Will AI Disrupt My Job? The Business Case for Humans and AI

Yes, AI will disrupt many jobs. It will change the tasks people perform, the skills companies value, and the economics of delivering products and services. Some roles will shrink. Others will expand or take a different shape.

That disruption is easier to understand, and to act on, when you look past the individual job description to the work a business needs done. Start with the two extremes: all the work done by AI, or all of it done by people. Neither holds up on its own, but each brings something the other lacks.

What the business needs

A business wants to serve more customers, build useful products, offer more services, and respond when customers need help. To grow that way sustainably, its work must be fast, accurate and reliable, someone must be accountable for each result, and the cost of delivery must still allow a competitive price. The practical question is how to organize work so the company delivers those outcomes, and how employees contribute as that work changes.

How to divide the work

The strongest operating model assigns work at the task and decision level. An entire department rarely belongs in a single category. Three modes cover most of it.

Design the work around three modes

Mode	Accounting example	Development example	What makes it suitable
AI executes within defined boundaries	Process eligible routine invoices after required checks and approvals	Run a verified routine maintenance workflow	Clear rules, dependable inputs, limited consequences and observable results
Humans lead and decide	Resolve a disputed obligation or authorize a policy exception	Resolve conflicting requirements or accept a release tradeoff	Ambiguity, material consequences, relationships or organizational authority
Together AI and humans	AI investigates a reconciliation difference; an accountant evaluates the explanation	AI proposes a fix and supporting tests; an engineer validates the behaviour	Automation can reduce effort while judgment remains necessary

“AI executes” should mean that an eligible task can proceed without a person reviewing every instance. It still needs monitoring, escalation rules, and a named owner.

“Humans lead” means the consequential decision remains with a person. Supporting research and preparation can still be automated.

In a hybrid workflow, the handoff must be explicit. The reviewer should receive the evidence, unresolved questions, proposed action, and reason for escalation. Putting a person at the end of every AI workflow does not automatically make it effective: if that person must reconstruct the work from scratch, the company has simply moved the effort into review.

Accounting: follow an invoice

Consider an accounting team processing invoices, reconciling accounts, closing the books, and answering questions from the business. Follow one supplier invoice through it: an invoice that exceeds its purchase order.

The outcome the business wants is easy to state. The right supplier is paid the right amount, on time, the transaction is recorded correctly, and there is a clear reason for any exception.

Where time goes

Some work involves entering data. Much of the difficulty lies in finding information, resolving inconsistencies, and getting decisions from other people.

An invoice arrives without a purchase order. A receipt has not been recorded. The amount differs from the agreement. Someone must determine which entity incurred the expense. A payment covers several invoices. An expenditure might belong in fixed assets, prepaids, or operating expenses. Meanwhile, an approver is unavailable, a supplier is asking about payment, and the controller needs the month-end numbers.

The work moves between email, spreadsheets, the ERP, expense systems, procurement, and conversations. Each handoff introduces an opportunity for delay or misunderstanding.

So before introducing AI, measure four kinds of time: doing the work, waiting, resolving exceptions, and correcting earlier mistakes. They are different problems. Faster invoice extraction helps with data entry. It will not, by itself, resolve an unclear approval policy or persuade someone to confirm whether goods were received.

The longest delay may sit outside the task being automated. If an invoice takes minutes to enter but several days to approve, cutting entry time produces a modest improvement unless the approval delay is addressed too.

Where an invoice’s elapsed time goes

TodayMinutes to enter, days to approve

EnterWaiting for approval

Faster entry onlyAI extraction, same approvals

Waiting for approval

Entry and approval both addressedClear routing, limits and delegation

Approval

That is why an accounting AI initiative should begin with the full journey from transaction to trusted financial information.

What AI can handle Extreme 1 · only AI

Imagine a system that reads invoices, identifies suppliers, proposes accounting codes, checks for duplicates, matches supporting documents, routes approvals, and posts eligible transactions, with no person involved.

For well-defined transactions, that removes substantial manual effort. A properly designed system can process a growing volume without someone inspecting every routine item.

Then the invoice that exceeds its purchase order arrives. The system detects the mismatch quickly. What happens next depends on why it exists. Perhaps the supplier made an error. Perhaps the company approved extra work during an urgent repair. Perhaps the purchase order is outdated. Perhaps the additional amount is disputed. Each needs a different response.

AI can retrieve the available correspondence, summarize the evidence, and propose alternatives. With the right integrations and permissions, it can contact the appropriate person or route the exception. But it may be missing a conversation, a commercial commitment, or a relationship that changes the decision.

The issue is whether the system has sufficient evidence, dependable reasoning, and authority for this particular action. A confident answer does not establish any of those things.

