AI Data Centers, Job Loss, and the Squirrel Problem
The expensive question in business AI is not whether a model can produce an impressive answer. It is whether the organization knows which problem it is solving, which data the system may touch, what success looks like, and what happens to the people who teach it the job.
Those questions matter before a company buys more software, connects more records, or joins the race to build more computing capacity. The global infrastructure numbers are large. The decision inside one business should still begin with one controlled workflow.
What the data-center forecast actually says
AI training and deployment take place mainly in data centers. The International Energy Agency projects that electricity generation serving data centers could rise from about 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030 in its base case.
That is a projection, not an instruction to every company. It describes the direction and scale of demand under the IEA's assumptions. It does not prove that every proposed facility is necessary, that every AI workload creates equal value, or that a small business needs the same infrastructure strategy as a frontier model developer.
The practical distinction is simple:
- The IEA can help leaders understand the larger energy and infrastructure environment.
- A local pilot must determine whether a specific workflow produces enough value to justify its own cost, complexity, and risk.
Read the IEA Energy and AI report.
The job question is a management decision
When employees hear that a company wants them to document their process for an AI system, they may hear a second message: teach us enough to remove you later.
That fear can ruin an implementation before the first useful test. Employees hold the exception cases, customer history, judgment calls, and informal knowledge that rarely appear in a procedure manual. If the organization hides its workforce intentions, the people with the most useful knowledge have a rational reason to protect it.
AI does not independently decide whether extra capacity becomes better service, more products, shorter wait times, different roles, or fewer employees. Management makes that choice. Leaders should state the intended employment treatment before asking a team to participate.
A credible pilot agreement can answer five questions:
- Which workflow is being tested?
- Which employee examples or records may be used?
- Who may review the system's output?
- Which measurements determine whether the pilot worked?
- How will role changes, retraining, or headcount decisions be handled?
This is not a promise that every job remains unchanged forever. It is a promise that the organization will not disguise a workforce decision as a neutral software experiment.
Vendor privacy language is only one layer
OpenAI states that it does not train its models on inputs and outputs from ChatGPT Business, Enterprise, Education, Healthcare, Teacher, or API customers by default. That is an important vendor commitment, but it is not a complete data-governance plan.
A business still needs to review:
- account administrators and their access;
- data-retention settings;
- connected applications and plug-ins;
- files copied into other systems;
- logging and backup behavior;
- outside contractors and subprocessors;
- the contract that applies to the exact product in use.
A broad statement about enterprise privacy should never be generalized to every AI product, consumer account, integration, or configuration.
Review OpenAI's enterprise privacy commitments.
Why narrow pilots beat vague transformation programs
"Use AI across the company" is not a measurable assignment. It is a slogan.
A bounded pilot gives the organization a way to learn without granting the system unlimited authority. Pick one repetitive workflow with a clear human owner. Establish a baseline before adding AI. Measure the same work during the pilot. Keep a manual fallback. Stop if the data boundary, output quality, or approval path fails.
Useful measures may include:
- minutes required per completed case;
- correction or rework rate;
- customer wait time;
- employee review time;
- cost per completed result;
- number and severity of exceptions;
- customer or employee satisfaction.
The measurement must fit the job. A faster answer that produces more corrections is not automatically a productivity gain. A cheaper response that damages customer trust can be an expensive failure.
A 30-day AI pilot that a business can audit
Days 1 through 5: Define the boundary
Choose one workflow. Name the human owner. List the allowed data, permitted tools, prohibited actions, approval points, and stop conditions. Record the current time, cost, quality, and error rate before automation.
Days 6 through 12: Build a draft-only test
Let the system produce recommendations or drafts without sending, publishing, charging, deleting, or changing official records. Keep every output attributable to the exact account and configuration that produced it.
Days 13 through 20: Test the exceptions
Give the workflow incomplete records, conflicting instructions, unusual customer cases, and an unauthorized request. Verify that it stops or escalates instead of improvising. A demonstration with perfect inputs proves very little about production reliability.
Days 21 through 27: Compare the evidence
Measure the same indicators used in the baseline. Count human review time and corrections, not just generation speed. Ask employees whether the system removed low-value repetition or merely added another layer of checking.
Days 28 through 30: Decide deliberately
Scale only if the measured benefit exceeds the cost and the controls survived the exception tests. Otherwise revise the workflow, narrow the scope, or stop. Ending a weak pilot is useful information, not failure.
Use NIST as a map, not a magic stamp
The National Institute of Standards and Technology publishes a Generative AI Profile for its AI Risk Management Framework. It helps organizations identify and manage risks associated with generative AI.
The profile does not certify that a particular product is safe. It gives leaders a structure for questions about governance, measurement, content provenance, privacy, security, and human oversight. A framework becomes useful when it changes a real decision, test, permission, or record.
Read the NIST Generative AI Profile.
The squirrel problem
Connor uses a squirrel as an analogy for a dramatic intelligence gap. A person can observe a squirrel, move its food, change its environment, or build around it. The squirrel cannot understand the planning meeting, the property line, or the long-term purpose.
The analogy is not evidence that today's systems possess superintelligence or independent goals. It is a way to ask a governance question: if systems become much more capable, will people still understand, challenge, and reverse the decisions that shape their work and lives?
The answer should be designed into today's deployments:
- humans retain the final irreversible decision;
- systems receive the least authority required for the task;
- important actions remain attributable and reviewable;
- people can stop the workflow and return to a manual process;
- the organization tests failures before granting broader access.
Those controls are useful even if the most dramatic future never arrives. They reduce ordinary mistakes, permission creep, vendor dependence, and silent automation now.
A better question than "How much AI can we buy?"
The best first question is: "Which result can we prove?"
Global demand may justify enormous data-center investment. A particular business still needs a local answer. Start with one workflow, one owner, one data boundary, one baseline, and one stop condition. Tell employees what the experiment means for them. Measure the entire result, including corrections and review. Preserve a human way back.
That approach does not reject AI infrastructure. It connects infrastructure to evidence. It also keeps the people closest to the work involved in deciding what the system should do, where it should stop, and whether it has earned the right to scale.
Sources
- International Energy Agency, Energy and AI
- International Energy Agency, Key Questions on Energy and AI
- OpenAI, Enterprise privacy
- OpenAI Help Center, How your data is used to improve model performance
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
Common questions
Does AI require data centers?
AI training and deployment run mainly in data centers, but one company's decision should be based on its actual workload, security needs, service requirements, and measured results rather than the global forecast alone.
Will an AI pilot eliminate jobs?
A pilot does not determine the employment outcome. Management decides whether additional capacity improves service, expands output, changes roles, or reduces headcount. That decision should be stated before employees are asked to teach a system their work.
Can OpenAI train on business data?
OpenAI says it does not train on Business, Enterprise, Education, Healthcare, Teacher, or API customer inputs and outputs by default. That vendor commitment does not replace a review of retention, administrators, connectors, third parties, and the exact contract for a deployment.
What is the squirrel problem?
It is Connor's analogy for an extreme intelligence gap. Humans can observe and alter a squirrel's environment while the squirrel cannot understand the plan. The analogy is a prompt to preserve human authority, not a claim that current AI is superintelligent.
Connor T. MacIvor · CalDRE #01238257 · Sync Brokerage, Inc. · DRE #02031490