AI Cannot Discover Your Intent. It Can Only Discover Your Drift.

There is an objection to authored Business Intent that sounds like progress.

It goes like this.

“Why should leadership spend weeks authoring its Business Intent? Our systems already know. Every transaction, every approval, every exception is in the data. Point AI at the event logs and it will reconstruct our operating model for us. Faster, cheaper, and based on evidence instead of opinions.”

It is a genuinely attractive argument.

It is also one of the most dangerous shortcuts in enterprise AI.

Because it rests on a quiet substitution.

It replaces the question “What did we decide?” with the question “What did we do?”

Those are not the same question.

And no amount of data can convert one into the other.

The distinction matters because enterprise AI is rapidly moving beyond analysis and recommendation. Organizations are now deploying agents, copilots, autonomous workflows, and decision-support systems that increasingly act on behalf of people. The quality of those actions depends not merely on whether AI understands historical behavior, but whether it understands the Business Intent leadership wants preserved.

That is where the shortcut breaks down.

What Process Mining Actually Sees

First, credit where it is due.

Process mining and process intelligence platforms are remarkable instruments. They reconstruct how work actually flows through the enterprise with a fidelity no workshop or interview can match. They expose variants, bottlenecks, rework loops, and shadow paths that leadership never knew existed.

If you want to know what is happening in your enterprise, mining the logs is among the most powerful tools ever built.

But look carefully at what the log contains.

It contains behavior.

Every workaround someone invented under deadline pressure in 2019.

Every exception that quietly became the norm because nobody closed it.

Every approval threshold that exists because a consultant needed a number by Friday.

Every regional variation that survived three reorganizations without anyone deciding it should.

Every default setting nobody ever examined.

The log records all of it with perfect neutrality.

What the log cannot tell you is which of those behaviors were decisions and which were accidents.

It cannot reveal whether a threshold exists because leadership deliberately created it, or because an implementation team chose a convenient configuration years ago. It cannot distinguish a carefully governed operating principle from a workaround that simply survived long enough to become familiar.

Behavior leaves evidence.

Business Intent does not.

Business Intent must be authored.

Interpretation Debt, Formalized

For years, every enterprise has been accumulating something I call interpretation debt: the growing gap between what leadership assumes the enterprise means and does, and what its people and systems have actually come to mean and do.

Interpretation debt builds silently.

A definition bends here.

An exception hardens there.

A threshold drifts.

An accountability shifts.

No single instance matters.

The accumulation does.

Over time, the enterprise develops a second operating model that nobody explicitly designed. It emerges from thousands of local interpretations, compromises, shortcuts, adaptations, and historical decisions.

Now consider what happens when you mine your processes and call the output your Business Intent.

You are not discovering Business Intent.

You are formalizing the debt.

Every unexamined workaround gets promoted to policy.

Every accidental threshold becomes a rule.

Every drift that crept in over a decade gets ratified in a single modeling exercise, stamped with the authority of “the data,” and handed to machines to execute faithfully.

The enterprise does not end up governed by its Business Intent.

It ends up governed by its history.

This Is Not My Objection. It Is Theirs.

What makes the shortcut untenable rather than merely risky is that the very disciplines being invoked have already recognized the boundary.

Process mining excels at discovering behavior. It reveals what happened and how work flowed through the organization. Its value comes from exposing reality, not defining the desired future state.

Likewise, the broader discipline of behavioral inference faces a similar challenge. Observed behavior can often support multiple explanations. The same pattern may be consistent with several different intentions, objectives, constraints, or assumptions.

Behavior alone cannot tell you which interpretation leadership intended.

Something outside the data must break the tie.

That something is always a set of assumptions.

Which raises the only question that matters.

There Is No Intent-Free Inference

Every inference engine must make choices.

Which data matters?

Which actions signal importance?

Which exceptions are meaningful?

Which patterns represent intent rather than noise?

Which outcomes are considered successful?

Those choices are not discovered.

They are supplied.

The uncomfortable reality is that there is no such thing as intent-free inference.

When data cannot uniquely determine meaning, assumptions fill the gap.

Someone decides which model is preferred.

Someone determines what constitutes success.

Someone chooses the boundaries.

Someone defines the labels.

Someone establishes the objectives.

And those choices are themselves expressions of intent.

The question is not whether intent exists.

The question is whose intent is governing the model.

Your Business Intent or Someone Else’s?

Once this becomes clear, the popular framing falls apart.

The choice is not authored Business Intent versus inferred intent.

The choice is your governed Business Intent versus someone else’s ungoverned intent.

If leadership does not explicitly define what the enterprise is trying to preserve, then that definition will emerge from somewhere else:

• Historical behavior

• Vendor assumptions

• Model design choices

• Configuration defaults

• Training data patterns

• Local interpretations

Every one of those sources exerts influence.

None of them were necessarily authorized by leadership.

No board would knowingly delegate its operating principles to undocumented assumptions buried inside a model.

Yet organizations do exactly that when they assume AI can infer their Business Intent purely from observed behavior.

What appears to be a shortcut is often an abdication.

Why AI Raises the Cost of Confusing Them

For decades, confusing behavior with intent was survivable.

Organizations documented processes imperfectly. Policies drifted. Definitions became inconsistent.

But humans remained in the loop.

A controller questioned an unusual result.

A manager challenged a strange exception.

An experienced employee recognized when a rule technically applied but should not.

Human judgment functioned as a continuous correction layer.