Where people must contribute Extreme 2 · only people

Now give the same invoice to a team doing everything by hand. An experienced accountant recognizes that the operations manager knows what happened. A controller authorizes an exception within policy, requires additional documentation, and preserves the reason for the decision.

That is what people contribute: they interpret ambiguity, challenge assumptions, negotiate, and accept responsibility. They recognize when a technically correct response would create an unnecessary business problem.

But in this extreme the same people also key, code and match every routine invoice, and human capacity is finite. Repetitive work consumes attention. People become fatigued, switch between competing priorities, and have different levels of experience. Asking a team to sustain the same concentration across every transaction, every day, is an unrealistic operating model.

As the business grows, it can add staff, and with them more handoffs, supervision, training, and coordination. The result may be an accounting function that spends most of its capacity processing the past while the business needs help planning the future.

How the work moves between them The combined approach

The better approach keeps AI’s capacity for routine volume and people’s judgment and authority, and designs out what breaks on each side: AI does not act on an exception it lacks the evidence or authority for, and people stop spending their attention on routine items.

Accounting: the two extremes, and the hybrid built from them

Extreme 1

Only AI, no people

What works

• Reads invoices, identifies suppliers, proposes coding, checks for duplicates and matches documents
• Handles a growing volume without someone inspecting every routine item

What breaks

• Can miss a conversation, a commercial commitment or a relationship that changes the decision
• A confident answer is not sufficient evidence, dependable reasoning or authority to act

Extreme 2

Only people, no AI

What works

• Interpret ambiguity, challenge assumptions and negotiate
• Accept responsibility, and see when a technically correct answer creates a business problem

What breaks

• Finite capacity: fatigue, competing priorities, uneven experience
• Growth adds handoffs, supervision and coordination; most capacity goes to processing the past

Keep what works on each side. Design out what breaks.

The hybrid

AI carries the routine volume. People own the exceptions and the decisions.

Following the supplier invoice that exceeds its purchase order:

1. AIPeople

2. Step 1

   Reads the invoice, proposes coding, checks for duplicates, matches the documents

   AI

3. Step 2

   Routes approvals and posts eligible routine invoices after the required checks

   AI

4. Step 3

   Flags the mismatch, gathers the correspondence and summarizes the evidence

   AI

5. Step 4

   Evidence, open questions, proposed action and the reason for escalation go to a person

   Handoff

6. Step 5

   The controller decides within policy, asks for documentation if needed and records why

   People

7. Step 6

   The time saved goes to exceptions, cash flow, financial interpretation and decisions

   People

In this workflow, AI executes the routine processing. Escalation happens when the documents disagree or the evidence is incomplete. An authorized person decides the exception, and a named owner in the accounting team is accountable for the outcome, including the record of why.

Early research supports this direction. A study described by MIT Sloan found that accountants using AI-enabled software supported more clients and shifted time from data entry toward communication and quality assurance. The researchers also emphasized the importance of experienced judgment and oversight. These findings describe a particular setting, rather than a guaranteed result for every accounting team.

Development: follow a change request

Now consider a development team, and a customer request to “add an approval workflow.”

The outcome the customer wants is not code. It is an approval process that behaves correctly in their business, works with their existing systems and permissions, and can be trusted in an audit.

Where time goes

The visible output is software, so it is tempting to assume that writing code is the main constraint. Yet delivery also depends on understanding requirements, navigating an existing system, obtaining access, resolving dependencies, testing integrations, and deciding what the customer actually needs.

The request sounds straightforward. Then the questions emerge. Who can approve? What happens when that person is unavailable? Are approval limits different by entity? Can the requester also approve? What happens after rejection? Must the decision be retained for an audit? How does this interact with the customer’s existing permissions?

A developer who asks the right question early can prevent days of rework. A solution lead may discover that the existing platform already supports most of the requirement. A support engineer may recognize that a configuration change solves the immediate problem more safely than new code. Those contributions are part of development productivity, even when they produce fewer lines of code.

What AI can handle Extreme 1 · only AI

Imagine an AI system receiving the requirement, writing the code, generating tests, and preparing a deployment.

Within a bounded environment, with clear requirements and strong verification, substantial portions of that workflow can be automated.

The difficulty increases when requirements are incomplete or the system has consequences beyond the immediate task. The workflow can pass its tests while implementing the wrong interpretation of who may approve. An integration can work with sample data and fail on an unusual customer transaction. A fix can solve the reported issue while introducing a permissions problem elsewhere. Tests generated alongside the code help, but they may share the same mistaken assumption as the implementation.

Someone must establish what correct behavior means, and evaluate the result against evidence that does more than repeat the original assumption: business acceptance criteria, integration behavior, security requirements, and the consequences of failure. AI can assist with each of these. The organization still needs an accountable owner for what reaches the customer.

Where people must contribute Extreme 2 · only people

People supply exactly what the first extreme lacks: they agree the behavior with the customer, recognize when the platform already does the job, and decide which tradeoffs are acceptable.

