Beyond the Hype: What AI Agents Actually Do in an Enterprise

 

IN THIS ARTICLE The second article of twelve. If Article 1 argued why Caribbean executives must engage with AI now, this article installs the vocabulary needed to engage well — and the discipline to tell a genuinely agentic product from one that has been rebranded as such. Most Caribbean executives have used consumer AI tools. Very few have a precise mental model of what distinguishes a chatbot, a copilot, and an agent — and that gap is the single largest source of confusion in current boardroom conversations about artificial intelligence. This article closes the gap. By the end of this article you will be able to: 1.   Distinguish with precision between the five tiers of enterprise AI systems — language model, assistant, copilot, agent, and multi-agent system — and explain the distinction to your board in plain language. 2.   Recognise, in any vendor demonstration, whether the product being pitched to you is genuinely agentic or a chatbot marketed as an agent — using a four-question diagnostic you can apply in the meeting itself. 3.   Identify at least three concrete agent use cases realistic for a Caribbean enterprise at your organisation’s scale, drawn from the workflows that are already in production elsewhere in the region.

A chief executive I have known for a long time — the head of one of the larger Caribbean financial services institutions — telephoned me a few weeks ago after a vendor meeting. He had just been shown what was described to him as ‘an agentic AI solution for relationship banking’. The demonstration, he said, had been impressive. The vendor had sophisticated slides, a well-rehearsed presentation, and what appeared to be a working product. His board was interested. He was about to commit meaningful capital to a pilot programme.

But something was nagging at him. He could not say precisely what. He asked me whether I could spare thirty minutes to look at the vendor’s materials before he signed the statement of work. I agreed. The materials arrived that evening.

What the vendor was actually selling, I told him when we spoke the next morning, was a chatbot. A capable one, with some useful integrations, but a chatbot nonetheless. It responded to prompts from bank staff; it did not execute multi-step work. Almost none of the so-called ‘agentic’ capabilities in the sales material were present in the product as sold. Some were on a roadmap. Others were frankly aspirational. The productivity claims in the business case rested on the product having capabilities it did not yet have.

This pattern — the gap between what is called an agent and what is actually an agent — is now so common in the Caribbean technology buying environment that I have decided to devote this entire article to closing it. I do not want my clients, my peers, or the institutions I care about to be misled. I do not want Caribbean boards to commit capital to products that will not deliver what has been promised. And I do not want the current AI cycle to be discredited in our region by deployments that fail for a reason that was preventable — which is, the buyer did not understand precisely what they were buying.

In Article 1 of this series, I argued that the current AI cycle is genuinely different from those that came before, that the productivity evidence is converging on effects in the 20 to 50 percent range, and that the Caribbean’s normal lag pattern will not provide cover this time because four compounding factors — capability build-up, talent scarcity, the widening data gap, and regulatory hardening — will make late adoption meaningfully more expensive than timely adoption. That argument assumes that when a Caribbean executive decides to act on artificial intelligence, they can distinguish what is real from what is marketing. This article provides the specific tools to make that distinction.

The structure of what follows is straightforward. I will take you through the five tiers of AI systems you will encounter in the market, with precise definitions and example products at each tier. I will explain what an agent actually does — the four specific capabilities that together define the category and without which ‘agent’ is just a marketing word. I will walk you through four concrete examples of agents in Caribbean enterprise contexts, so you can see what the abstract definitions look like in practice. I will then give you a four-question diagnostic — the D-AGENTICA™ Agentic Vendor Assessment — that you can apply to any vendor pitch to assess whether what you are being sold is genuinely agentic. And I will close with what this means for the evaluation discipline Caribbean boards should adopt over the next twelve months.

By the time you finish reading, you will be able to walk into a vendor meeting tomorrow and, within five minutes, tell the difference between a product that can do the work and a product that can only talk about the work. That distinction will save your organisation a great deal of money — and a great deal of disappointment — in the next two years.

 

The hierarchy of AI systems, in plain language

Let me start with the most common source of confusion. When an executive says, ‘our organisation already uses AI,’ they almost invariably mean one thing. When a vendor says, ‘our product is powered by AI,’ they almost invariably mean another. When a consultant says, ‘AI will transform your business,’ they almost invariably mean a third. Three different people, using the same word, are describing three different things. This is not a minor semantic problem; it is the reason so many Caribbean AI conversations go in circles.

