Stop Buying Custom AI Agents. Hire AI-Native Employees Who Can Build Them.

Companies are beginning to spend serious money on AI agents.

A sales agent. A customer-support agent. A finance agent. An HR agent. An operations agent.

This is creating a new procurement question:

How much does it cost to build an AI agent?

But for many businesses, that is the wrong question.

A better question is:

Who should we hire so that our company can continuously create, operate and improve its own agents?

The argument is simple:

Businesses should stop treating custom AI agents as finished products to buy from vendors. Instead, they should hire AI-native employees who understand the business and can develop agents around real work.

The competitive advantage will not come from owning a few agents.

It will come from having people who know how to create them.

A Custom Agent Is Rarely a Finished Product

Imagine two companies buying a “customer-service agent.”

For the first company, the agent searches a knowledge base and drafts replies.

For the second, it must:

• identify the customer,
• read previous support tickets,
• access the CRM,
• check ERP inventory,
• determine warranty status,
• create replacement orders,
• issue refunds within certain limits,
• escalate exceptions,
• maintain audit logs,
• and explain why it made each decision.

Both companies can call the product a customer-service agent, but the engineering complexity is completely different.

The cost is determined much more by:

• data quality,
• integrations,
• permissions,
• business rules,
• exception handling,
• security,
• governance,
• human approval,
• and the consequences of mistakes.

McKinsey reported in 2026 that although nearly two-thirds of enterprises had experimented with agents, fewer than 10% had scaled agents to produce tangible value. It also found data limitations to be a major barrier to scaling agentic systems.^[1]

That tells us something important:

The difficult part is usually not creating an agent demo. The difficult part is making the agent part of the real business.

The Outsourcing Problem

Traditional software outsourcing works reasonably well when requirements are stable.

Define the requirements.

Build the software.

Test it.

Deploy it.

Maintain it.

Agentic AI is different because almost everything underneath the agent keeps changing.

Your business process changes.

Your employees change how they work.

Your CRM changes.

APIs change.

Company policies change.

Customers create new exceptions.

Foundation models improve.

Agent frameworks improve.

The tasks that required complicated workflows six months ago may suddenly be achievable by a model with a relatively simple instruction.

A custom agent therefore should not be treated as something that is simply delivered and finished.

It is closer to a continuously evolving digital employee.

And if your business depends on an external consulting company every time that digital employee needs to learn something new, you have created another form of vendor dependency.

Hire the Capability Instead

There is another option.

Instead of paying a vendor to build five custom agents, hire people capable of creating and improving agents internally.

These people are AI-native employees.

An AI-native employee is not necessarily an AI researcher or machine-learning engineer.

They might be:

• an AI-native accountant,
• an AI-native marketer,
• an AI-native operations manager,
• an AI-native salesperson,
• an AI-native product manager,
• an AI-native software engineer,
• or an AI-native customer-support specialist.

Their defining characteristic is not that they know everything about artificial intelligence.

It is that they naturally think:

Can I delegate this work to AI?

And then:

How do I turn what I know into instructions, tools and workflows that an agent can execute?

What Makes an Employee AI-Native?

An ordinary employee may use ChatGPT to help write an email.

An AI-native employee goes much further.

They can take a recurring business process and gradually transform it into:

Prompt → Skill → Workflow → Tool → Agent

They understand how to work with:

• prompts and context,
• reusable skills,
• agent instructions,
• APIs,
• MCP and external tools,
• automation,
• structured outputs,
• memory,
• evaluations,
• permissions,
• human approval,
• error handling,
• and multi-agent workflows.

But technical ability alone is not enough.

The most important advantage is that they also understand the business domain.

They know what a good result looks like.

They know which exceptions matter.

They know when the AI is wrong.

They know what can be automated and what should still require human judgment.

That combination is extremely valuable.

The Market Is Already Moving Toward AI-Native Hiring

This is not just a theoretical idea.

Microsoft's 2025 Work Trend Index found that 78% of leaders were considering hiring for AI-specific roles, rising to 95% among what Microsoft calls Frontier Firms. Roles being considered included AI trainers, AI agent specialists, data specialists, security specialists, ROI analysts and AI strategists across functions such as marketing, finance and customer support.^[2]

The World Economic Forum's Future of Jobs Report 2025^[3] surveyed more than 1,000 employers representing over 14 million workers. It found that AI and big data are the fastest-growing skills, while 63% of employers identified skills gaps as one of the biggest barriers to business transformation.