AI changes the equation.

An intelligent agent does not ask whether a rule was accidental.

It does not wonder whether a threshold still makes sense.

It does not question whether an exception reflects deliberate policy or historical residue.

It executes.

Consistently.

Quickly.

At scale.

And it executes whatever was provided.

Including the drift.

Especially the drift.

This is the AI governance problem that many organizations have not fully recognized.

The greatest risk is not that AI ignores instructions.

The greater risk is that AI faithfully operationalizes assumptions nobody intended to preserve.

Inference does not create the problem.

It industrializes it.

The Enforcement Layer Enforces. It Does Not Author.

At this point, a reasonable executive will object that this is exactly what modern AI governance is for.

And the machinery is genuinely impressive.

Enterprises now deploy an entire enforcement layer around their AI: deterministic controls that prevent what is out of bounds, probabilistic guardrails that screen for likely harm, and monitoring that detects and records what occurred. Much of it is now customer configurable. Policies, permissions, thresholds, and evaluation rubrics can all be tuned to the enterprise.

This layer is necessary. It is improving quickly. Nothing here argues against it.

But notice what every one of those mechanisms has in common.

Each applies the logic it is given.

A configurable guardrail is an empty form. A permission system governs what a system can do, not whether what it does within its authority is what leadership intended. A monitor records behavior, but it cannot judge that behavior without a standard to judge it against.

The enforcement layer enforces. It does not author.

Which means the enforcement layer cannot rescue inferred intent. Hand it logic reconstructed from your history and it will enforce that history with precision: the workarounds, the accidental thresholds, the drift, all of it, applied consistently and at scale.

Better enforcement of an unauthored standard is not better governance.

It is faster drift.

Nothing Can Be Validated Against Nothing

There is a reason every discipline that proves correctness works the same way.

Auditors test against approved policy. Quality teams test against a specification. Controllers reconcile against the ledger. In every case, objective evaluation requires a reference: an explicit standard that exists prior to, and apart from, the thing being evaluated.

No reference, no evaluation. Only opinion.

Now apply that to the enterprise deploying AI.

Ask what the intended state is, the reference against which system behavior, agent decisions, and implementation choices should be evaluated, and in most organizations the honest answer is that no such reference exists. There are strategies, process documents, requirements lists, and tribal knowledge. There is no governed, decision-grade statement of what the enterprise intends that execution can be objectively compared against.

This is why authored Business Intent is not documentation.

It is the reference.

Leadership must author the Business Intent of the enterprise, including the definitions, boundaries, Conditions of Success, exceptions, accountability requirements, evidence requirements, and outcomes that must be preserved through implementation and operations.

Only then does evaluation become objective. Only then can technology, testing, and monitoring determine whether reality conforms to what was decided, rather than to someone’s recollection of it.

Without the reference, AI systems can only optimize the past.

With it, every downstream layer, enforcement included, finally has something legitimate to enforce.

And the Reference Cannot Come From the Players Being Measured

One more property matters, and every controller already knows it.

A reference is only trustworthy if it is independent of the parties measured against it. No auditor certifies their own books. No delivery team writes the acceptance test for its own work and calls the result assurance. When the same party produces both the work and the standard it is judged by, passing was never in doubt.

Now look around the transformation table.

The software vendor has a financial interest in how your intent maps to their product. The implementer has a financial interest in how gaps get interpreted, and in the change orders those interpretations produce. The platform has a financial interest in its own defaults becoming your standards.

I have written elsewhere that every party at the table has a financial interest in the outcome except you. The reference is where that observation stops being a talking point and becomes structural: whoever authors the standard controls what conformance means.

Which is why the reference must be authored by the enterprise itself, on the sponsor’s side of the table, and held apart from every party whose work will be measured against it.

Not as a matter of preference.

As a matter of validity.

The Right Place for Inference

None of this argues against process mining.

It argues for putting it in the position it was always meant to occupy.

Mining is a diagnostic instrument, not an authoring instrument.

Used before Business Intent is authored, it provides discovery. It reveals behaviors leadership may not realize exist.

Used after Business Intent is authored, it becomes even more valuable. It acts as a conformance instrument, showing where actual behavior diverges from the authored reference.

The gap becomes measurable.

Drift becomes visible.

Corrective action becomes possible.

The sequence matters:

1. Observe to discover.

2. Author to decide.

3. Observe again to prove conformance.

The common mistake is collapsing the first step into the second and allowing a reconstruction of the past to substitute for a decision about the future.

Observing behavior and authoring Business Intent are complementary activities.

They are not interchangeable.

The Question Every Executive Should Ask

When someone proposes inferring your Business Intent from enterprise data, ask a simple question:

Since multiple interpretations are consistent with the same behavior, which assumptions did your system use to choose among them, and who in this company authored those assumptions?

There are only two possible answers.

Either those assumptions were authored and approved by leadership.

Or they were not.

If they were, then the enterprise has already done the work the shortcut claimed to eliminate.

If they were not, then the organization is preparing to let machines operate according to assumptions nobody deliberately chose, and to let an enforcement layer apply them faithfully.

Observed process is evidence of what happened.

Business Intent is a decision about what should happen.

Your data can tell you what you did.

AI can reveal your behavior.

AI can reveal your patterns.

AI can even reveal your drift.

But only leadership can author the Business Intent that determines what the enterprise is supposed to become.

Author your Business Intent before implementation begins.