Epic's 'AI-First' Mandate Delayed Its Disney Game

By Lazar Milicevic · Published October 15, 2026 · 8 min read

You're a founder or small-team lead, and someone just told you to "use AI and ship faster." No task list, no definition of done, no named checkpoint. That is the setup Bloomberg described at Epic Games, and it's worth studying because the part you can copy is the part most likely to hurt you.

Here's exactly how I'd read this report, what it does and doesn't say, and the audit I'd run instead of issuing a mandate.

What Bloomberg actually reported (and what it didn't)

Bloomberg reported on October 12, 2026 that Epic Games delayed its upcoming Disney collaboration and told developers to use more AI tools to finish it. The sources are anonymous people familiar with the project, and the report was relayed via Yahoo Finance.

The confirmed facts:

• The game is a shooter featuring Marvel and Star Wars characters.
• Developers were told to adopt an "AI first" production model using tools such as Anthropic's Claude, with humans managing and guiding the work.
• An internal message from a game director, reviewed by Bloomberg, reportedly said "No more hand-written code" and that all art would be "generated with full AI workflows."
• Epic's communications director Jake Jones confirmed the general direction with a statement: "AI-accelerated development, led by Epic developers."
• Epic has communicated no new release date.

What the report does not establish matters just as much. Most outlets place the original target at March 2027, so the date hasn't passed yet; the delay was announced ahead of time (heise online). Bloomberg's stated reason for the slippage is slow progress: Epic underestimated the staff and time the game would need. Engadget reports the date had already moved from November to March before the mandate arrived.

So the sequence is: the schedule slipped for planning and staffing reasons, then the AI mandate came as a response. The report lists the two side by side but never says one caused the other (Happy Gamer). Anyone claiming this proves AI works, or proves AI fails, in game development is guessing. We don't know the team size, the budget, or what share of the work is AI-generated.

The mandate is the story

If you strip the headline noise, the interesting detail is how the directive was delivered. According to heise, it was announced to staff on a video call in early October. Some developers were surprised, some hoped it wasn't serious, and some doubted it would bring the hoped-for success. Per Quartz, staff also questioned how errors in AI-generated work would be handled.

That last question is the whole ballgame. It's a workflow question, not a tool question.

One caveat I want to be upfront about: whether Epic gave the team a workflow is not something any source states. "Handed down without a workflow" is my interpretation, not Bloomberg's reporting. What's on record is that staff were asking who catches the errors. That's the question a workflow answers and a slogan doesn't.

Context also matters. Epic announced more than 1,000 layoffs in March 2026, with Tim Sweeney telling employees the company was "spending significantly more than we're making," and Epic said it had identified more than $500 million in savings (Yahoo Finance). Disney invested $1.5 billion in Epic in 2024 for an equity stake (Quartz). When the budget is tight and the partner is large, "use AI more" is an attractive thing for management to say. It's also an easy thing to say without specifying anything.

Sweeney, for his part, posted on X about a "huge pivot" to retool developers for "maximum AI acceleration," saying the aim is not to do the same work with fewer people but to build more and better (WDWINFO). That's a strategy statement. It still doesn't tell an engineer on Tuesday morning which task to hand to which tool.

Why "use AI more" fails at small-team scale too

I build AI automations for small businesses, and the pattern repeats at a much smaller scale. A policy ("everyone use AI more") pushes the discovery work onto the people already under deadline pressure. They have to figure out where the tool helps, where it hurts, and who checks the output, all while the original deadline keeps moving.

The wins I've seen come from a different shape of decision:

1. Pick one specific, repetitive job.
2. Define what "done" looks like in checkable terms.
3. Put a human checkpoint where mistakes are expensive.