PwC's 2026 AI Jobs Barometer^[4] reports that skills in the most AI-exposed jobs are changing more than twice as fast as in the least exposed roles. It also finds that new tasks added to AI-exposed roles are 2.5 times more likely to rely on empathy, judgement and creativity.

The labour market is therefore starting to move from:

“We need somebody who can do this job.”

to:

“We need somebody who can do this job while using AI to multiply their capability.”

One AI-Native Employee Can Operate Multiple Agents

This is where the economics become interesting.

You should not think of:

1 employee = 1 agent

The relationship may increasingly look like:

1 AI-native employee = several specialised agents

OpenAI described this directly while discussing its internal use of Codex. Engineers commonly opened several agent sessions, assigned different jobs, reviewed outputs and redirected them when necessary. OpenAI found that people could comfortably manage roughly three to five concurrent Codex sessions before context-switching became problematic.^[5]

OpenAI's newer multi-agent tooling is explicitly designed around the same principle: one coordinating system can delegate independent work to several specialist agents operating in parallel.^[5]

This gives us a glimpse of the emerging employee model.

An AI-native marketing employee might supervise:

• a market-research agent,
• a competitor-monitoring agent,
• a content agent,
• an SEO agent,
• and a lead-research agent.

An AI-native operations employee could supervise:

• an invoice-checking agent,
• a supplier-monitoring agent,
• a reporting agent,
• a scheduling agent,
• and an internal knowledge agent.

The human does not disappear.

The human moves upward.

Their job increasingly becomes:

define → delegate → evaluate → correct → improve

So How Many AI-Native Employees Should a Company Hire?

There is currently no credible universal industry benchmark saying:

“Every company needs exactly X AI-native employees per 100 employees.”

The technology is too new, and different industries have very different automation potential.

But there are useful signals.

Citi provides an interesting real-world example. In June 2025, Citi said that more than 2,000 employees had joined its AI Champions and AI Accelerator programmes^[6] to support adoption and create a feedback loop across the firm. Citi did not present this as a universal staffing ratio.

Google also recommends creating distributed groups of generative-AI champions rather than leaving AI adoption entirely to a central technology department.^[7]

Microsoft describes a similar employee-AI model in which employees use agents to research, analyse, draft and automate their own workflows while humans remain accountable for decisions and outcomes.^[2]

Based on those signals, a reasonable pilot operating model is:

One AI-native builder/operator for every 20–50 knowledge workers.

This is a provisional starting range for testing, not a published industry standard.

It corresponds to roughly 2–5% of knowledge-work headcount.

For example:

• Knowledge workers: 10–25 · Initial AI-native employees: 1
• Knowledge workers: 25–50 · Initial AI-native employees: 1–2
• Knowledge workers: 50–100 · Initial AI-native employees: 2–5
• Knowledge workers: 100–200 · Initial AI-native employees: 3–10
• Knowledge workers: 200–500 · Initial AI-native employees: 5–20
• Knowledge workers: 500–1,000 · Initial AI-native employees: 10–50

The exact number matters less than where they are located.

Do not put all of them into an “AI department.”

Embed them inside the business.

Have an AI-native person in:

sales, marketing, operations, finance, customer service, product and engineering.

Because the person closest to the work usually understands best what should become an agent.

This Is Different From Hiring an AI Engineer

An AI engineer might know how to build sophisticated infrastructure.

That remains important.

But companies also need people who sit between traditional business operators and AI engineers.

Think of three layers.

Layer 1 — AI Platform Engineers

They manage:

• infrastructure,
• security,
• authentication,
• permissions,
• model access,
• data architecture,
• observability,
• compliance,
• and shared agent platforms.

Layer 2 — AI-Native Employees

They understand a business function and can convert real work into AI workflows and agents.

They become agent builders and agent managers inside the business.

Layer 3 — AI-Enabled Employees

Everyone else increasingly uses the agents, workflows and capabilities created by Layers 1 and 2.

