Data does not remove humans from a decision. It makes it harder to pretend that a preference, fear, or bet was a fact from the beginning.

TL;DR

I want to use AI and data to reduce the space occupied by impulse, selective memory, and the loudest opinion in the room. That is useful. But data does not choose a goal, a value, acceptable risk, or a human consequence. Metrics, samples, and models carry choices made before the chart appears. A data-informed decision is not a decision without humans or emotion: it is one in which evidence, hypothesis, preference, and responsibility are separate, documented, and verifiable. Agents can look for contrary evidence, compare scenarios, and expose uncertainty. That does not make them moral arbiters.

There is a sentence that sounds mature in any meeting room:

“Let’s decide based only on data.”

I understand the intention. Most of the time, I agree with the discomfort it is trying to solve.

Too many decisions come from convenient memory, theatrical urgency, hierarchy, vanity, a hallway conversation, or the person who can sound certain for fifteen minutes without producing evidence. Organized data can cool down a room, recover context, test projections, and show that the comfortable story had a hole the size of the budget.

The problem is the word “only.”

Data does not decide. It reduces the room you have to lie to yourself.

A chart does not choose what matters

Imagine a company deciding where to reduce cost. A dashboard shows that support is expensive. It seems objective: cut there.

But before it became a decision, someone already chose what counts as cost, which period matters, how quality is measured, what later revenue loss appears, what reputational risk is acceptable, and which customer promise may be broken. The chart did not choose any of that. It made one particular view of the problem visible.

That does not make the number useless. It makes the conversation honest.

A metric can show that average response time fell. It cannot, by itself, tell you whether the customer solved the problem, whether hard cases were pushed into another queue, or whether the team started ending conversations early to protect the metric. A model can estimate a higher chance of delay. It does not decide whether the consequence should be a message, human review, a process change, or an automatic penalty.

Goals, limits, and consequences remain human decisions.

The National Institute of Standards and Technology (NIST) published the AI Risk Management Framework: a voluntary guide for organizing how an organization identifies, measures, monitors, and reduces AI-system risks. It does not solve that by decree. It is broad and does not certify an architecture. It does reinforce the right discipline: scope, expected benefits, costs, risk tolerance, and human oversight need to be defined and documented throughout the lifecycle.

Data has backstage choices too

There is a comfortable fantasy that data arrives raw, neutral, and ready to rescue the meeting from imperfect humans.

It does not.

Someone decides what to record. Someone fails to record something else. Someone defines a category, removes duplicates, chooses a time window, handles missing values, and decides that a set of fields represents “quality,” “risk,” or a “healthy operation.” A model, when one exists, learns and calculates on top of those choices.

That is not a moral accusation against spreadsheets or statistics. It is a description of the work.

Datasheets for Datasets proposes recording a dataset’s motivation, composition, collection process, and recommended uses. The goal is not to turn every table into paperwork. It is to prevent the recipient from treating a collection of choices and limitations as a natural fact.

AI can tighten the right screw. It should not hold the company keys.

This is where I see the promise.

A good agent can do, in minutes, work that often dies between the meeting and the decision: gather scattered documents, compare versions, find numbers that contradict the initial thesis, expose hidden assumptions, simulate scenarios, and list what remains unknown.

It can ask:

• Which source supports this number?
• What contrary data was left out?
• Does the projection change if the conversion rate — the share of people who complete the expected action — falls by two percentage points?
• Are we measuring a cause, a symptom, or only what is easy to measure?
• Which hypothesis became “fact” because it appeared in three reports?

That is much better than using AI to write an elegant paragraph confirming a decision already made.

It is the same requirement I described in The best answer isn’t the one that pleases me most: I do not want a system that agrees with me in good grammar. I want one that can withstand the question, “How do you know?”

An agent can organize alternatives, show the cost of each, discover an incomplete base, and register that a projection depends on a fragile premise. It should not decide alone what harm is acceptable for a person, team, or customer. That is not a processing gap. It is responsibility.

