The Race to Replace: Why AI Needs to Start Doing the Work

For more than two decades, one of the world's most successful internet business models has been built around the same simple transaction. Someone has a question. They type it into a search engine. The search engine shows results — including advertisements. The user clicks through to a website. A business wins their attention.

Generative AI is already disrupting this. Users increasingly ask an AI system directly rather than searching, reading, comparing and then acting. Instead of searching for hotel recommendations, reading reviews and making a booking through several separate steps, an AI system can potentially understand requirements, compare options and in some cases complete the transaction.

But that disruption — significant as it is — may be only the first stage. The more consequential shift may come not from AI answering questions better, but from AI taking responsibility for the work that follows the answer.

The Quick Answer

Why Search Was Such a Powerful Business Model

To understand what AI agents could disrupt, it helps to understand what the internet model they are challenging actually is.

The internet created a fundamental problem: enormous information existed, but users needed help finding it. Search engines solved discovery. Businesses paid to appear at the point of that discovery — buying the attention of someone whose intent was already known. This created one of the most efficient advertising models ever built.

The chain was: user intent → search → advertisement or result → website → business action (purchase, booking, enquiry, subscription). Each link in that chain was a commercial opportunity.

Generative AI compresses parts of this chain. Instead of searching, reading and comparing, a user may ask an AI system directly and receive a synthesised answer. The steps between question and answer collapse.

But agentic AI may compress it further still. The chain becomes: user states objective → AI agent → action completed. The biggest disruption may not be fewer search clicks. It may be fewer steps between intention and transaction.

Why AI Search Alone May Not Be Enough

AI companies have several viable revenue models: subscriptions, enterprise licences, advertising, API consumption, transaction fees, commerce and the ability to charge for autonomous workflow execution. Advertising may remain valuable. Enterprise licences are already substantial. The existing models are not broken.

The argument is subtler than that. An extraordinarily expensive global computing platform — one requiring enormous investment in data centres, GPUs, electricity, networking, model development and engineering talent — ultimately needs to create more value than generating slightly better search results.

The economic case for current levels of AI investment depends significantly on AI eventually doing things that search cannot: performing real work, automating substantive processes and delivering measurable productivity. That is the pressure behind the transition from AI assistance to AI agency.

The Infrastructure Economics Behind AI

Building the infrastructure that makes this possible is a significant undertaking. AI requires major investment across data centres, specialised processors, electricity, cooling, networking, model development and the engineering teams to build and operate it. The economics of that investment create pressure for AI to do economically useful work — not merely make existing processes marginally faster.

The level of investment currently flowing into AI infrastructure — from hyperscalers, from governments and from private capital — reflects a belief that AI agents performing real business tasks are a genuine near-term possibility, not merely a speculative future.

What Is an AI Agent?

There is a meaningful difference between a chatbot, a copilot and an agent. The distinction matters because they represent different economic models and different levels of governance risk.

Chatbot

You ask. It answers. The conversation ends. The human then decides what to do with the answer and does it themselves.

Copilot

You perform the work. The AI assists in real time — suggesting, completing, translating, checking, drafting. The human remains in control of each step.

Agent

You define an objective. The AI may perform several steps using available tools and systems to complete it. The human may not be involved in each individual step.

Compare: a chatbot you ask to write an email reminding a customer about an overdue payment. An agent might instead review all overdue accounts, identify customers requiring follow-up based on defined criteria, draft suitable emails for each, update the CRM and present anything unusual for human approval before sending.

Agents may reason, plan, retrieve information, use applications, call external APIs, create or modify files, update records, monitor outcomes, retry failures, escalate exceptions and return completed work. The human supervision requirement shifts from each individual step to the objective and the exceptions.

Chatbot, Copilot and Agent: A Plain Comparison

Current Evidence of the Shift

OpenAI has reported that agentic AI usage in its research is moving users from short AI interactions towards longer delegated tasks — users assigning objectives rather than questions. The company describes this as an emerging pattern of AI handling complex, multi-step work with reduced per-step human involvement.

Google Cloud similarly describes what it calls the 'agentic enterprise' — organisations redesigning operations around AI agents and human experts working together. Its research frames agents not as experimental features but as a coming operational model for enterprise knowledge work.

The UK Competition and Markets Authority published research in 2025 and 2026 examining consumer-facing agentic systems capable of acting on behalf of users — booking, purchasing, managing preferences — and the significant questions this raises about competition, accountability and consumer protection.

