AI Should Not Be Part of Every Task

The real challenge of enterprise AI transformation may not be using more models, but knowing where a model is genuinely useful.

If we wanted to measure how advanced a company’s use of AI is, the first thing that came to mind would probably be the number of models.

How many employees use ChatGPT? How many agents are running? How many processes include AI? Which large language model has the company chosen?

Yet Zapier’s data for the second quarter of 2026 contains a striking finding: even in the AI-enabled workflows of the companies using AI most intensively, only 18% of the steps are performed by AI.

The remaining 82% consists of conventional automation: moving data, rules, filters, API calls, recordkeeping, sending messages, scheduling, and similar tasks.

At first, that seems strange.

Why is 82% of the workflow at companies leading in AI transformation not AI?

Perhaps that is exactly the question we should be asking.

Zapier’s research does not represent the average company. From 1,500 mid-sized and large customers, it examined the top 25% using AI workflows most intensively: 375 organizations. So this is not a statistic that tells us how all companies operate. It offers something more interesting: how the leaders in AI adoption design their systems.

And apparently, the most mature architecture does not send everything to a model.

It uses the model where it is needed.

This distinction may seem small, but it could significantly reshape companies’ AI investments over the next few years.

You do not need AI to retrieve an exchange rate

You need to retrieve outstanding receivables from an ERP.

There is no point asking an LLM to do that.

You need to check whether an invoice is overdue.

You do not need one for that either.

Running a report at nine, updating a CRM field, moving data from an API to another system, calculating an amount with a formula…

These are deterministic tasks.

Given the same data, we want them to produce the same result as consistently as possible.

AI’s value begins somewhere else.

Is a customer’s late payment a genuine risk signal or an operational hiccup?

Which team should receive an incoming request?

Is a particular clause in a contract unusual?

How urgent is the problem in a customer’s message?

What is the likely cause of a variance in a report?

Here, we are no longer merely moving data.

We are interpreting it.

That is why the four roles in Zapier’s study are useful: Communicator, Clerk, Analyst, and Coordinator.

The first two are relatively familiar.

The Communicator reads, summarizes, or writes something for people. The Clerk takes unstructured information and enters it into a system, for example extracting a customer name, action, date, or sentiment from a conversation and writing it to CRM fields.

At the Analyst stage, the task changes. The model produces a decision, score, or classification, and that result determines the next step.

With the Coordinator, the result turns into action: a task is opened, a ticket is created, or a team is mobilized.

Among the advanced users in the study, 84% have a Communicator and 79% a Clerk. Some 45% use an Analyst and 26% a Coordinator. Moreover, most companies first establish roles that write or record information, then move on to roles that make decisions and initiate work. The median time from the first AI role to the second is 143 days. Subsequent transitions fall to 111 and 99 days.

This sequence may be no coincidence.

If a summary is poor, a person corrects it.

If the wrong value is entered in a CRM field, it can be reversed.

But if a model’s decision directly activates another system, we are no longer talking only about output quality.

We are talking about authority.

What may the model decide?

Up to what amount?

At what confidence level?

When is human approval required?

Who will notice a wrong decision?

How will it be reversed?

Beyond a certain point, AI stops being a technology issue and becomes a question of how authority is designed.

The employee is getting faster. What about the company?

This is one of the most important distinctions in enterprise AI today.

An employee producing a report in thirty minutes instead of three hours is a real productivity gain. We should not dismiss it.

But it is not the same thing at company level.

In McKinsey’s 2026 research, 80% of respondents say AI has improved their individual productivity. By contrast, 37% report at least some positive effect from AI use on company EBIT. Only 6% are defined as high performers and report an EBIT impact of at least 5% from AI.

That gap matters.

An employee may be writing emails faster.

Researching more quickly.

Preparing presentations in a few hours.

But if a customer request still circulates among four departments, collection risk is still tracked manually in someone’s spreadsheet, and pricing decisions still wait in three separate approval queues, the employee has become faster.

The company has not.

A CEO survey published by BCG in July draws a similar distinction. Many participating companies report cost or revenue benefits from AI in particular areas, yet struggle to extend the impact across the business. High-performing companies are about seven times more likely than others to redesign workflows end to end.

To me, this also explains why AI transformation is moving more slowly than expected in many companies.

Companies invest in the model.

But often, they do not change the work itself.

The old process remains.

AI is inserted at a few points.

The result is an old process running faster.

That is not a bad outcome.

It simply is not transformation.

Perhaps no new interface is needed

There is another misunderstanding here.

AI transformation can sound as though all of a company’s existing systems must change.

The ERP must go.

The CRM must be replaced.

A new agent interface must arrive.

People must work somewhere else.

In practice, I am not convinced that this is always necessary.

For many companies, the real opportunity may lie in adding a separate intelligence layer over their existing systems.