The fix is to install a clear hierarchy. There are five tiers of AI systems that you will encounter in your market. Each tier has a distinct definition, a distinct capability profile, and a distinct commercial meaning. Once you know the hierarchy, every conversation about AI in your organisation becomes more precise.

 

Tier	Category	What it is	Example products
Tier 1	Language Model	The underlying AI — Claude, GPT-4, Gemini Ultra. Pure capability. Not a product.	Claude Opus 4
Tier 2	Assistant	A model wrapped in a chat interface. Responds to prompts, stops. No memory of tools beyond the conversation.	ChatGPT, Claude.ai
Tier 3	Copilot	An assistant embedded in a specific application with access to that application’s data. Still waits for user prompts but operates inside context.	Microsoft 365 Copilot, GitHub Copilot
Tier 4	Agent	A system that takes a goal, plans, calls tools, executes multi-step work, and returns a finished output. Human at the boundary, not in every step.	Claude Agents, Salesforce Agentforce, Workday AI
Tier 5	Multi-Agent System	Multiple specialised agents coordinating on a shared goal. One agent calls another. Emerging, mostly still in controlled deployment.	Early-stage: research previews from Anthropic, OpenAI, Google

 

I want to draw out three things from this hierarchy that matter for Caribbean executives.

Tier 1 is a capability, not a product

A language model — Claude, GPT-4, Gemini — is not something you buy as an enterprise. It is the underlying intelligence that sits beneath everything else in the hierarchy. When a vendor tells you their product ‘uses Claude’ or ‘is powered by GPT-4’, they are telling you which Tier 1 model sits inside their product. That is useful information, but it does not tell you what tier their product itself operates at. A chatbot built on Claude and an agent built on Claude are different products with different capabilities, even though they share the same underlying model. Do not confuse the engine with the vehicle.

Tier 2 and Tier 3 are what most organisations currently use

When a Caribbean executive tells me ‘we already use AI’, they almost always mean that their staff use ChatGPT (a Tier 2 assistant) or Microsoft 365 Copilot (a Tier 3 copilot). This is a reasonable starting point. It is not, however, agentic AI. The productivity gains I cited in Article 1 — 30 to 40 percent uplifts on professional work — are not being achieved by assistants and copilots alone. They are being achieved by agents. Organisations that have deployed Tier 2 and Tier 3 systems and concluded that ‘AI is mildly useful’ have not yet deployed AI in the form that creates the competitive effects I described in Article 1.

Tier 4 is where the current cycle is creating competitive effects

Almost every significant productivity uplift being reported by credible sources in 2025 and 2026 is being produced by Tier 4 agents, not by Tier 2 or Tier 3 systems. Claude Agents, Salesforce Agentforce, Workday AI, and the emerging agent layers inside SAP, Oracle, and Microsoft are the products that are actually reshaping enterprise economics. Tier 5 — multi-agent systems — is where the research frontier sits today and where the next wave of capability is being built. Most Caribbean organisations will be operating at Tier 4 by 2028. The organisations that are ahead of that timeline will be at Tier 5.

 

The productivity gains are being produced by Tier 4. If your organisation is still at Tier 2, you are not behind on AI. You are not yet on it.

 

What an agent actually does — the four capabilities

The word ‘agent’ is now used so loosely in the market that the term has almost lost its meaning. Any product that uses AI in any form is at risk of being described by its vendor as agentic. This is a problem for buyers, because the word is doing work that the product often cannot. I therefore want to be precise about what an agent actually is — by specifying the four capabilities that, taken together, define the category. A product that demonstrates all four is genuinely agentic. A product that demonstrates only some of them is not.

Capability one — planning

Given a goal, an agent decomposes that goal into a sequence of sub-tasks. ‘Prepare a reconciliation of the January accounts receivable’ is not a single task; it is a goal that requires multiple sub-tasks — pull the accounts receivable ledger, pull the bank statements, match transactions, identify variances, flag anomalies, produce a summary document. An agent works out what these sub-tasks are, in what order they should be done, and which ones can be done in parallel. This is called ‘planning’, and it is the first characteristic that distinguishes an agent from an assistant. An assistant executes whatever single instruction you give it. An agent works out what instructions need to be given, and then gives them to itself.