This is much more scalable than asking a central AI department to interview thousands of employees and manually automate every workflow.

Why Hiring Wins Over Custom-Agent Procurement

Suppose a company spends $100,000 commissioning several custom agents.

After delivery, those agents begin ageing.

Every major change creates another integration request.

Every new workflow generates another statement of work.

Every unusual case goes back to the vendor.

Now consider using that budget to hire an AI-native operations employee.

During the first month, that employee may create one useful workflow.

Then another.

Then another.

They learn the company's systems.

They understand customer behaviour.

They accumulate prompts, skills, integrations and evaluation datasets.

Their agents improve because the employee improves.

The company gradually builds an internal AI capability rather than purchasing isolated pieces of automation.

The difference is:

Custom agents depreciate.

AI-native organisational capability compounds.

The Evidence Suggests the Biggest Gains Come From Changing How People Work

Microsoft's 2026 Work Trend Index found that 66% of AI users said AI allowed them to spend more time on higher-value work, while 58% said they were producing work they could not have produced a year earlier.^[8]

Among Microsoft's most advanced group of users—what it calls Frontier Professionals—that second figure rose to 80%.

The important transition is therefore not simply:

Human → AI Agent

It is:

Traditional Employee → AI-Native Employee → Human + Agent Team

That is a much more powerful organisational model.

Agents Are Becoming Cheap. People Who Know What to Do With Them Are Scarce.

Foundation models will continue improving.

Agent frameworks will become easier to use.

APIs will become easier to connect.

Model costs are likely to continue falling.

Agent creation itself will therefore become increasingly commoditised.

The scarce resource moves somewhere else:

business judgment.

Knowing what should be automated.

Knowing how the company really works.

Knowing which results are acceptable.

Knowing when an agent needs human approval.

Knowing how to redesign a process around capabilities that did not exist six months ago.

And knowing how to continuously turn that understanding into AI-executable workflows.

That is why the most important AI investment may not be another agent contract.

It may be the next employee you hire.

Don't Ask Candidates Whether They “Know AI”

There is also an important implication for recruiting.

Putting “ChatGPT experience” on a CV means very little.

Instead, interview candidates by asking them to demonstrate how they work.

Give them a real business problem.

Then observe whether they can:

1. break the problem into tasks,
2. determine what AI should handle,
3. delegate work to multiple agents,
4. provide the necessary context,
5. connect tools or data where necessary,
6. evaluate the output,
7. identify failure cases,
8. improve the workflow,
9. document it as a reusable skill,
10. and eventually automate the process.

You are not primarily testing whether someone knows how to write prompts.

You are testing whether they understand AI leverage.

That may become one of the most important hiring signals of the next decade.

The Better Enterprise AI Strategy

The first generation of enterprise AI strategy looks like this:

Find use case → Buy agent → Integrate agent → Deploy agent

The AI-native model looks different:

Hire AI-native talent → Give them secure AI infrastructure → Let them solve real problems → Capture reusable skills → Build internal agents → Scale successful workflows

Professional AI engineering teams still matter.

They should build the secure foundation:

• identity,
• permissions,
• data access,
• governance,
• auditability,
• model routing,
• infrastructure,
• evaluations,
• and agent security.

But the thousands of business workflows sitting on top of that infrastructure should increasingly emerge from the people actually doing the work.

Don't Buy Five Agents. Hire Someone Who Can Create Fifty.

This is ultimately the economic argument.

A custom agent solves a defined problem.

An AI-native employee can keep discovering problems.

A custom agent implements today's workflow.

An AI-native employee can redesign tomorrow's workflow.

A custom agent is an asset.

An AI-native employee becomes an agent factory.

As AI capabilities continue improving, this distinction becomes increasingly important.

Companies should therefore reconsider the question:

“Which agents should we buy?”

And start asking:

“Which AI-native people should we hire?”

Because in the long run, the strongest company may not be the company with the largest collection of purchased AI agents.

It may be the company where a relatively small group of exceptional AI-native employees can continuously create, supervise and improve hundreds—or eventually thousands—of agents around the organisation.

Don't just buy the agent.

Hire the person who can build the next hundred.