First, a word of vocabulary. A retainer is the advertising industry's long-term service model: a brand pays a fixed monthly or yearly fee and hands ongoing work — running its social media accounts, producing content — to an agency, as opposed to one-off, per-project campaigns. A typical retainer engagement has the agency permanently responsible for the brand's Instagram, TikTok, and X accounts: content planning, copy and design, publishing, community interaction, and performance reviews.

Now picture such a retainer team. All of them "use AI" — yet their mornings can look like four entirely different things.

Level 1 morning: everyone clicks away by hand

The company has no system. AI is a chat window each person opens on their own. Copywriters draft with AI, designers generate images with AI, the PM summarizes meetings with AI — but every task is still remembered, created, and pushed forward by a human.

Level 2 morning: the PM decides what to produce, then works the system

The company now runs a workflow system: brand knowledge base, content calendar, generation tools — all in place. And yet:

Level 3 morning: the agent has already decided and produced — the PM reviews

What the PM sees in the morning is not a piece of software waiting to be operated, but a report on work already advanced:

The human no longer needs to know in advance how much to produce today. Behind this stands a task loop the agent runs on its own (all 11 steps in 1.2 below).

Level 4 morning: the agent doesn't just run the accounts — it lands a hit

The level 4 morning goes one step further: the PM picks up the phone and finds that a piece the agent made has gone viral.

The viral hit itself is not the point. The point is that this piece was on no agreed schedule: the system found the opportunity, validated it, and then distilled the success into a reusable pattern — turning virality from a lucky accident into a manageable capability (what it takes, in 1.6 below).

The dividing line of this whole essay

The difference between these four mornings is the dividing line this essay is about:

One thing needs saying up front: an AI-native enterprise is not a rating scale for how much of the old organization you have automated — it is a redesign of the relationship between people, agents, information, and decisions. You cannot measure a company's AI-native level by how many models it has integrated, how many agents it has shipped, or how many employees use AI tools. What actually determines the level are six things — context reach, result closure, attention ownership, continuous learning, permissions to act, and whether processes and value get redefined (each unpacked in Part III · the six dimensions).

On that basis, this essay divides enterprise AI-native maturity into four levels, with software products mapping level for level:

The essay starts with the concrete retainer case, generalizes to the value levels of software products, and ends with the full enterprise-side framework.

Part I · A close look at the Retainer agent

1.1 What makes retainer work hard

A typical retainer engagement staffs one PM, half a copywriter, and one to two designers, plus shared hours from strategists and managers.

The pain is not just slow asset production. It includes:

In other words: the cost of a retainer is not just writing copy and making images — it is the cognitive burden of a PM continuously remembering, scheduling, coordinating, chasing, and closing open items.

1.2 The four product levels of a Retainer agent

Level 1: content generation tools

PMs, designers, and copywriters use scattered AI tools: writing headlines, generating copy, generating images, summarizing trends, drafting report outlines. Humans remain responsible for: remembering tasks, supplying context, organizing the process, editing outputs, managing versions, owning the outcome. What gets delivered: individual content assets.

Level 2: Retainer copilot / workflow SaaS

The PM enters the brand, the month, and the product focus, picks topics; the agent generates a few directions of copy and design; the PM selects, splices, edits, and reviews. The system covers trend-watching, topic selection, copy, artwork, publishing, and retrospectives — product capabilities include a brand knowledge base, monthly roadmaps, topic and content calendars, copy generation, visual briefs, review and version management, data rollups, report generation — but the PM is still the engine of the whole project.

The typical way of working is exactly the seven steps of the level 2 morning above.

Level 3: the Retainer digital employee

The real change at level 3 is not adding more generation modules. It is:

The agent proactively and continuously owns the operating responsibility — for example, a standing duty like:

Every day, or whenever an event fires, it checks content inventory, publishing schedule, client feedback, approval status, trends, and performance data; decides how many posts and image sets today needs; and completes production, QA, and process advancement on its own. One full round of work is a closed loop (the dashed line means: when done, return to step 1 and wait for the next wake-up):

This is the complete transfer of attention:

What level 3 delivers: a portfolio of social accounts that runs steadily with minimal human involvement. (What level 4 looks like: see 1.6.)