Capability two — tool use

An agent executes its plan by calling software tools — the accounting system, the document repository, the customer relationship management platform, the tax authority’s API, the email system, whatever is required. Each tool call is a concrete action in the real world: data is retrieved, a record is updated, an email is drafted, a document is generated. The agent chooses which tools to call, in what order, with what inputs. The agent then reads the outputs of each tool call and uses them to inform the next step. This is called ‘tool use’, and without it an agent is not doing any work in the world — it is producing text about what work could be done. A product that cannot call external tools is not a Tier 4 agent, regardless of what it is called.

Capability three — memory

An agent remembers what it has already done within the current task. It remembers what tools it has called, what outputs those tools returned, what intermediate conclusions it has drawn, what steps remain. A system without memory has to start over with each prompt; a system with memory can execute a forty-step process in a single goal statement from the user. Note that memory here is not the same as persistent memory across different tasks — that is a separate capability. What I mean by memory is the ability to hold the state of a single multi-step task in working context from start to completion. Without this, tool use produces chaos rather than work.

Capability four — self-correction

When an agent takes an action that does not produce the expected result — an API call fails, a piece of data is missing, a calculation produces an anomaly — the agent notices and tries a different approach. It reassesses its plan. It retries with different inputs. It escalates to a human if it cannot resolve the problem. This is self-correction, and it is what turns an agent from an impressive demo into a reliable production tool. A product that cannot handle its own failures will not survive in a Caribbean enterprise environment, where the integration boundaries are messier and the exception rates are higher than in the more polished environments where most agents are first developed. Self-correction is the capability that separates a product that works in a controlled demonstration from a product that works in the real world.

These four capabilities — planning, tool use, memory, and self-correction — are the minimum for a product to be genuinely agentic. Not all four present means not an agent. Let me repeat that, because it is the single most important sentence in this article. If a product cannot plan, cannot call external tools, cannot remember what it has done, or cannot recover from its own failures, it is not an agent, regardless of how it is marketed.

 

Planning, tool use, memory, self-correction. If all four are not present, you are not looking at an agent, regardless of how it is marketed.

Four agent examples in a Caribbean context

Abstract definitions are necessary but insufficient. Let me make this concrete with four examples of agents that are now operating, or could be operating within the next eighteen months, inside Caribbean organisations. Each example shows how the four capabilities I just described play out in a real workflow.

Example one — the audit workpaper agent

This is an agent that prepares substantive testing workpapers for an external audit engagement. Given a goal — ‘prepare the accounts receivable confirmation testing workpapers for the December year-end audit of Client X’ — the agent plans the work: identify the customer balances to be confirmed, retrieve the customer master data from the client’s system, draft the confirmation letters, generate the accompanying schedules, prepare the statistical sampling documentation, and produce the audit programme coverage mapping. It calls the audit software (CaseWare or equivalent) to pull templates, calls the client’s ERP to pull balances, generates PDFs of draft letters, and returns a complete working paper file for senior auditor review. When a customer balance fails a data quality check, it notices, flags it, and proposes a remediation path. A human senior auditor reviews the output, signs off on the approach, and the confirmation process executes. This is not a theoretical example. Agents of this kind are now in production inside Big Four firms and are beginning to appear in the mid-tier audit market, including at Dawgen Global.

Example two — the GCT compliance agent

This is an agent that prepares monthly General Consumption Tax compliance filings for a Jamaican enterprise — or equivalent filings across the VAT regimes of other Caribbean territories. Given the goal ‘prepare the January 2026 GCT return for Client Y’, the agent retrieves the sales data from the accounting system, classifies each transaction by its GCT treatment (standard-rated, zero-rated, exempt), calculates the output tax, retrieves the purchase data and calculates the input tax credits, prepares the reconciliation schedules, drafts the GCT return in the format required by the Tax Administration Jamaica system, and produces the working papers supporting the return. When a transaction has an ambiguous GCT classification, the agent flags it for human review rather than guessing. The tax professional reviews the return, signs off, and the submission proceeds. What took a week of manual work now takes an hour of review. This is the specific kind of workflow where agents are creating real competitive advantage for Caribbean firms that have deployed them.

