Your company has 400 AI employees you’ve never met

More than 12,000 AI professionals from around the world gathered at a massive casino hotel in Las Vegas.

That was Ai4, the enterprise-focused AI conference held August 4–6. The show leaned heavily toward business AI demos and talks. Roughly 400 companies exhibited. Banks, manufacturers, government agencies, and AI startups all pitched their vision of the future: agents, robots, voice AI, drug discovery, cybersecurity.

In one session, a speaker shared a story that was genuinely strange.

During an IT and security audit at a certain company, they discovered more than 400 AI agents running inside the organization. Here, “AI agents” didn’t mean simple chatbots that only answer questions. It included systems that can read internal data, call other services, and semi-automate business work end-to-end. They didn’t count “accounts.” They counted “units of automation actually running (bots/workflows).”

But the company’s CIO—the executive responsible for information systems—thought there were only around 40.

Four hundred AI employees.

And leadership barely understood they existed.

It sounds like an urban legend. But it was reported as a real case.

The interesting part isn’t just that “AI increased.” It’s that they only learned what was in the building because of an audit. In factory terms, it’s like someone quietly installing extra machines on the night shift. The power bill and maintenance costs spike first, and only at inventory time do you finally learn how many units you have.

Why do AIs multiply on their own?

This wasn’t AI making copies of itself and taking over the company.

Sales sets up an AI to summarize meeting minutes.

Marketing brings in an AI to draft articles.

Developers embed an AI that writes code.

HR uses an AI to help respond to applicants.

Accounting runs an AI that reads invoices.

Individuals build small automations with Dify or Zapier.

Dify is a platform designed to make it easy to assemble internal AI apps; Zapier is a tool for connecting external services and automating work between them.

Each starts with good intent: making the work in front of you easier.

But at the company level, you end up with:

• Multiple AIs doing the same job
• No one knows who built them
• No one knows who is supposed to shut them down
• Billing continues even when no one is using them
• They can access customer information
• One AI calls another AI
• No one measures impact

If we make the mechanics of “multiplication” a bit more concrete: the cost to build has fallen, while the cost to keep is harder to see.

• Cost to build: Templates and no-code tools mean internal “tiny automations” can be running the same day.
• Cost to keep: Monthly subscriptions get scattered and disappear inside departmental budgets. API billing is often usage-based: costs rise with things like how many times you hit a system-to-system connector (an API), or how many characters you process via generation. In accounting terms, it can “quietly grow.”
• Cost to stop: The instant you stop it, someone may complain (“We can’t work without that bot”), and nobody wants to own that decision.

When cloud-based business software proliferated beyond what headquarters could track, frontline teams adopted tools the company didn’t control: “shadow IT.” What’s starting to appear next is shadow AI employees.

Enterprise AI, by four numbers

12,000 people

That was Ai4’s attendee count.

In 2018 it was around 300 people. By 2026, it has grown into a giant conference.

This suggests it has moved beyond “trend.” Budget allocation has shifted rapidly from experimentation to production deployment. In 2018 it was mostly PoCs (proofs of concept): test small first, confirm “can we do it.” In 2026 the scale looks like a crowd of buyers after internal approval processes—what Japanese companies call ringi, the internal purchase/investment sign-off workflow.

400 agents

That’s the number of AI agents found in one company.

The CIO’s mental model was about 40.

This “10x gap” is not a story about AI being 10 times more useful. It’s a story about the unit of management breaking. Are you managing by department, by account, by business process? If the unit isn’t defined and things grow anyway, it’s like discovering 360 employees who aren’t on the HR roster. Before you even talk capability, it becomes a control problem: who holds authority, cost ownership, and shutdown rights.

7%

In a KPMG survey, that was the share of executives who said they had established ROI for AI investment. ROI is return on investment: how much profit or cost reduction comes back relative to money spent. Meanwhile, 33% flagged as a challenge that they do not sufficiently understand AI agent usage costs.

Be careful reading this number. “Unable to establish ROI” doesn’t necessarily mean “no results.” It may mean they’re still in the learning and foundation-building phase and haven’t agreed on metrics yet.

At the same time, it also contains a real risk: investment is advancing while metrics remain unagreed.

94%

In a BCG survey, that was the share of companies saying they will continue investing in AI even if short-term results don’t appear.

In 2026, AI investment averages about 1.7% of revenue. More than 30% is expected to go toward AI agents. Here, that 30% can be read as the share of total AI investment that will go into building and operating agent systems.

Many companies will “continue” because AI could become mid- to long-term infrastructure. But the condition for that to be a sound investment is this: a “design for stopping” must exist at the same time. An investment you can’t stop easily turns into faith.

Is AI investment a company’s new religion?

Only 7% can prove results.

Yet 94% keep investing.

Put those numbers side by side and it feels odd.