But a team relying entirely on manual effort also spends substantial time searching documentation, producing standard components, investigating logs, drafting tests, and answering recurring support questions. That consumes capacity that could go toward architecture, customer discovery, difficult debugging, and product improvement.

AI can help accelerate this work, but the gain must be measured across delivery. Field experiments run with Microsoft found increased task completion with an AI coding assistant. METR’s early-2025 study found a slowdown among experienced developers working in familiar repositories; its February 2026 update said newer results were difficult to interpret reliably and should not be treated as a clean estimate of current performance. The tools, tasks, developers, and working environment matter.

How the work moves between them The combined approach

The better approach lets AI speed up the searching, drafting and checking, while people define what correct means before the work starts and verify it before anything ships.

Development: the two extremes, and the hybrid built from them

Extreme 1

Only AI, no people

What works

• Writes code, generates tests and prepares a deployment quickly
• Automates much of the workflow in a bounded environment with clear requirements and strong verification

What breaks

• Can pass its tests while implementing the wrong interpretation of the requirement
• Generated tests can share the same mistaken assumption as the code

Extreme 2

Only people, no AI

What works

• Ask the right question early and prevent days of rework
• Spot that the platform already supports the need, or that configuration is safer than new code

What breaks

• Time goes to searching documentation, standard components, logs and recurring questions
• Less is left for architecture, customer discovery, hard debugging and product improvement

Keep what works on each side. Design out what breaks.

The hybrid

AI speeds up the building. People decide what correct means and own what ships.

Following the request to “add an approval workflow”:

1. AIPeople

2. Step 1

   Agree the behaviour: who approves, delegation, limits by entity, rejection, audit, permissions

   People

3. Step 2

   Searches documentation and logs, drafts the implementation and supporting tests

   AI

4. Step 3

   The proposed change, its tests and the open questions go to an engineer

   Handoff

5. Step 4

   The engineer validates against acceptance criteria, integrations and security, not only the generated tests

   People

6. Step 5

   Runs the verified checks and prepares the deployment

   AI

7. Step 6

   A named owner accepts what reaches the customer

   People

Here, people decide the behavior and accept the release, AI executes the drafting and the verified checks, and the handoff carries the proposed change, its tests and the open questions to an engineer. For a business, the useful measure is how quickly the customer receives a dependable solution. More generated code is valuable only when it contributes to that outcome.

Turn capacity into growth

Suppose an accounting firm can support additional clients with its existing team, or an internal finance team can support more entities and business units. Suppose a development company can release improvements sooner and offer more responsive support.

Those are opportunities. They become business results when the company has demand, a credible offering, and the capacity to deliver consistently. So leaders should decide where the newly available time goes: additional customers, faster onboarding, a new service, better financial analysis, product development, or support closer to the moment a customer encounters a problem.

For just-in-time support, AI could gather the relevant information, diagnose familiar issues, and carry out approved routine remedies. A person takes over when the issue requires a consequential decision, an unfamiliar investigation, or coordination across teams.

Just-in-time support: one service, one clear handoff

A customer hits a problem

At the moment they need help

AI

Gathers the relevant information, diagnoses familiar issues and carries out approved routine remedies

People

Takes over when the issue needs a consequential decision, an unfamiliar investigation or coordination across teams

One continuous service

Clear ownership through the handoff

The customer should experience one continuous service, with clear ownership through that handoff.

Count the total cost of delivery

The cost of delivery includes software subscriptions, model usage, integration, supervision, review, rework, and support. A cheap AI tool that creates expensive corrections can increase the total cost. A more expensive solution that reduces delays and reliably handles routine work may improve the economics.

The appropriate price must support customer value, dependable service, and a sustainable margin.

What you should do next

None of this guarantees that every existing role will remain. Automation can reduce staffing needs, and growth may not absorb all the capacity released. Leaders should be honest about that possibility. But designing solely around headcount reduction leaves the larger opportunity unexplored.

The larger goal is a business that can serve more customers, deliver more useful products and services, respond sooner, and maintain trust at a sustainable price. AI contributes execution capacity. People contribute expertise, judgment, relationships, and accountability. Their combined value depends on how thoughtfully the company organizes the work.

Will AI disrupt your job? Very likely. The opportunity is to help shape how the work changes, and become better at delivering the outcomes the business and its customers need.

Sources

• Roger L. Martin, The Opposable Mind: How Successful Leaders Win Through Integrative Thinking (Harvard Business School Press, 2007)
• MIT Sloan: How generative AI can make accountants more productive (August 2025)
• Microsoft Research: The Effects of Generative AI on High-Skilled Work, three field experiments with software developers
• METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (July 2025)
• METR: We are changing our developer productivity experiment design (February 2026)