Example three — the credit union member-service agent

This is an agent that handles routine member service queries for a Caribbean credit union. When a member calls or emails with a request — ‘what is my current loan balance?’, ‘can I make an additional payment?’, ‘when is my next direct debit scheduled?’ — the agent retrieves the member’s account information from the core banking system, answers the query directly where possible, executes the requested action where authorised (scheduling a payment, generating a statement, sending a confirmation), and escalates to a human member-service officer for anything that requires judgment or exceeds its authority (loan restructuring requests, dispute handling, anything involving financial advice). The agent operates within clearly defined boundaries — what it may and may not do, what must be escalated, what must be logged for compliance review. Human member-service officers move from handling routine queries to handling the exceptions the agent escalates, which is the work their training actually equipped them for. This example demonstrates the capability-and-boundary principle I will return to later in this series: the agent extends human capacity, it does not replace human judgment.

Example four — the Virtual CFO agent

This is an agent that supports a small-or-medium Caribbean enterprise in place of a full-time chief financial officer. Given the goal ‘prepare the monthly management accounts for Client Z’, the agent retrieves transaction data from the accounting system, reviews the trial balance, proposes and posts routine adjusting journal entries (prepayments, accruals, depreciation), prepares the management accounts in the format the client uses, generates the KPI dashboard, produces a brief written commentary on variances against budget, and identifies the two or three items that the client’s owner-manager should focus on in the monthly finance conversation. The agent works alongside a part-time professional accountant who reviews the output, challenges the agent’s conclusions where warranted, and takes the client meeting. The result is a level of finance function sophistication that small Caribbean enterprises have historically not been able to afford. This is, in my view, the single most economically important agent use case for the SME layer of the Caribbean economy — because it democratises access to senior finance expertise in a way that could materially change the operating quality of the region’s SME sector if it is adopted at scale.

 

WHAT WE ARE ALREADY SEEING IN THE CARIBBEAN Across the last eighteen months, Dawgen Global’s advisory practice has observed early-wave agent deployments producing documented results that align with the global evidence base. In professional services, audit workpaper preparation cycles have been compressed by 40 to 55 percent where the workpaper agent pattern has been properly implemented. In finance operations for mid-sized Caribbean enterprises, month-end close has moved from ten to fifteen working days to three to five working days with agent-augmented processes. In credit union member service, routine-query resolution time has dropped from minutes to seconds, allowing human staff to concentrate on the member situations where judgment actually matters. These are not theoretical claims. They are what is happening, in our region, right now, in organisations whose leadership made the decision to act early.

A four-question diagnostic for your next vendor meeting

Now that you know the five tiers and the four capabilities, you can apply them. The D-AGENTICA™ Agentic Vendor Assessment is a four-question diagnostic designed for use in the vendor meeting itself — not afterwards, not during procurement review, but in the conversation where the vendor is pitching to you. Each question probes one of the four capabilities that together define an agent. Ask them in sequence. Listen carefully to the structure of the answers.

If the vendor gives confident, specific answers to all four questions, the product is probably genuinely agentic and worth further evaluation. If the vendor hedges, redirects, or answers a different question, the product is probably not yet what it is being described as — and your organisation is at risk of paying Tier 4 prices for a Tier 2 capability. The discipline of applying this instrument in the meeting itself will save your organisation substantial amounts of money and avoid substantial amounts of reputational exposure over the coming two years.

 