Take a small invoicing business hand-typing 200 invoices a week (an illustrative round number, not a client figure). You don't tell the bookkeeper to use AI more. You automate field extraction, validate that line items sum to the stated total, and route only the exceptions to a person. Here's the kind of validation gate that makes it safe:

from decimal import Decimal

def validate_invoice(extracted: dict) -> tuple[bool, list[str]]:
    """Return (auto_approve, reasons). Anything flagged goes to a human."""
    problems = []

    line_sum = sum(Decimal(str(i["amount"])) for i in extracted["line_items"])
    stated_total = Decimal(str(extracted["total"]))
    if line_sum != stated_total:
        problems.append(f"line items sum {line_sum} != total {stated_total}")

    if not extracted.get("vendor_id"):
        problems.append("vendor not matched to known vendor list")

    if extracted.get("confidence", 0) < 0.90:
        problems.append("low extraction confidence")

    return (len(problems) == 0, problems)

The model does the tedious extraction. Deterministic code does the checking. A person only sees the invoices that fail a rule. That's how you get hours back without introducing silent errors, and it's the opposite of "use the tool more."

Where AI-generated work breaks: the seams

Games are a brutal version of this problem, so it's a useful stress test even though I have no inside knowledge of Epic's pipeline. Code and art have to fit together. Generated art has to match a style the partner will approve. Generated code has to coexist with a codebase that thousands of hours of human work depend on.

I'm reasoning from general engineering experience here, not from anything the report says about Epic's process. But the failure mode is one I see constantly: AI makes the first 90% feel instant, then the last 10% (integration, consistency, approval) takes longer than it used to, because nobody owns the seams.

You can see how this plays out in a smaller setting. Suppose you use an agent to draft client proposals:

The draft step is identical in both columns. The difference is everything after it. That's the part a mandate never specifies.

Your customers are Disney in miniature

The Disney angle holds a second lesson. When a client with strict brand and quality standards sits at the other end, AI output doesn't get graded on speed. It gets graded on whether you'd stand behind it.

If you sell marketing, bookkeeping, or software, your customers behave the same way. They won't care that a model wrote the thing. They'll care that it's right. Staff at Epic reportedly raised exactly this concern about handling errors. A customer will raise it too, just later and less politely.

This is also why I'd treat "who approves this before it leaves the building" as the first design decision, not the last.

Run an artifact audit instead of issuing a mandate

Here's the one practical thing you can do today. Don't announce that you're AI-first. Audit your artifacts.

Take your last week of work and list every deliverable you produced: invoices sent, emails triaged, reports built, proposals written. For each one, record three things:

• How long it took.
• How much of it is repeatable pattern versus real judgment.
• What happens if it's wrong.

A simple sheet is enough:

deliverable,minutes_per_unit,units_per_week,pattern_vs_judgment,cost_if_wrong,check_effort
invoice,6,40,high_pattern,medium,low
email_triage,2,150,mostly_pattern,low,low
client_report,90,3,mixed,high,high
proposal,60,2,mostly_judgment,high,high

Then pick exactly one deliverable that is high-repeat, low-judgment, and cheap to check. In that example it's the invoice or the email triage, not the client report. Automate that one first. Measure the hours before and after with real numbers, not estimates. If it saves time and the error rate holds, pick the next one. If it doesn't, you lost an afternoon, not a release date.

That is the difference between adopting AI as a workflow and adopting it as a slogan.

Where to go from here

I build AI automations for solopreneurs and small teams, and the artifact audit above is the starting point I'd use for any of them. Invoice extraction with total-checking, inbox triage, and lead follow-up with a review step are the kinds of repetitive jobs that tend to fit this pattern, but treat those as illustrations, not a menu. The tooling around the model should be deliberately boring, because boring is what keeps the error rate flat.

The takeaway

My view, and it is a view, not a reported fact: an AI-first mandate is a management shortcut dressed up as strategy. Telling a team to use a tool more shifts the burden of finding where it works onto people already under deadline pressure. We don't know how Epic's project will turn out, and the reporting doesn't let us conclude AI helped or hurt. What we can say is that the teams most likely to win with these tools will be the ones who map the work first and apply the tool where the math is obvious.

Tools don't ship products. Defined workflows do.

Want more like this?

I publish practical AI automation and GenAI engineering content regularly.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, senior engineer.

Visit bizflowai.io for more.