These are evidence of direction, not proof of outcome. The scale and speed of any labour-market effect remains genuinely uncertain. What the research does confirm is that the shift from AI answering questions to AI completing tasks is being taken seriously by major AI providers, enterprise technology companies and governments.

Jobs Are Bundles of Tasks

The most useful reframe for understanding AI's effect on work is this: jobs are not indivisible units. They are collections of different tasks, each with different characteristics, different automation potential and different timelines.

Consider what an accountant actually does across a working week: collecting documents from clients, reconciling records from multiple sources, investigating anomalies, preparing reports and financial statements, interpreting results in the context of the client's situation, speaking to clients and providing professional judgement. Some of these tasks are already highly AI-assisted. Others require professional judgement, accountability and relationship that are substantially harder to delegate.

A solicitor receives documents, creates chronologies, conducts legal research, drafts documents, analyses evidence, advises clients, negotiates on their behalf and represents them in proceedings. A salesperson researches prospects, updates CRM records, prepares outreach messages, arranges meetings, understands customer needs, negotiates and builds the relationships that enable deals to close.

An IT support engineer reads tickets, collects diagnostic information, checks alerts, runs diagnostic scripts, investigates, communicates with the user, decides on the appropriate remediation, fixes the problem and escalates what they cannot resolve. Each of these tasks has a different profile. Some are already automated. Some are close to being automatable. Some require the engineering judgement that only comes from experience with the specific environment.

Different tasks have very different automation potential. Understanding which tasks within a role are automatable, and on what timeline, is more useful than asking whether an entire job will be replaced.

The Race to Replace Tasks: A Practical Framework

Not all tasks are equally exposed. This framework describes six types of work and their approximate automation profile.

From People Doing Work to People Managing Work

The shift from AI as a tool to AI as a worker changes what some roles involve. Today, one employee performs ten tasks. In an increasingly agentic model, the same employee might delegate six of those tasks to AI agents, personally handle the remaining four — particularly those involving judgement, relationships and accountability — and manage exceptions from the delegated work.

If this model becomes common, several things may follow. Output per employee may increase. Some teams may become smaller while maintaining similar or greater capacity. The nature of remaining human work may shift upward towards judgement, oversight and client relationships. Junior administrative tasks may reduce. The skills required to manage effectively — including managing digital work — may change significantly.

None of this is inevitable. The speed of transition will vary enormously by industry, organisation, task type and the actual reliability of the AI systems involved. But the direction of travel is already visible in forward-looking organisations.

One Person, Several AI Workers

An emerging operating model places one capable person at the centre of several AI agents, each handling a different category of work. A commercial manager might supervise a research agent that monitors markets and compiles briefings, a sales preparation agent that researches prospects and prepares outreach, an admin agent that manages scheduling and document processing, and a reporting agent that collates data and produces performance summaries.

The person sets priorities, approves consequential actions, handles exceptions that the agents escalate and applies the judgement and relationship skills that make those capabilities commercially valuable.

This is already a working model in some technology-forward businesses. It is not yet the norm. The reliability, tool access and governance infrastructure required to run it well are still maturing. But the economics of the model — one person leveraging several AI workers — are compelling enough to explain the investment.

The Junior Work Problem

There is an uncomfortable dimension to the automation of tasks that deserves honest discussion. Many professional careers traditionally begin with the very work AI is most likely to automate first.

Junior professionals in law, accountancy, finance, consulting, marketing and many other fields traditionally learn by performing research, preparing documents, entering data, creating first drafts, carrying out routine analysis and handling the administrative work that senior professionals delegate. These are precisely the task types — information retrieval, content preparation, structured processing — where AI assistance is already strong and where agentic automation is advancing most quickly.

This does not mean junior roles will disappear universally or quickly. But organisations that automate entry-level work without redesigning how people develop expertise may find themselves with a skills pipeline problem in five or ten years. The organisations that handle this well will probably redesign graduate and early-career roles deliberately — combining earlier client exposure, supervised AI work, simulation and structured coaching rather than assuming development happens through volume of routine tasks.

Why Full Job Replacement Is Harder Than It Sounds

For all the genuine progress, there are substantial constraints on how quickly AI agents can replace entire roles rather than automate individual tasks.

