughout Degree 2 and Degree 3 – wherever finished proper – will show extraordinarily helpful within the close to time period for organizations which are dipping their toes into the house, however the selection of use circumstances and finesse in technical execution will likely be essential to success.
- Degree 4: AI brokers, or ‘agentic methods’. Programs at this degree exhibit each company and autonomy inside broad contexts. Like a extremely expert human colleague or supervisor, they don’t want a step-by-step flowchart; they’re goal-oriented, utilizing their data and contextual consciousness to find out the very best plan of action. AI brokers price excessive on management and execution dimensions, with restricted monitoring capabilities. A number of examples of true AI brokers are coming into being. Just a few examples would come with Devin, a programming agent, or AI Scientist for analysis and scientific discovery. Now we have seen a number of enterprise use circumstances for these true AI brokers in areas corresponding to drug discovery, advanced know-your-customer processes or superior insights era (to call a couple of of a number of). That mentioned, actually agentic methods function at a degree of functionality that could be a step-function increased than ‘agent-ish’ methods in enterprise worth created.
- Degree 5: AGI (Synthetic Basic Intelligence), or no matter comes subsequent. We don’t know the place AI may evolve within the subsequent 5 years. Whereas AGI is aspirational and poorly outlined at this time, it does describe a future the place AI methods self-govern and handle not solely targets but additionally their evolving goal.
What it means
It’s not unrealistic to think about organizations designed within the type of hierarchies whereby agentic methods handle different types of autonomy throughout Degree 1, 2 and three ‘ (together with agent-ish’ methods), both changing or augmenting human labor in these roles.
Nevertheless, most organizations are at very early phases of this journey. So, it is necessary that expertise consumers and resolution makers take a clear-eyed view to the hype and to know that these ‘agent-ish’ methods will not be the Promised Land of enterprise autonomy, however simply an intermediate (however nonetheless necessary) step alongside the journey.