A NAMED INSTRUMENT The D-AGENTICA™ Agentic Vendor Assessment Four questions a Caribbean executive can apply to any AI vendor pitch to assess whether the product being sold is genuinely agentic — or a chatbot that has been rebranded for marketing purposes. Ask them in the vendor meeting. The pattern of answers will tell you, within five minutes, what you are actually being sold. QUESTION 1 Can you show me a task your system completes end-to-end, not one it merely assists with? Signal of a genuinely agentic product: The vendor demonstrates a full task: a trigger, a plan, multiple tool calls, intermediate checks, and a completed output. They can point to the artefact the system produced. Signal it is not: The vendor shows you a conversation. The system responded to prompts. You the user did the rest of the work. That is a chatbot, regardless of how it is marketed. QUESTION 2 What tools and systems does the agent actually call, and what happens when a call fails? Signal of a genuinely agentic product: A specific named list: the ERP, the CRM, the document repository, the email system. A clear account of retry logic, fallback behaviour, and human escalation when calls fail. Signal it is not: Vague references to ‘integration capabilities’ or ‘plug-and-play connectors’. No concrete system list. No clear account of failure handling. This is a product that is not yet in production. QUESTION 3 Where is the human in the loop, and where is the human at the boundary? Signal of a genuinely agentic product: A clear risk-tiered policy: low-risk tasks the agent completes autonomously, medium-risk tasks where a human approves before execution, high-risk tasks where a human reviews the plan before the agent acts. Signal it is not: Either ‘the human is always in control’ (which means a chatbot, not an agent) or ‘the system runs autonomously’ (which means no governance model, which means unfit for any regulated Caribbean environment). QUESTION 4 Show me a production deployment of this agent in a regulated environment. Not a demo, not a pilot — production. Signal of a genuinely agentic product: A named client, a specific workload, an operating period measured in months. Ideally, someone you can call. Case study discipline similar to what audit firms provide as evidence. Signal it is not: A ‘flagship pilot’, ‘innovation partnership’, or ‘select early access’. These are euphemisms for ‘not yet in production’. A vendor selling production capability has production references.

A practical note on using this instrument. The vendor will not necessarily be attempting to deceive you. Many vendors genuinely believe their product is more agentic than it is — partly because the industry vocabulary is so loose that even well-intentioned teams use the word loosely, and partly because roadmap features are often discussed as though they are delivered features. The point of the diagnostic is not to catch a vendor in a lie. The point is to establish, with precision, what the product actually does today — so that your business case is based on delivered capability and not on a mixture of delivered and promised capability.

 

A CARIBBEAN VENDOR PATTERN WORTH RECOGNISING In a recent advisory engagement with a Caribbean commercial bank, we ran the four-question diagnostic across eight AI vendor pitches the bank was considering. Four of the eight vendors passed all four questions with clear, production-grade answers. Three passed two of the four questions but hedged materially on tool integration or self-correction. One passed none of the four — their ‘agent’ was a well-dressed chatbot. The bank’s procurement team had previously been weighing all eight vendors on a similar ‘capability score’ because their evaluation framework did not distinguish between the tiers. After the diagnostic, the shortlist of genuinely agentic products was four, not eight. The bank avoided committing meaningful capital to capabilities that did not exist. This is the specific kind of advisory work the discipline in this article is designed to enable, whether or not our firm is in the room.

 

What agents do not do — the boundaries

I want to be equally precise about what agents do not do, because overstating the current state of the technology is the fastest way to lose credibility with a Caribbean board. An agent that has been oversold and then underdelivers will discredit the technology category for that organisation for years. I would prefer Caribbean boards to commit to agentic AI with accurate expectations — even if those expectations are more modest than the marketing claims — than to commit with inflated expectations that are almost certain to be disappointed.

Agents do not replace professional judgment

An agent cannot replace the judgment of a qualified accountant, auditor, lawyer, doctor, or senior executive. What an agent does is compress the execution of the work those professionals would have done — the preparation, the cross-referencing, the drafting, the calculation — so that the professional’s judgment is applied to the output rather than expended on the preparation. A firm that tries to use an agent to replace professional judgment is misusing the technology. A firm that uses an agent to amplify professional judgment is using it well. This distinction matters because it determines where the human oversight sits in the workflow and where the professional liability attaches.

Agents are not reliable without controls

Agents, like all AI systems, produce outputs that are sometimes wrong. They occasionally miscalculate. They occasionally retrieve the wrong record. They occasionally draft text that sounds correct but is factually inaccurate. In a controlled environment with appropriate human review — the audit senior reviewing the workpapers, the tax professional reviewing the GCT return, the member-service officer reviewing the escalated cases — these errors are caught before they cause harm. In an uncontrolled environment with no human review, these errors accumulate. The agent is a productivity tool, not a judgment-replacement tool, and the organisation that forgets this distinction will pay for the lesson.

Agents are not yet appropriate for the highest-risk decisions

In 2026, for a Caribbean enterprise operating under regulatory supervision, the responsible deployment pattern is to use agents for preparation and execution of well-defined, reviewable work — not for final decisions that affect customers, employees, regulators, or financial outcomes. Credit decisions, hiring decisions, medical decisions, disciplinary decisions, strategic investment decisions — these remain human decisions. An agent may prepare the analysis. A human must make the call. This is not because the agent will necessarily make a worse decision. It is because the accountability structure in which your organisation operates — legal, regulatory, fiduciary, reputational — is built around human decision-making, and the governance model for agent-made high-stakes decisions has not yet been established in Caribbean law. Until it has, caution is the correct posture.