1.3 Is level 3 just level 2 plus a scheduler

Structurally, yes — level 3 can be built on top of level 2. It reduces to:

If the system merely "generates three posts at nine o'clock every day", it is still scheduled automation. A digital employee is:

So the timer only wakes the agent up. The real capability of a digital employee is:

1.4 Does a Retainer agent need to run nonstop

No. The right concept is not "continuous execution" but:

A retainer agent can wake on:

On waking, the agent can reach one of three conclusions: it must act now; it must wait for a condition; or there is nothing worth doing.

A mature digital employee must be able to judge that "now is not the time to keep working." It should never manufacture content and tasks just to look busy.

1.5 How a Retainer agent keeps learning

Start with concrete things worth learning, then sort them into layers. The system should remember, for instance: this client dislikes loud emoji; "premium" means more restraint, not more complexity; a certain product-exposure ratio always gets bounced; the client wants two complete directions per review — that is client-preference learning. Layered, a level 3 retainer digital employee needs at least four kinds:

Level 3 requires at least the first three; level 4 demands the fourth.

1.6 Level 4: the Retainer value-discovery engine

What the brand originally asked for may be nothing more than: keep up the social presence, publish on time, support campaigns, avoid PR incidents, send a monthly report.

The agent doesn't just maintain the brand's social presence — it keeps finding and realizing more value across the account portfolio. It proactively spots content directions, user sentiments, and brand narratives with breakout potential; grows one high-performing piece into a recurring series, a content IP, and a stable brand persona; turns virality from a lucky outcome into a capability that can be learned, reused, and managed; and ultimately produces verifiable lift in brand awareness, search, user preference, and sales.

"Make every post go viral" works as an internal ambition, not as a per-post contractual promise. The more accurate definition of level 4 is:

1.7 What justifies $20K–200K a year

Level 2 enterprise SaaS can also reach $20K a year, especially bundled with customization, deployment, data integration, and services. But to charge on the logic of a "digital employee" rather than "expensive SaaS," you must prove three things:

A suggested core dashboard for a digital employee:

The pricing logic in one sentence:

1.8 Team and resources at each stage

Part II · The value levels of software products

The four retainer levels are not specific to that industry. Generalize "generation tool → copilot → digital employee → value-discovery engine" and you get the value levels of software products at large:

The four levels also differ in their basic unit:

Level 1: point AI tools

Typical products: copy generators, image generators, video tools, summarizers, simple chatbots, one-shot automation plugins. The product delivers a result, but the user still holds all context, tasks, and responsibility. Priced against software tools, model calls, and convenience.

Level 2: ordinary SaaS / copilot

The internal workflow products of leading companies sit roughly here — in retainer terms, the copilot form of 1.2.

Typical pricing: $20–200 a month, more for complex enterprise editions. What this level sells: software capability and process efficiency.

Level 3: the digital employee

The basic definition of a digital employee:

A digital employee is not required to be busy all the time. It should be: continuously responsible, intermittently executing, waking on demand, sleeping when done.

Typical pricing: $2K–10K a month, benchmarked against role cost, capacity, and management attention (how to prove you deserve it: the three proofs in 1.7). What this level sells: sustained role capacity, closed-loop responsibility, attention transfer, changed staffing, delivery reliability.

Level 4: the value-discovery engine

A value-discovery engine carries the original responsibility and, beyond it, proactively finds: which goals are more worth pursuing, which work has no value, which opportunities the company has not seen, which experiments could create new revenue and profit.

Its pricing no longer benchmarks only software and labor. It can benchmark: incremental revenue, margin improvement, cost reduction, new business opportunities, avoided bad investments. Business models can include: base service fees, deployment fees, ongoing platform fees, revenue share on increments, profit-improvement share.