The ERP can remain the system of record.

The CRM can continue to hold the core customer record.

Teams or Slack can remain the communication environment.

Excel can continue to serve some analyses.

AI’s role may be less about replacing them and more about reducing the blind spots between them.

A recent McKinsey study gives this an apt name: the ‘coordination tax.’

For years, companies optimized the boxes in their process charts.

Sales.

Finance.

Operations.

Procurement.

Legal.

But work often slows down not inside a box, but as it moves from one box to another. Information has to be passed on, explained again, interpreted, and approved; context gets lost…

One of AI’s biggest opportunities may lie here.

Then a fundamental problem emerges: Where is the company’s memory?

An approach recently described by Zapier CEO Wade Foster is interesting in this context.

Zapier has long been a remote company. Foster encourages internal Slack conversations to take place in accessible shared channels rather than private messages wherever possible. They have even created a transparency ranking among executives.

According to the latest figures he shared, 79% of his own messages are in shared channels. For the company’s CMO, the figure is 98%. He had previously said that the share of public Slack messages across the company rose from 33% to 46%.

At first glance, this looks like a corporate culture story.

For AI, it means something else.

Foster puts it plainly: when a decision is made in a private message, an AI agent cannot see that context.

The outcome may exist in a system.

But it does not know the reason.

Companies’ institutional memory has suffered from this problem for years.

The ERP does not record why an investment was postponed.

The CRM does not always explain why a customer received unusual payment terms.

The rationale behind an executive’s acceptance of a particular risk may be absent from the meeting notes.

An experienced sales manager may know that a customer is riskier than they appear, but that knowledge may exist in no database.

Human organizations have handled this informally for years.

We call someone.

Ask in the corridor.

Say, ‘What was the story behind this?’

AI has no corridor.

It knows the company only through the data and context it can access.

Of course, it would be wrong to conclude that every conversation should be open to everyone. Foster specifically notes that legal, HR, and sensitive matters should remain private.

The real issue is different:

We now have to consciously design which parts of institutional memory should become the company’s shared memory.

Because the usefulness of AI inside a company will not be determined solely by the model’s intelligence.

It will also depend on how much accurate context about the company the model can access.

Finance makes this distinction especially clear

Consider a finance workflow.

I would not want to use an LLM to pull bank balances into a system.

Nor to retrieve outstanding invoices from the ERP.

For calculating DSCR or due dates, creating debt repayment schedules, and aggregating foreign-exchange positions, I want the architecture to be as deterministic as possible.

If the numbers are the same, the result should be the same.

Then come different kinds of questions.

Why has this customer’s payment behavior deteriorated?

What are the two most important risks that could create a cash shortfall over the next three months?

In a currency shock, which payment decision would have the least operational impact?

Is this variance temporary or structural?

This is where AI begins to make sense.

Reading data, finding relationships, evaluating scenarios, listing possible causes…

But when it is time to decide which customer should receive credit terms, which investment should be postponed, or which financing offer should be accepted, another layer comes into play:

Human judgment.

Then the decision returns to the system.

A task is opened.

An alert is sent.

A collection action is recorded in the CRM.

The scenario is run again.

Perhaps this is where the most important feature of good AI architecture emerges:

The model does not do everything.

It occupies the right place in the work.

Why does 18% matter more than we think?

In separate cost modeling across 24 example AI-agent workflows, Zapier also calculated that execution costs could be 71% lower if deterministic steps were handled by conventional automation rather than an LLM. This was not a finding from actual customer bills; it was a simulation based on specific model and pricing assumptions. The 71% should therefore not be read as a universal savings rate.

But the underlying idea is powerful.

Sending everything to a model is not merely expensive.

Sometimes it introduces unnecessary uncertainty.

We do not need an API call to be creative.

We do not want an accounting entry to offer interpretations.

A payment date saying, ‘I think the 15th would be better,’ would be a disaster.

In enterprise systems, it is good for some things to be boring.

It is even necessary.

AI creates value where we add reasoning on top of that boring infrastructure.

That is why I think the definition of an AI-first company may also change over the next few years.

It may not be the company with the most models.

Nor the one running the most agents.

What may matter more is knowing which work to give the model, which work to keep deterministic, which decisions to bring to people, and how to monitor all of it.

Seen this way, Zapier’s 18% does not mean AI is being underused.

Quite the opposite.

It shows that we are beginning to learn where to put it.

And perhaps technology is not even the hardest part of enterprise AI transformation.

It is the company having to describe how it works clearly enough for the first time.

Where is each piece of information?

Why is each decision made?

Who has which authority?

When does a human step in?

Which tasks truly require judgment?

If we do not know these things, we can still connect the most powerful model to the system.

The model will be fast.

The company will remain the same.

Real transformation will begin not when AI starts to understand the company, but when the company makes its own operations clear enough for AI to understand.