Jobs involve exceptions, incomplete information, responsibility, legal accountability, physical activity, interpersonal relationships, tacit knowledge accumulated over years and organisational context that is rarely documented. They also involve the ability to handle novel situations — the category of problem that was not anticipated when the process was designed.

AI agents also introduce new work of their own. Errors require investigation. Hallucinations need catching. Security risks need managing. Permissions need reviewing. Output needs checking. Failures need cleaning up. The net productivity gain from an AI agent depends on whether the cost of managing and checking it is genuinely lower than the cost of the original human work.

The Checking Problem

AI may produce output quickly. But someone may still need to verify it, correct it, approve it, investigate anomalies and explain the results to a client or colleague who holds the organisation accountable.

This is the condition that makes AI agents genuinely valuable rather than merely impressive. If the cost of checking approaches the cost of the original task, the productivity gain disappears. If the agent produces output that requires complete human recreation, the gain is negative. Agents improve this equation only when they can reliably execute tasks, check intermediate steps themselves, identify exceptions before a human has to find them, provide evidence of their reasoning and escalate with appropriate context.

Current AI agents are inconsistently reliable at this. Some tasks — structured data processing, well-defined document production, routine communication from templates — are already viable for autonomous execution with manageable checking cost. Others require more human supervision than the productivity gain justifies. Identifying which is which, for a specific task in a specific organisation, is where AI strategy currently requires the most careful thought.

The AI Productivity Ladder

A useful way to think about where AI creates value is to consider five levels of AI engagement, each representing more delegation and more economic value — with correspondingly more governance responsibility.

What Happens to Software?

Today, the standard model for using business software is: a human interacts with an interface, navigates through screens, enters data and submits requests. The software validates, processes and stores the result.

In an agentic model, the interaction may look very different. The human states an objective to an AI agent. The agent calls the relevant APIs, updates the relevant records, queries the relevant databases and returns a completed outcome. The human never touches the software interface directly.

This shifts where software value lives. User interfaces become less important for a class of automated tasks. What matters instead is the quality of the API, the reliability of the data, the clarity of business rules and the robustness of the permissions model. Software that is difficult to integrate with — that has poor API coverage, inconsistent data models or complex permissions — will be poorly served by AI agents. Software that is well-structured for programmatic access will be more valuable in an agentic world.

What Happens to Search?

Search is unlikely to disappear. But it may move behind the agent rather than remaining in front of the human. Instead of a person deliberately opening a browser and typing a query, an AI agent performs the search as a step within a larger task.

Search becomes infrastructure rather than destination. This does not mean search revenue disappears — search providers may charge AI platforms for high-quality index access in ways that are different from but not necessarily smaller than traditional advertising revenue. But the user experience and the commercial model around that experience will both change substantially.

What Happens to Advertising?

Traditional digital advertising depends heavily on human attention — on a person seeing an advert, forming an impression and choosing to act. When an AI agent selects a product or service on behalf of a user, the conventional advertising impression may not occur.

This could shift competition towards being machine-readable — having structured product information that agents can evaluate, strong reputation in sources agents consult, competitive pricing and availability, and agent recommendations from trusted review platforms. The question 'How do we rank in Google?' may eventually be joined by 'How do we perform in AI agent comparisons?'

Advertising will not simply disappear. But businesses that depend heavily on pay-per-click advertising should be monitoring how agentic commerce develops and whether their customer acquisition model remains robust as agent-mediated purchasing grows.

What Happens to Jobs?

There is no single answer to this, and any source that claims certainty should be treated sceptically. Several different outcomes may occur simultaneously, in different proportions in different industries.

• Augmentation — the same workforce produces substantially more, with AI handling the routine and people focusing on the high-value
• Role redesign — people stop doing repetitive tasks and take on different responsibilities, often requiring new skills
• Headcount reduction — some processes genuinely require fewer people as tasks are automated and not replaced with new work
• Demand expansion — cheaper, faster services create more demand, and the employment displaced in one place is absorbed elsewhere in the economy
• New businesses — smaller teams create products and services that previously required larger organisations, generating new employment

Historical technology transitions offer some guidance. The introduction of word processors did not eliminate secretarial roles in one step — it changed what those roles involved and eventually reduced them over decades as organisations found they could operate differently. The internet created enormous new categories of employment while disrupting established ones. The pace and nature of the current transition may be different, but the pattern of gradual task-level automation rather than instant job-level elimination is consistent with how previous general-purpose technologies have worked.