What this level sells: the sustained ability to discover and realize growth opportunities.

Part III · Enterprise AI-native maturity

Finally, the enterprise side: where does a whole company sit? First watch two departments across the four levels, then the full definitions of the four levels and the six dimensions that decide the answer.

3.1 Four levels through two departments

HR, level by level

Marketing ops, level by level

3.2 The four maturity levels defined

Level 1: the personal-tools company

Employees use general-purpose LLMs, AI coding tools, generation tools, and automation scripts on their own. AI capability is not yet an organizational asset: results depend on individual skill; prompts and experience leave with employees; data never flows back; workflows do not change; the financial contribution cannot be measured. At bottom it is still a traditional company in which some employees happen to use AI.

Level 2: the department-workflow company

Leading companies sit roughly here. The company can speed up each department's existing workflows with agents, producing cost, capacity, and financial improvements visible at group level — the level 2 of the HR and ops examples above. But the premise of this level is:

Level 3: the process-rebuilt AI-native company

Level 3 stops treating existing departments and processes as immovable premises. It redefines from scratch: how information flows, how work gets created, how departments collaborate, how humans and agents divide labor, which processes stay, which processes go — the level where, in the examples above, HR discovers "the growth ceiling is senior-PM scarcity" and ops proactively owns account responsibility.

Break the department walls; build enterprise-level context. The reason those level 3 examples are impossible at level 2 is that a traditional company's information is scattered across nine domains: sales, operations, product, R&D, production and delivery, support, finance, HR, management. Each department sees only local facts, so it can only optimize locally. The first thing level 3 builds is not more agents but a unified enterprise context, so the system can continuously understand:

• Why customers buy or churn
• Which projects earn revenue but lose money
• Which people and capabilities actually drive delivery
• Which problems keep recurring across departments
• Which internal goals conflict
• Which link constrains the whole company's growth
• Which business assumptions have expired

This is fundamentally a question of CEO resolve and CTO field of view. The CTO and the AI product lead must: ① observe every department's real workflows; ② map the information and efficiency gaps; ③ redefine the data structures; ④ work with the CEO to break departmental isolation; ⑤ turn information assets into durable context agents can use.

Continuous responsibility and the autonomous task loop. A level 3 system no longer depends on people entering software to create tasks. One full round of its work (isomorphic to the retainer employee's 11-step loop):

So the core change of level 3 is:

Continuous, autonomous learning. A level 3 agent must keep learning from: human edits, content acceptance and rejection, project wins and failures, client satisfaction, business results, execution anomalies, and its own errors. What it learns must be written into: long-term memory, business rules, prompts, eval sets, skills, model routing, dynamic workflows, decision policies. Early on the model need not modify its own weights — but one thing must hold:

The talent ladder level 3 requires:

Organizational preconditions: strong CEO backing; a CTO with a cross-department field of view; accessible information assets; walls that can actually be broken; solid permissions and auditing; a real business as a continuous training ground; a management team willing to redesign roles and processes.

Level 4: the value-discovering AI-native company

Level 4 does not merely hand tasks to digital employees — it redefines, from first principles, why each department exists: the level where, in the examples above, HR becomes a "capability allocation system" and ops becomes a "market-cognition, value-validation, opportunity-occupation system".

The level 4 enterprise loop:

Preconditions: the level 3 talent ladder is in place; company-wide information assets are connected; department walls are down; the CEO is willing to re-examine the existing business; the CTO tracks the moving frontier of AI capability; agents are allowed to question existing tasks; agents are allowed to drive experiments and resource reallocation; performance shifts from "how many tasks completed" to "how much value created."

3.3 Six dimensions that decide your level

Strip away all the examples, and what actually determines where a company (or a product) sits are six things:

The progression of the six dimensions across the four levels, as one matrix:

Dimension 1: context reach

Level 1: AI has only what the user pastes in. Level 2: AI reads one department's knowledge, processes, and history. Level 3: AI connects the information assets of sales, operations, product, delivery, R&D, support, finance, and HR. Level 4: AI understands not just the inside of the company but continuously absorbs the market environment, competitors, user behavior, technology shifts, culture, supply chains, and industry change.

