AI Agents Aren't Employees: Building an AI Workforce That Actually Works

There is a slide that has appeared in a great many presentations over the past two years. It shows an organisation chart in which some boxes contain people and others contain AI agents with job titles: AI Sales Development Representative. AI Finance Analyst. AI Customer Success Manager.

The slide is persuasive because it maps something unfamiliar onto something familiar. It is also misleading, for the same reason.

An employee arrives with judgement, context, accountability, an understanding of what is not written down, and the ability to notice when an instruction no longer makes sense. An AI agent arrives with none of those by default. It has a model, a set of instructions, whatever tools and data it has been connected to, and a tendency to produce confident output whether or not the underlying task went well.

That does not make agents useless. Used well, they are among the most practical productivity tools a smaller business can adopt. But the organisations getting value from them are not hiring digital employees. They are designing systems with clear responsibilities, constrained permissions, reliable information and human checkpoints.

Three things that are often called "AI agents"

Much of the confusion begins with terminology. At least three different things are commonly sold under the same label.

AI assistant Workflow automation Autonomous agent with AI steps

What it does              Responds to a person's     Follows a predefined      Decides its own
                          request: drafts,           sequence; AI handles      sequence of steps and
                          summarises, analyses       specific steps such as    tool calls to reach a
                                                     classification or         goal
                                                     extraction

Who controls the flow The person The workflow designer The model, within limits

Predictability High (a person reviews High Variable every output)

Best for                  Knowledge work with a      Repetitive, well-         Open-ended tasks
                          human in the loop          understood processes      where steps can't be
                                                                               scripted in advance

Typical failure           Inaccurate content         Unhandled edge cases      Wrong actions taken
                          accepted without                                     confidently, cost
                          review                                               overruns, loops

A useful rule follows from this table: use the least autonomous design that solves the problem. Many tasks described as "agentic" are better handled by a workflow with one or two AI steps, which is cheaper, easier to test and far easier to audit.

Industry analysts have made the same point. In a June 2025 forecast, Gartner predicted that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It also warned of "agent washing": existing chatbots, assistants and robotic process automation tools being rebranded as agents.

Demonstration versus deployment

Agent demonstrations are usually impressive. An agent reads an inbox, researches a prospect, drafts a tailored email and updates the CRM, all in under a minute.

What the demonstration rarely shows is the second hundred runs. In operation, agents encounter incomplete records, ambiguous requests, conflicting data, documents in unexpected formats and customers who say one thing and mean another. Each run that fails quietly — a wrong field updated, a misread quantity, an email sent to the wrong contact — creates cleanup work that may cost more than the time saved.

The gap between demonstration and deployment is mostly made up of unglamorous engineering:

• Defined inputs and outputs, so the agent knows exactly what it receives and what "done" looks like.

• Exception handling, so uncertain cases are routed to a person rather than guessed.

• Testing on real historical cases, not curated examples.

• Monitoring, so failure rates, costs and drift are visible over time.

A useful planning assumption is that building the demonstration is perhaps a fifth of the work of building something a business can rely on.

What a well-defined agent role looks like

If the employee analogy is worth keeping at all, it is for one reason: good employees have job descriptions, and so should agents. A practical agent specification answers six questions.

• 1. Responsibility. What single outcome is this agent accountable for? "Categorise and route inbound customer enquiries" is a responsibility. "Handle customer service" is not.

• 2. Inputs. What information does it receive, from where, and in what format?

• 3. Tools and permissions. What systems can it read? What, if anything, can it change? Can it send external communications?

• 4. Knowledge. Which documents, records or databases does it rely on, and who keeps them accurate?

• 5. Escalation. Under what conditions must it stop and hand off to a person?

• 6. Measures. How will accuracy, speed, cost and business value be tracked?

The third question deserves particular attention. Security frameworks treat over-permissioned agents as a distinct risk category. The OWASP Top 10 for LLM Applications (2025) lists excessive agency — giving AI systems more functionality, permissions or autonomy than they need — alongside prompt injection,

in which instructions hidden in emails, documents or web pages manipulate a model's behaviour. An agent that can read customer emails and also issue refunds is an agent that can, in principle, be instructed by a malicious email to issue refunds.

The safest principle is the oldest one in information security: least privilege. Agents should be able to read what they need, write only where necessary and act externally only with approval.

How agents coordinate, and where they shouldn't

Multi-agent systems, in which several specialised agents pass work between them, are an active area of development. A common pattern uses an orchestrating agent that breaks a task into parts and assigns them to specialists: one researches, one drafts, one checks.

There are good reasons for this design. Specialised agents with narrow instructions and limited tools tend to perform more reliably than one agent asked to do everything. Separation also makes permissions easier to control.

But coordination introduces its own failure modes. Errors compound as work passes between agents. Costs multiply as each agent consumes its own processing. Diagnosing failure becomes harder when no single component is obviously wrong. For most smaller organisations, the most robust "coordination layer" is still a shared system of record — the CRM, the ticketing system, the job management platform — where each agent reads and writes, and where people can see what happened.

In practice, agents should coordinate through structured handoffs: a record changes status, a ticket is created, a document is placed in a defined folder. Free-form conversation between agents is better kept to experiments.

Human approval checkpoints

The question is not whether people should be involved, but where. A useful way to place checkpoints is by the reversibility and external impact of an action.

