Your AI Rollout Is a System of Inspection. That Is Why Nobody Is Using It.

In 2006, salespeople hated CRM, and they were right to. The software asked them to type out what was already in their heads, handed their pipeline to management for inspection, and gave them almost nothing back. That reaction was not technophobia. It was an accurate reading of who the system had been built for, and it is the same reading most employees are quietly making of the AI tools being rolled out to them right now.

There is one test that predicts whether a technology wave lands inside an organization, and it has nothing to do with the technology. Does the person being asked to use it gain leverage, or do they only see compliance? Every wave that stuck delivered leverage to the people doing the work. Every rollout that stalled asked them to feed a system that watched them.

I spent an hour on The Integration Layer with Mykle McKiernan, who has led enterprise technology platforms at Amazon and Wayfair, and earlier at Citrix and McKinsey. He has now watched finance react to spreadsheets, operations react to ERP, sales react to CRM, and everyone react to AI. Asked whether this time is genuinely different, his answer was "yes and no is almost always the answer with technology." The no is the useful half.

Four waves, one arc

Line the waves up and the shape is hard to miss.

Spreadsheets. Finance never asked for them. Accountants trusted paper, ledgers, and centralized reporting, and early spreadsheets felt dangerous because suddenly anyone could build a model and anyone could be wrong at scale. Mykle's read is that finance both lost and won: "They lost control, but they gained leverage. Instead of 10 analysts serving the business, everyone became a part-time analyst."

The labor data has been bearing that out ever since. The US Bureau of Labor Statistics projects employment of bookkeeping, accounting, and auditing clerks to decline about 6 percent from 2024 to 2034, while accountants and auditors grow about 5 percent, and it attributes the split directly to automation absorbing routine work while raising demand for analytical and advisory skills. The clerical layer thinned. The profession moved up.

The danger was real too, and it never went away. Raymond Panko's review of operational spreadsheets found that 94 percent of those studied contained at least one error. A spreadsheet mistake was a documented contributor to JP Morgan's multibillion dollar London Whale loss. In October 2020, Public Health England lost track of roughly 16,000 COVID cases because an old Excel file format ran out of rows. We democratized analysis forty years ago and we are still paying for it annually, which is worth remembering the next time someone argues that AI can be made safe by slowing it down.

ERP. This one arrived openly as control. "ERP wasn't really a technology project," Mykle said. "It was a business transformation disguised as software." The software effectively announced that there is now one way to do purchasing, one way to close the books, one way to manage inventory. It unified operations and it transferred power from local business units to standardized processes, which is why the person who had been the only one who knew how to configure a new widget fought it. Their resistance was rational. They were being asked to hand over the thing that made them necessary.

CRM. The purest version of the pattern, because the value flowed one direction for years. Management got forecasting, pipeline visibility, and customer insight. The salesperson got data entry, lost the ability to manage their own numbers, and watched accounts get reassigned to newer people. "What's in it for the salesperson?" Mykle asked. "Not a whole lot." He is blunt that the hatred was earned: "It was a rational reaction from the salespeople."

CRM only worked when that changed. The platforms that won became, in his phrase, "systems of engagement rather than systems of inspection." The turn came when a rep who used to cover ten enterprise accounts could cover twenty, and could pull their own probability of a deal closing rather than reporting it upward. The moment the tool started answering questions for the salesperson instead of only about them, adoption stopped being a fight.

Which brings us to now

Most enterprise AI deployments resemble early CRM with uncomfortable precision. Management sees visibility and is delighted by how much data there suddenly is. Employees see additional work and new reports to generate. The dashboards measure adoption rather than delivering it.

The survey data reads exactly like a workforce doing the math on that. PwC's 2025 Global Workforce Hopes and Fears survey found 54 percent of workers had used AI for their job in the past year. Pew, surveying US workers, found 52 percent worried about future AI use in their workplace and 33 percent saying they feel overwhelmed by it, against 36 percent who feel hopeful.

Those numbers are usually presented as an emotional problem to be managed with communications. They are better read as an accurate assessment. A large share of workers have used these tools, which means their worry is informed rather than superstitious. They have looked at what the rollout gives them and what it asks of them, and they have noticed the direction of travel.