Agents and Small Businesses

For small businesses, the agentic opportunity may be significant. Historically, a small business competing in markets occupied by larger organisations has been limited by the cost of building capacity — hiring people for marketing, administration, customer support, sales preparation, research and operations.

A two-person consultancy with well-configured AI agents might produce the research, documentation, client communications and administrative output that previously required a substantially larger team. The competitive advantage goes to those who identify the right work to delegate, configure it reliably and maintain the quality that clients experience.

But agents still require systems, integration, data, permissions, governance and human judgement. The overhead of managing AI tools properly is not zero. Small businesses that deploy agents without appropriate configuration, oversight or quality checking may find they have created new problems — incorrect client communications, errors in financial records, missed escalations — rather than saved time.

Agents and Large Businesses

Large organisations have a different opportunity. They already have staff, processes, systems and data. Agents could reduce the friction between those elements — automating the hand-offs, data transfers and coordination steps that currently require human intermediaries.

Likely early areas for agentic automation in larger organisations include finance operations, HR administration, procurement processes, IT service management, customer service and internal reporting. Each of these involves significant volumes of structured processing tasks — the category AI handles most reliably.

The governance challenges are proportionally larger. Legacy systems may have poor API coverage. Organisational boundaries can make agent deployment politically complex. Permissions in large systems can be difficult to scope correctly. And when an agentic system fails at scale, the clean-up can be substantial. The businesses that deploy agents most successfully will be those that invest in the governance infrastructure before they invest in the automation.

Risks of Agentic Work

The move from AI answering questions to AI taking actions changes the risk profile substantially. A chatbot that produces a wrong answer requires a human to do the wrong thing with it. An agent that takes a wrong action may have already done it.

• Actions may be irreversible — deleted records, sent communications, committed transactions cannot always be undone
• Errors can compound — a mistake in an early step propagates through subsequent automated steps
• Security risk increases — an agent with broad system access is a more valuable target for attack than a chatbot
• Prompt injection attacks — where malicious content in documents or emails manipulates agent behaviour — are a specific agentic risk
• Permission drift — agents accumulate access over time that was never explicitly approved
• Over-automation — the agent handles an exception it should have escalated
• Runaway execution — a loop or retry mechanism that cannot be stopped quickly
• Business dependency — a pilot system that quietly becomes business-critical before controls are in place

The Human Approval Problem

For AI agents to create productivity, they need enough autonomy to complete tasks without requiring human approval at every step. But meaningful actions — communications to customers, changes to financial records, purchasing decisions, deletions — require human accountability. The design challenge is identifying where approval gates should sit.

The answer is not 'approve everything' (which eliminates the productivity gain) or 'approve nothing' (which removes accountability and creates serious risk). It is identifying the specific actions that are consequential, irreversible or high-risk, and building approval requirements at exactly those points while allowing the agent autonomy everywhere else.

Approval gates should be: explicit (not implied), logged (for audit), tested (to confirm they actually work), owned (a named person is responsible) and maintained (as the agent's scope evolves). The existing IT Club article on AI agent safety covers the full governance checklist. The approach connects to the broader principle: governance should be proportionate to consequence.

Agent Governance

When agents move from answering to acting, governance becomes correspondingly more important. The controls that matter most include:

• Identity — each agent should have its own identity, not share credentials with a person or another agent
• Permissions — scoped to the minimum required for the task; regularly reviewed
• Data access — limited to what is genuinely necessary; customer data appropriately separated
• Tool access — each tool the agent can use should be deliberately authorised, not default-on
• External actions — communications, transactions and external API calls should require explicit approval
• Audit logs — all agent actions should be logged in a form that cannot be modified by the agent
• Spending limits — agents that can trigger financial transactions need hard budget limits
• Rate limits — to prevent runaway execution
• Stop controls — a tested way to pause or shut down the agent quickly
• Workflow ownership — a named person who is accountable for what the agent does
• Monitoring — unusual behaviour should trigger alerts, not just be logged
• Incident response — a documented process for when something goes wrong

These controls should be verified, not merely documented. Infrastructure-level limits — what the agent cannot reach, not just what it has been told not to do — are the most reliable.