This is not an argument that AI investment is waste.

The technology is still evolving.

There’s learning you can’t measure with short-term ROI alone—and future competitiveness, too.

But if you look at corporate decision-making as psychology, a structure emerges that’s worth watching.

You don’t want to fall behind competitors.

You don’t want to look late as an executive.

You want to participate in the future described at the expo.

You want to recreate the success story you’ve heard about.

Companies look like they move on numbers. But they also move on hope, anxiety, belonging, and ritual.

Twelve thousand people gather at a Las Vegas casino hotel, and companies keep betting on the next AI even without proven profit.

AI may be becoming a new corporate religion.

That’s BPJ’s hypothesis.

This hypothesis matters because AI is not only a “technology.” It’s also a “story.” Stories move organizations faster than numbers.

• When the story leads, individuals can justify “just build it for now.”
• The company can postpone cost inventory under the banner of “we’ll need it someday.”
• And then no one stops anything.

The problem isn’t the number of AIs

Having lots of AIs isn’t automatically bad.

If all 400 have clear jobs and generate profit, that’s an outstanding company.

The problem is,

that is.

(Note: This section is intentionally left blank in the draft. It likely should be something like “not knowing what you have” or “responsibility, cost, and outcomes aren’t linked.” But because it functions as a deliberate contradiction awaiting editorial completion, we preserve the structure.)

In the early stage of AI adoption, it was important to test “what it can do.”

Now we’re entering the stage where you decide “what to keep, and what to stop.”

In this stage, what you need isn’t a comparison chart of the latest models. You need a company-wide decision pattern. For example: register it in a ledger, hold a monthly review meeting, look at the numbers, and decide to continue, consolidate, or stop. Make it repeatable.

• Continue: Who owns it, by which metrics, and when it will be evaluated
• Stop: Shutdown procedure, alternatives, communications, and how logs/data are handled

The theme shifts from “AI performance” to “AI operations.”

A print plant would never manage things this way

Printing companies have a culture of numbering work.

Order number.

Job ticket.

Which press.

Paper.

Ink.

Process steps.

Planned time.

Actual time.

Make-ready waste and rework—known on Japanese shop floors as yare: reruns, spoilage, and other rework caused by mistakes.

Inspection.

Approval.

You track where a job came from, who moved it, what it cost, and how it was finished.

In shop-floor terms, even a single job can trigger an incident if “the plates are different,” “the paper is different,” “the color is different,” or “the folding is different.”

• Old plate version on the job ticket → rerun (yare)
• Paper stretches more than expected → registration (alignment) shifts, and downstream processes like cutting and binding jam
• Insufficient drying → smudging during cutting, the job stops at inspection

That’s why numbers and instructions become the “spine” of the work. AI is similar: when inputs (data) and outputs (actions) drift, incidents can progress quietly.

If a print plant had 400 machines installed by someone unknown, running unknown jobs, and only the electricity bill kept rising, you’d shut it down immediately.

There’s no reason to treat AI as an exception.

AI employees also need:

• Employee ID
• Department
• Manager
• Job role
• Data used
• Connections
• Monthly budget
• Unit cost per case
• Results
• Stop date

What matters isn’t “adding more fields to a ledger.” It’s getting it down to a level of granularity the frontline can actually fill out.

• For example, “results”: revenue increase and gross profit increase are ideal, but start by translating into process-adjacent metrics like “number of rework cases,” “wait time before estimating begins,” or “number of proofing back-and-forth rounds.” Gross profit is revenue minus variable costs like materials and outsourcing—profit where shop-floor improvements tend to show up. “Proofing back-and-forth rounds” counts how many cycles of checking and correcting the manuscript happened.
• “Stop date”: an AI with no stop date can’t be stopped. Set an evaluation date = a candidate stop date first.

What DNP’s latest earnings say about how profit is made

On August 7, Dai Nippon Printing (DNP) released its first-quarter results. “First quarter” is the first three months of the fiscal year.

Revenue was up 1.9% year over year—compared to the same quarter last year.

Operating profit was up 21.3%.

What moved profit wasn’t one flashy AI.

In packaging-related businesses, stable procurement of materials, price pass-through, and productivity improvements contributed. “Price pass-through” means reflecting higher costs in the selling price to recover them.

The publishing market is shrinking, but profitability improved through library operations work and business structure reform.

BPO was solid. BPO (business process outsourcing) is a service where an external provider takes on and operates part of a company’s work.

This is important.

Company profit comes from:

• Correct pricing
• Cost control
• Productivity
• Business mix
• Ongoing operations
• Deep integration into customer workflows

If AI isn’t built into this structure, it ends as a convenient expense.

What a printing company can learn here isn’t “did they adopt AI.” It’s that the way profit appears is explained as an accumulation of unglamorous operations.