Four layers that change the conversation

Layer	Question it forces	Example
Evidence	What was observed, and where did it come from?	“Median response time rose 18% in period X, according to source Y.”
Hypothesis	What explanation is plausible but still needs testing?	“The increase may be related to the queue change.”
Preference	What do we want to prioritize?	“We prefer to preserve quality even at higher cost.”
Decision and responsibility	Who chooses, accepts the risk, and answers for the consequence?	“Leadership approves a limited pilot, with review in 30 days.”

When these layers become one sentence—“the data shows we should cut support”—a decision has already entered disguised as a discovery.

Separating them returns disagreement to the right place. You can agree with evidence and reject its interpretation. You can accept a hypothesis and reject the decision. You can prefer another value, as long as you own the cost instead of calling your preference technical inevitability.

That is the discipline behind A hypothesis does not become fact because AI repeated it. Repeating an inference in three dashboards does not make it a discovery. It only makes the problem better designed.

Less emotion does not mean less human

I defend using data to reduce decisions captured by impulse—especially when money, deadlines, political pressure, or the need to defend an old choice are involved.

But “removing emotion” is a bad formula if it means pretending emotions, bonds, and values do not exist.

In a personal decision, fear may signal real risk or a protection that no longer serves. Desire may show priority. Discomfort may reveal a conversation that is necessary. In a company, a team’s concern can expose an exception the dashboard cannot see. None of that becomes proof by itself. Neither is it noise an algorithm has the right to sweep out of the problem.

The mature work is to identify when emotion is being sold as evidence—and when evidence is being used to hide a value choice.

A business example without stage magic

Consider an operation that wants to use AI to prioritize support tickets.

The data can show volume, waiting time, affected product, customer history, recurrence, and estimated impact. An agent can group patterns, flag contradictions, and prepare capacity scenarios for the team.

But someone still needs to decide whether a large customer always goes first, whether a security issue is escalated regardless of revenue, how personal data is protected, when a recommendation must stop for human review, and how a customer can contest an error.

If nobody owns those choices, “the system prioritized it” becomes a very convenient sentence. It is also an expensive way to outsource responsibility to an architecture that never asked for the job.

An agent as a useful adversary

If I am building a system to support me, I want it to have permission to disagree within clear limits.

Not through automatic moralizing. Not through false neutrality that lists five sides and calls that depth.

I want useful contradiction:

1. record the question and the decision it is intended to support;
2. separate observed source, transformation, and produced hypothesis;
3. look for evidence that weakens the initial thesis;
4. compare scenarios with explicit assumptions;
5. state what could not be measured or verified;
6. return a recommendation with risk, alternatives, and a next step;
7. require the responsible person when the choice involves value, rights, material impact, or a hard-to-reverse consequence.

This does not turn AI into an oracle. It turns it into a verifiable part of a conversation that used to happen entirely inside someone’s head—or, worse, in a presentation with handsome transitions.

At i-9.ai, that is why a useful conversation begins with the bottleneck, available data, exceptions, and who owns the result—as I explain in Your company does not need to discover where to put AI. The tool comes later. Responsibility comes first.

Data is not an absolution

In the end, the most useful question may not be “what does the data tell us to do?”

It is this:

What is evidence here, what is hypothesis, what is preference, and who is willing to answer for the choice?

When those answers appear, the decision remains human. It simply becomes less comfortable to lie to yourself.

References and limits of use

• NIST AI Risk Management Framework 1.0: supports defining scope, risk, cost, tolerance, and oversight across the lifecycle; it is voluntary, not a certification or a replacement for contextual judgment.
• Datasheets for Datasets: supports documenting data origin, composition, collection, and use; documentation does not eliminate bias or choose which values matter.
• Model Cards for Model Reporting: supports declaring model uses, evaluations, and limitations; transparency is not a performance guarantee or a correct decision.
• Does Automation Bias Decision-Making?: supports the lens of undeserved authority given to automated aids; the experiment did not study current language models and should not be transferred to them automatically.

These sources do not prove that every data-informed decision will be better, that an agent is neutral, or that human responsibility can be automated. That conclusion is my synthesis of architecture and governance.