Operational Heartbeat

Agentic workflows change as models improve, staff leave, permissions expand over time, tools change, costs shift and business dependence on the agent grows. A system that was well-governed at deployment may drift without a regular review that covers: what the agent is doing, who owns it, what tools and permissions it has, what errors have occurred, how many human corrections were required, what it cost, what productivity it actually delivered and whether that remains positive. The next review date should be agreed at deployment and treated as a commitment.

What Businesses Should Do Now

A practical eight-step approach:

1. 1Break jobs into tasks — do not begin with 'Which jobs can we remove?' Begin with 'What work is repeatedly being done?' Map the actual tasks within roles rather than thinking at the job-title level.
2. 2Identify repetitive work — look for research, copying, checking, updating, scheduling, summarising and drafting. These are the most immediate candidates.
3. 3Classify each task — use the AI Productivity Ladder: is this an Ask, Assist, Delegate or Orchestrate task? Or not yet ready for automation?
4. 4Pick a low-risk workflow — avoid immediately automating payments, legal submissions, safety-critical work or anything that cannot be reversed. Start with internal, low-consequence, well-defined tasks.
5. 5Measure the current process — understand time, cost and error rate before you automate. Without this, you cannot know whether the AI version is actually better.
6. 6Pilot the agent — keep scope narrow. One task, one workflow, one team. Not a full department transformation.
7. 7Measure total productivity — include checking time, correction time, failure handling and any new overhead the agent creates. Time saved on the task itself is not the same as net productivity gain.
8. 8Expand carefully — increase scope, tools and autonomy only when evidence supports it. A pilot that works well in a controlled test may behave differently at scale.

Warning Signs

Be sceptical of AI agent deployments where:

• Management begins with headcount reduction rather than workflow mapping
• Nobody has documented the actual workflow being automated
• The ROI calculation includes only theoretical time saved and excludes checking, failures and overhead
• The AI output requires complete human recreation rather than verification
• Agents have been given administrator permissions to make configuration simpler
• External actions — emails, transactions, API calls — occur without human approval
• Nobody is named as the owner of the workflow
• Staff are measured by AI adoption rates rather than business outcomes
• Junior role development has not been considered
• Tool licensing costs have been excluded from the productivity calculation
• Every automation failure becomes a human clean-up task with no remediation plan
• A pilot system has quietly become business-critical before controls were established
• There is no tested way to stop the agent quickly
• The process being automated was already broken before AI was applied

Practical Business Implications

• AI search is only the first phase — answers improve information access, but the larger economic opportunity lies in AI performing work
• Agents change the unit of work — AI can potentially take responsibility for multi-step tasks, not just responses to individual prompts
• Tasks will move before whole jobs — work is easier to decompose and automate than entire occupations
• The business model may shift — value may move from attention towards transactions and completed outcomes
• Software may become less visible — agents interact with systems; the user interface becomes less central
• Small teams may become more capable — AI provides access to digital labour that previously required headcount
• Junior roles may need redesign — career development cannot rely exclusively on the routine work AI now performs
• Human oversight still matters — more autonomy creates more governance risk, not less
• Productivity must be measured — AI adoption is not an economic outcome

The IT Club View

The current AI conversation focuses too heavily on a single question — 'Will AI replace jobs?' — and in doing so skips the more immediate transition that is already underway.

That is the transition most relevant to most businesses right now. Individual tasks — research, drafting, data processing, scheduling, routine communication — are already moving to AI assistance and will increasingly move to AI delegation. For many businesses, this will change what people spend their time on before it changes how many people they employ.

But there is an uncomfortable consequence that should not be avoided. If enough tasks inside a role disappear, some organisations will eventually find they can operate with fewer people in that role. This is neither inevitable nor instant, but it is a real possibility for some roles in some industries. Pretending AI will never reduce headcount is no more credible than claiming every office job is about to disappear.

Equally, cheaper and more capable businesses may expand, create new services, hire elsewhere and deliver more to the same clients. The eventual outcome is genuinely uncertain and will vary significantly by sector, organisation and the quality of the AI governance in place.

The sensible business strategy is: identify work, automate selectively, protect judgement, measure results, redesign roles, train people and govern agents.

Plain-English Takeaway

AI Task Delegation Checklist

Before delegating a task to an AI agent, use IT Club's AI Task Delegation Checklist to assess whether the task is ready, what controls are needed and how to classify it on the AI Productivity Ladder.

Download the AI Task Delegation Checklist →

Administrator and Governance Technical Note

Related Business Questions