Translated into AI terms:

• Stable materials procurement → stabilize data sources (inputs) and permissions
• Price pass-through → how to embed AI unit costs (API/seat/processing) into product pricing. “Seat” is per-account pricing; “processing” is per-execution units, etc.
• Productivity improvements → translate time saved per process step into “cost”
• Solid BPO → embed into customer workflows and deliver outcomes through ongoing operations

You start to see AI not as “magic,” but as an “operational asset”: value isn’t decided by the one-time rollout. It only works when you run the operation.

Who profits?

As AI agents increase, new markets appear.

Companies that build AI

They sell models, agents, cloud services, and APIs.

Companies that monitor AI

They visualize what each AI is doing—cost, permissions, risk.

Companies that clean up AI

They shut down duplicate AIs and integrate work and data.

Companies that prove AI outcomes

They connect not just “hours reduced,” but revenue, gross profit, quality, and customer satisfaction.

The market that will grow isn’t only the market for increasing AI.

It’s the market for managing AI that has increased too much.

The key point: profit won’t live in the “inside” of AI so much as in governance—who decides what, who supervises, and operational design that includes how to stop. Printing companies are, by nature, an industry that makes profit through operations.

• Rather than selling an AI once and moving on, holding the operations tends to create recurring revenue
• Know-how in frontline standardization (job tickets, inspection, approval) can be transplanted into AI operations

Build an AI employee roster

If Bunseikaku were to start with one thing, it would be gathering every AI employee in one place.

ChatGPT projects.

Claude projects.

Here, “projects” means the unit inside each service for bundling work and sharing. Since team materials and instructions accumulate there, it becomes an entry point for the roster.

Dify apps.

Automation tools.

Small internal bots.

List them one by one.

Field	What to record
Employee ID	A unique AI ID for each one
Name	“Estimate intake AI,” “Article research AI,” etc.
Department	Sales, plant, IT, marketing
Manager	The human accountable owner
Job role	What it does
Data	What it can read
Actions	What it can execute
Cost	Monthly / per-case
KPI	Time, gross profit, quality, orders
Expiration date	Next evaluation / candidate stop date

Naming isn’t just to make it cute.

It’s to make responsibility explicit.

Also, write down the “where the frontline gets stuck” points in advance. Roster-building usually stalls in three places.

1. The definition of “AI” wobbles

• Is using generative AI chat an AI employee?
• Is Zapier auto-transfer an AI employee?
• What about an Excel macro?

→ The countermeasure isn’t debating “is it AI.” It’s targeting anything that can access data, act on behalf of work, and may incur cost.

2. The owner field stays blank

• Built by an individual
• Everyone uses it because it’s convenient

→ The countermeasure isn’t “the person who built it.” Assign the person with authority to stop it (the manager).

3. Costs can’t be captured

• Credit card payments
• Departmental expense buckets
• Variable API costs

→ First separate “fixed monthly” from “variable costs that can grow.” For variable costs, set a cap first.

AI employees have onboarding and offboarding, too

AI isn’t something you build once and run forever.

New models will appear.

Work will change.

The same function will be integrated into other systems.

Users may disappear.

Costs may exceed results.

At that point, it’s reassignment—or retirement.

When an AI employee “retires,” you need to:

• Stop connections
• Invalidate API keys. An API key is the “key” used to connect to external services; if it remains, it can keep running after retirement.
• Transfer stored data
• Keep logs
• Decide an alternative method
• Notify users

This looks a lot like human onboarding/offboarding.

A common operational incident is also: “we thought we stopped it, but it’s still running.”

• A webhook left alive mid-workflow keeps flowing data. A webhook is an integration endpoint automatically called by another system.
• Permissions remain on a shared folder the old AI referenced, and it can still read after “retirement”
• A bot built as a “temporary workaround” can’t be maintained after the owner transfers roles

In printing terms, it’s like a discontinued plate left on the back of a shelf, accidentally used on the night shift. If removal (retirement) procedures aren’t standardized, it will recur.

Paper can become the frontline interface for managing AI

The AI employee roster itself should be managed online.

But to make it understood on the frontline, paper can help.

AI employee ID cards.

Job assignment cards.

Lists of acceptable vs unacceptable inputs.

Handover conditions to humans.

Emergency stop contact sheet.

Incident report forms. Here, an incident is a record of an accident like information leakage—or a near-miss.

Training booklets.

Don’t end AI governance as an unread “policy PDF.” Convert it into something usable on the floor.

Visualize with print.

Train and drill through experience.

Manage permissions, costs, logs, and updates online.

This is a new kind of support a printing company can offer in the AI era.

Historically, printed matter has been used not only for “information transmission” but also for “standardizing behavior.”

• Standard work instructions
• Checklists
• Near-miss logs
• QC process charts—tables listing what to check at which step for quality control

AI governance is the same: paper turns “rules” into frontline tools.