Action type Example Suggested control

Internal, reversible, low impact Tagging a support ticket; Automate; sample-check drafting meeting notes periodically

Internal, material Updating a customer record; Automate with logging; review changing an inventory status exceptions

External communication              Sending an email or quotation       Human approval before
                                    to a customer                       sending, at least until accuracy
                                                                        is proven

Financial or contractual            Issuing refunds, approving          Human decision; AI prepares
                                    payments, committing delivery       the recommendation
                                    dates

Irreversible or regulated Deleting records; regulatory Human only submissions

Controls should be relaxed deliberately, based on measured performance, rather than by default because the agent "seems to be doing fine."

Privacy, security and accountability

For any business handling customer, employee or commercial data, a few basics are non-negotiable:

• Know where data goes. Understand which AI providers process your data, under what terms, and whether it is retained or used for training.

• Keep audit logs. Record what each agent received, what it did, which tools it called and what it produced. Without logs, there is no meaningful accountability.

• Assign a human owner. Every agent should have a named person responsible for its configuration, performance and errors.

• Separate credentials. Agents should use their own service accounts with scoped permissions, not a manager's login.

Governance guidance is maturing. Singapore's Infocomm Media Development Authority launched a Model AI Governance Framework for Agentic AI in January 2026. It is organised around bounding risks upfront, keeping humans meaningfully accountable, implementing technical controls and enabling end- user responsibility. The framework is guidance rather than law, but it is a practical reference for organisations designing agent deployments, wherever they operate. The US NIST AI Risk Management Framework offers a broader foundation for AI risk.

Measuring cost, reliability and value

An AI workforce should be measured like any operational investment. Four measures matter most:

• 1. Task accuracy. The percentage of runs completed correctly without human correction, measured on real work.

• 2. Exception rate. How often the agent escalates, and whether those escalations were appropriate.

• 3. Total cost per task. Model usage, software subscriptions, integration maintenance and human review time. The last is often forgotten and frequently decisive.

• 4. Business outcome. Response times, conversion rates, invoice cycle time, hours released for higher-value work.

If an agent saves ten minutes per task but every output needs eight minutes of review, the true saving is two minutes, not ten.

A sensible starting architecture for a small company

Consider an illustrative twelve-person industrial services business with a sales team, an operations coordinator, a part-time finance function and no dedicated IT staff. A realistic first architecture might look like this.

Foundation: one system of record per domain. CRM for customers and opportunities; a job or project system for operations; accounting software for finance. Agents are added around these systems, not in place of them.

Layer 1 — Assistants for staff (low risk, fast value).

A secure AI assistant with access to a curated internal knowledge base: product information, pricing rules, standard terms and procedures. Staff use it for drafting, summarising and answering internal questions.

Layer 2 — Two or three narrow workflow agents.

• An enquiry triage agent that reads inbound emails and web forms, classifies them, extracts key details and creates or updates a CRM record. A person assigns follow-up.

• A document extraction agent that reads RFQs, purchase orders or supplier invoices and pre-fills structured fields for review.

• A meeting follow-up agent that turns call notes into action items in the CRM.

Layer 3 — Human checkpoints. All customer-facing messages are drafted, not sent. All financial entries are proposed, not posted.

Governance. One named owner, a simple log of agent actions, a monthly review of accuracy and cost, and a written list of what each agent may and may not do.

This architecture is not ambitious. That is the point. It can be built with commercially available tools, maintained without an in-house engineering team and expanded as confidence grows.

A phased approach

Phase Focus Exit criterion

1. Assist (months 0–3)              AI assistants with curated           Staff use tools regularly;
                                    knowledge; staff training; data      knowledge base is accurate and
                                    clean-up                             owned

• 2. Automate steps (months 3– One or two narrow workflow Measured accuracy on real 6) agents with full human review work at an agreed threshold

3. Reduce review (months 6–         Move proven low-risk actions to      Stable accuracy and cost over
12)                                 sample-based review; add             several months
                                    exception handling

4. Extend carefully (12             Additional agents; limited           Clear business case and
months+)                            coordination through shared          governance for each new agent
                                    systems; selective autonomy

The timeline is illustrative. Some businesses will move faster; many will find that Phase 1 reveals data problems that must be solved first.

What is real now, and what is still speculative

It is worth being clear about the boundary.

Reliable today, with good design: drafting, summarising, classification, information extraction, retrieval from well-maintained knowledge bases, and narrow multi-step workflows with defined tools and human review.

Workable with care: agents that take limited actions in business systems, coordinated through structured handoffs, with monitoring and escalation.

Still speculative for most businesses: fully autonomous agents managing open-ended responsibilities over long periods, negotiating commercially on a company's behalf, or replacing roles that depend on judgement, relationships and accountability.

Capabilities are improving quickly, and that boundary will move. But businesses are better served by building on what works reliably now than by designing around what might work later.

Conclusion: design the work, not the job title

The organisations that will benefit most from AI agents are not those with the most agents or the most autonomous ones. They are the ones that understand their own work well enough to break it into clear responsibilities, connect those responsibilities to trustworthy data, and decide deliberately where human judgement remains essential.

That is less a technology project than an operating-model project. The technology is the easier part.

Omnidite Advisory helps businesses connect operational processes, software platforms and automation, including designing practical AI-enabled workflows that fit the resources a company actually has.

This article provides general information and commercial commentary, not legal, investment, regulatory, medical or certification advice. Illustrative examples and calculations are not actual customer engagements or quotations.