Agents require infrastructure you may not yet have

An agent works against your organisation’s data, documents, and systems. If your data estate is scattered, poorly classified, or inaccessible via APIs, the agent will not be able to do useful work — and no amount of vendor magic will compensate for a data estate that is not ready. This is why the D-AGENTICA™ Maturity Model, which we will unveil in Article 11, weights data readiness and infrastructure at 25 percent of the overall assessment. If your organisation is weak on these two dimensions, the sensible first investment is not in agent licences. It is in the foundations that would make agent deployment successful.

 

What this article has established, and what comes next

This article has done three things. It has installed a five-tier hierarchy of AI systems that you can use to cut through the vocabulary confusion in the Caribbean technology market — language model, assistant, copilot, agent, multi-agent system. It has specified the four capabilities that together define an agent — planning, tool use, memory, and self-correction — and made clear that a product missing any of these is not an agent regardless of how it is marketed. It has walked through four concrete agent use cases in Caribbean contexts that show what the abstract definitions look like in real workflows. And it has equipped you with the D-AGENTICA™ Agentic Vendor Assessment, a four-question diagnostic you can apply in any vendor meeting to assess what the product being sold to you actually is.

Taken together with the D-AGENTICA™ Three-Question Board Diagnostic from Article 1, you now have two instruments. The first tells you where your organisation stands. The second tells you whether the products you are being offered will actually move that position. These two instruments will do more, in practical terms, to improve the quality of your organisation’s AI decisions in the next twelve months than any piece of external advisory work you could commission, because they are applied in the room where the decisions are actually being made.

Next week’s article shifts from the question of what AI is to the question of what AI costs — and what it returns. In my experience, the single most common reason Caribbean boards hesitate on AI is not scepticism about what the technology can do. It is uncertainty about what the economics actually look like at Caribbean small and medium enterprise scale, where budgets are measured in hundreds of thousands rather than tens of millions of dollars, and where the case studies published by global consultancies do not translate directly. Article 3 will address this gap. It will present a specific ROI model built for Caribbean SME economics, walk through three anonymised SME vignettes with honest cost-and-benefit numbers, and introduce the D-AGENTICA™ SME AI Sequencing Framework — a five-step sequence that the Caribbean firms deploying agents successfully are following, whether they know it or not. By the time you finish Article 3, you will be able to tell whether an AI investment case presented to your board is numerically serious or numerically aspirational.

If there is one reflection you take from this article into your next executive meeting, let it be this. The single most expensive thing your organisation can do in the next twelve months is to commit capital to an AI product under the mistaken belief that it is a different tier of capability than it actually is. The language matters. The tier matters. The four capabilities matter. Know what you are buying. That discipline alone will separate the Caribbean organisations that get this cycle right from those that get it expensively wrong.

 

FOR THE BOARD AGENDA This article has specified how to tell a genuinely agentic product from a chatbot that has been rebranded as an agent. A board chair or audit committee chair reading this article has earned the right to ask their leadership team one specific question, and to propose one specific decision that will materially improve the quality of the organisation’s AI procurement over the next twelve months. THE QUESTION For each of the AI vendors currently being evaluated by our organisation, can management demonstrate, using the D-AGENTICA™ Agentic Vendor Assessment, that the product is genuinely agentic and not a chatbot marketed as one — and if not, can the business case be re-examined against what the product actually does? THE DECISION That no AI vendor contract above a materiality threshold set by the board will be executed without the D-AGENTICA™ Agentic Vendor Assessment being formally applied, with the results documented in the procurement file, and that the board will receive a summary of these assessments as part of its quarterly governance reporting.

 

THE CARIBBEAN AI ADOPTION IMPERATIVE A 12-Article Series from Dawgen Global NEXT IN THIS SERIES Article 03 — The Productivity Economics What the return actually looks like below US$50m in revenue MEASURE YOUR ORGANISATION’S AI READINESS Request the free D-AGENTICA™ AI Maturity Self-Assessment Email : [email protected]

About Dawgen Global