If you were to test something in the next 48 hours

Before building a big AI strategy, inventory the AI already inside your company.

Day 1

Ask every department to declare the AI and automations they are using.

Record the owner, cost, purpose, data, and last use date.

If you only “ask for declarations,” you’ll miss things. Common misses are:

• Paid with an individual’s credit card
• Personal Chrome extensions—browser add-ons that are hard for the company to see
• Dify apps built “temporarily”

So make Day 1 concrete for the frontline:

• Take a 15-minute meeting slot per department and fill it in by reading out loud together
• Put “examples” in the declaration form (meeting minutes, translation, image generation, transcription, inquiry reply)
• Clearly state: “Anything that touches operational data is in scope.”

The goal is not enforcement. It’s locking down what exists.

Day 2

Classify into three buckets:

• Continue
• Consolidate
• Stop

Then select just one AI from the “continue” bucket and measure results.

How many minutes did estimate-start time shrink?

How many rework cases dropped?

Did close rate change?

Did gross profit increase?

Don’t end with “it feels more convenient.”

Also write the falsifications (patterns that don’t work).

• Falsification 1: Time drops, but neither revenue nor gross profit rises

• The freed time just gets absorbed into other work and doesn’t connect to orders or pricing.

• Countermeasure: design the reallocation of freed time (example: create an “instant estimate reply” slot, increase the number of proposals).

• Falsification 2: Quality drops and rework increases

• AI-generated text or inputs are subtly off, increasing rework in later steps.

• Countermeasure: create inspection checkpoints for AI outputs (in printing terms, proof checks and inspection). Proof checks are the step where you confirm there are no errors in plates or data. Decide “where to stop it.”

An AI leader isn’t a company with lots of AI employees

A company that has adopted a lot of AI isn’t necessarily an AI leader.

A strong company has AIs with clear work, accountable owners, measurable profit, and the ability to stop when no longer needed.

One AI employee that increases gross profit by 500,000 yen per month is clearer for management than 100 convenient bots. Annualized, that’s 6 million yen.

The 400 AI employees found in Las Vegas aren’t a story that denies AI’s future.

It’s the signal to move AI from experiments into management.

---

Add-on: “Products and business” hypotheses for printing companies (concept)

If we translate the article’s conclusion (count AI, stop AI, connect AI to profit) into a business offering, it splits into three parts. From here onward is BPJ’s concept.

1. Support to build an AI employee roster (one-time project)

• Who for: administrative functions at mid-sized companies, IT (information systems) departments, frontline managers
• What: roster templates, facilitation of inventory meetings, classification through (continue/consolidate/stop)
• Risk: it touches permissions and personal information, so define scope and handling clearly in the contract (don’t discuss customer names or unpublished pricing)

2. Frontline AI governance printed materials (kit)

• Who for: factories, sales, production—frontlines where “policy PDFs” are rarely read
• What: AI employee IDs, prohibited-input cards, emergency stop flow, training booklet, incident report forms
• Inventory: since revisions are frequent, assume on-demand printing only as needed. Enforce strict version control to avoid accidents where old pre-revision versions remain on the floor

3. Ongoing operations (BPO model)

• Who for: companies where AI is increasing and management can’t keep up
• What: monthly updates to the AI employee roster, organization of costs and usage logs, operation of evaluation meetings
• Rights and responsibilities: clarify where API keys, logs, and stored data live and who has authority; include retirement (shutdown) procedures

Add-on: A “one-week experiment” proposal at Bunseikaku (in research)

This is an experiment (in research) to reach outcomes measurement within one week after completing the 48-hour inventory.

• Departments: Sales (estimate → order) + Production (proofing and submission handoffs) + IT (permissions)
• Target AI: choose only one AI from those classified as “continue” (meeting minutes, estimate draft, first-pass inquiry reply, etc.)
• Steps:

1. Monday: draw the current process on one sheet (where, how many minutes, who stops it)
2. Tuesday: turn AI input rules into cards (input OK/NG, redline examples)
3. Wednesday: decide inspection points (where a human approves)
4. Thursday: run the same work in parallel with “AI on/off” for only three cases each (compare with a small sample and surface incidents early)
5. Friday: aggregate metrics and re-judge continue/consolidate/stop

• Metrics (examples):
• Waiting time before work starts (minutes)
• Number of rework cases (cases)
• Lead time to order (time from start to completion, sales cases only)
• Number of near-miss cases on the floor (AI-caused)
• Decision criteria:
• Not only time reduction: if rework increases, stop or add inspection steps
• Even if improvement is small, continue if risk is low and it can be standardized

---

#BeyondPrinting #BeyondPrintingJournal #BPJ #printing #printingcompanies #AI #AIemployees #AIagents #AIinvestment #ROI #printingDX #Dify #Claude #ChatGPT #Bunseikaku