Dimension 2: result closure and pricing class

What decides a product's level is not how smart the model is, but:

Commercial value can be understood as:

Dimension 3: who holds the working attention

This is the decisive boundary between levels 2 and 3 — the abstract version of the second and third mornings that opened this essay.

Which means AI has taken over not just the labor, but the attention of management.

Dimension 4: the capacity to keep learning

Continuous learning is not just online training of model weights. An enterprise agent's learning has at least five layers (the retainer agent's four kinds in 1.5 are the concrete form of the first four):

An enterprise digital employee needs at least the first three; the fourth underpins the value-discovery engine. The fifth remains frontier research on base models — not a precondition for the product to work.

Dimension 5: permissions and org structure

What an agent can create is bounded by the information and permissions it holds. An HR agent that can only read résumés can only optimize screening. Give it access to performance data, project margins, client reviews, delivery quality, collaboration graphs, and future business plans — and it may redefine what the company actually lacks.

That is why moving from level 2 to level 3 is not merely a technical upgrade but a joint shift in data governance, permission governance, department boundaries, incentive design, and decision rights.

Dimension 6: observe, distill, define

Genuine AI-native product definition is not collecting feature requests from business units. It is three things:

Final summary

The four levels of the AI-native enterprise, the four levels of software products, and the retainer agent's levels compress into one table:

So the most accurate definition of a digital employee:

And a value-discovery engine:

Appendix A · How the market and frontier labs define the digital employee

A.1 There is no unified definition

The previous generation of automation vendors defined digital workers as software robots that execute enterprise processes end to end. Automation Anywhere emphasizes a "virtual workforce" of AI, machine learning, RPA, and analytics completing sequences of business tasks; UiPath now describes a single AI agent as an autonomous digital worker that completes a specific task from start to finish. (Automation Anywhere)

Salesforce uses the concept of "digital labor," stressing agents with defined roles, enterprise knowledge, the ability to act, and around-the-clock availability; Microsoft stresses managing agents like employees: define the role, limit the permissions, supervise continuously. (Salesforce)

So today's de facto market floor is roughly:

This essay uses a stricter definition:

The heart of it is not "a model running 24 hours a day," but:

A.2 What the frontier labs are building toward

Frontier agent research is no longer chasing single-shot reasoning alone; it is filling in the capabilities a long-lived working entity requires:

Which is to say:

Appendix B · From discovery to operations: how the product gets built

B.1 Finding the need

Need-finding requires the CEO, the CMO, or someone with deep market experience. The goal is not collecting feature suggestions but understanding: who hurts most; whether the problem is frequent; how it is solved today; why existing solutions fail; who owns the budget; who bears the risk of failure; how much the problem costs in profit; and why AI can solve it now.

The final deliverable is not a feature list but: a set of testable value hypotheses.

B.2 Defining the product

Product definition takes the CTO, the AI product lead, AI-native engineers, and business experts together. The core method: observe → distill → define (the same method as dimension 6, here as the working version for product definition).

This stage requires every CXO to converge on which level the product is.

B.3 Design and engineering

A senior product-engineering team covering four blocks:

B.4 Running it in production

An AI-native product needs continuous operations after launch — and operations here is not sales plus support. It is running:

The core work: measure the unedited-into-review rate, the unedited-direct-publish rate, the rejection/rework rate; analyze failure traces; keep an error ledger; fold repeat errors into the regression set; update brand memory, rules, and skills; study why humans take over; prove value with real project financials.

Early go-to-market should lean on hard numbers from lighthouse projects, not agent demos: how many labor hours dropped, how many management minutes per brand dropped, how many more brands one team can serve, how much gross margin improved, whether delivery quality held.