Every software vendor now claims to offer AI agents.
A chatbot that answers questions is called an agent. A workflow tool that routes forms is called an agent. A coding system that investigates bugs, writes fixes and runs tests is also called an agent. So is a customer-service bot that retrieves account information and drafts a response.
The problem is that these are not the same thing.
This is the central argument in our latest whitepaper, The Agent Platform Trap: How smart companies make shortsighted AI decisions that are expensive to undo, authored by Netsurit CIO, Louis de Klerk. The paper warns that many organisations are making AI platform decisions that appear sensible in the short term, but could become expensive and difficult to reverse within 12 to 18 months.
For CEOs, this is a capital allocation issue. For CTOs, it is an architecture issue. For the broader business, it is a strategic risk. AI platforms are not only software purchases. They influence how work is redesigned, how knowledge is captured, how governance is applied, what can be automated, and how quickly the organisation can improve over time.
This whitepaper argues that one of the biggest mistakes in the market is treating all AI agents as if they sit on one simple maturity ladder, with chatbots at the bottom and multi-agent systems at the top. That view may be easy to understand, but it is misleading. A highly governed assistant can be more valuable than a more autonomous agent in a regulated workflow. A shallow multi-agent system can be less useful than a single deeper agent. The real question is not which platform appears most advanced, but which platform fits the organisation’s work, controls and future needs.
Another common mistake is focusing too heavily on the large language model itself. Businesses often compare models, benchmarks and token pricing, as if the model is the agent. It is not. The model provides the reasoning capability, but the surrounding architecture, or harness, determines whether that reasoning turns into useful work.
That harness includes how context is managed, how tools are called, how errors are recovered from, how state persists, how knowledge is loaded, how the agent interacts with its environment, and how much autonomy it is given. In practical terms, two organisations could use the same underlying model and achieve very different results because the agent architecture around the model is different.
This is where the whitepaper introduces one of its most important distinctions: mediated execution versus environmental execution.
In mediated execution, the agent performs work by calling approved tools and receiving results back as text. This can be easier to control and audit, which makes it useful for structured, policy-bound work. However, it can also limit the number and depth of tasks an agent can complete, because every tool call and every result consumes part of the model’s working memory.
In environmental execution, the agent has its own persistent working environment. It can read and write files, run code, inspect outputs, install packages, test results and iterate inside a continuous workspace. This makes it better suited to long-running tasks, deeper synthesis, messy exceptions and work that involves large bodies of organisational knowledge.
Neither approach is automatically better. The issue is that they are not interchangeable. A business that chooses the more constrained pattern for convenience may later discover that moving to a more capable architecture requires changes to governance, security, operator workflows and the way teams structure work.
For many organisations, the right AI strategy starts with understanding the work itself. The whitepaper identifies three practical classes of work that matter most: structured workflow work, unstructured knowledge work, and cross-source synthesis or profile-building.
- Structured workflow work includes defined operational processes such as service desk operations, case routing, access requests, HR operations and policy-bound internal workflows. Here, control and auditability may matter more than maximum flexibility.
- Unstructured knowledge work is different. It involves ambiguity, judgement and synthesis across incomplete or conflicting information. This includes investment analysis, due diligence, policy analysis, underwriting support, market research and document-heavy advisory work. These tasks benefit from agents that can plan, compare, refine and adapt.
- The third category, cross-source synthesis and profile-building, may be the most underestimated. This is the work of assembling and maintaining a living view of a client, company, household or case using information from emails, meeting notes, CRM records, invoices, account statements, recommendations, filings and regulatory requirements.