Mykle thinks the deeper reason this wave feels sharper is that the previous three were about process and this one is about identity. Spreadsheets democratized analysis, ERP standardized process, CRM standardized relationships. AI, he argues, democratizes intelligence, and that sits closer to how knowledge workers understand their own worth. Earlier technologies automated activities. This one appears to automate thinking, and whether or not that perception is accurate, it is why the reaction is louder.

The test: harmony, not balance

The useful framework he offers is not about employees alone. Ask what each stakeholder gets, and check that none of them is purely paying.

The standing set is employees, customers, shareholders, and vendors, plus regulators in industries where they bind. Then look at the distribution. If a program exists mostly to satisfy a regulator, it is a compliance activity and it will be treated like one. If customers get all the value and employees work twice as hard, it will not hold. If the vendor's entire margin has been negotiated away, it is not sustainable either, and Mykle is direct that there has to be something in it for the vendor.

His important qualifier is that this is not an even split. "It's not about balance," he said. "It's about harmony." Sometimes the regulator is irrelevant. Most of the time the customer and the shareholder matter. The requirement is not that everybody gets 20 percent. It is that no stakeholder is asked to be a pure payer over a multi-year horizon.

Run any AI initiative currently in flight through that. Name what each of the five gets this year, and what they get in three. If the employee column is empty, or contains only the word "productivity," the program has an adoption problem that no amount of enablement content will solve.

The Wayfair story that shows what real adoption looks like

The counterexample in the conversation is worth more than any framework, because it describes a tool whose value was invisible until someone tried to remove it.

Wayfair had deployed Glean for knowledge discovery, and by Mykle's account it was genuinely useful, especially for the volume of new employees who needed to get up to speed and could not work out where anything lived. Then margin pressure arrived and the seat count was cut from 3,000 to 1,000. His description of the result: "There was almost a religious war to say, I can't do without that."

Then the line that should stay with anyone running an AI program: "That was a fantastic tool that we didn't really know where it was getting adopted and who was dependent on it until we looked to take it away."

That is what leverage looks like from the inside. Not a usage dashboard, not an enablement campaign, not a mandate. People fighting to keep something because removing it would make their work harder. Most organizations have no idea which of their AI tools would produce that reaction, and there is a cheap way to find out that does not involve a survey.

The governance fight is the ERP fight wearing new clothes

The other thing that repeats is the argument about who decides.

Sit in enough executive rooms and the AI governance debate turns out not to be about models at all. One team is waiting on Copilot credits. Another has been requesting access to a frontier model for months while leadership deliberates. Every department is arguing about how much freedom it has. That is the ERP standardization fight almost line for line: how much power sits with the local unit and how much with the center.

Mykle's answer is the same one that was right then. One size does not fit all, and there is still real value in formal standardized governance. Those are not contradictory. Some analysis gets federated, some does not, and the split is situational rather than doctrinal. Not every decision needs five nines of reliability. "A lot of things are just directionally representative."

The practical version, which lands squarely on the model proliferation everyone is currently celebrating: not every process deserves a frontier model. Some deserve a small model you host yourself. Choosing a flagship model for a job that a local one would do is the same category of error as buying a camper van to drive ten kilometers to the office. It will work. You will pay for it every day.

Two cautions he adds that belong in any procurement conversation. First, these commitments behave like the ERP ones did: many are one-way doors on a two or three year horizon, so map who your vendors partner with and who they quietly refuse to integrate with before you sign. Second, and this one is free: be suspicious of any vendor whose case studies are all successes. "Anybody that only has success stories is giving you the tip of the iceberg, the top 15 percent." Ask what failed, what they learned, and how they applied it. A vendor with no failures is not a vendor with no failures.

What to do Monday

The finish line is invisibility

The closing idea is the one that should reframe how you read every AI announcement this year.

"The technology is never the destination," Mykle said. "The destination is adoption. The final stage of every successful technology is invisibility." Nobody says they are leveraging client-server computing. Nobody announces that they are consuming cloud infrastructure. People simply do their jobs.

Robert Solow made the same observation from the other side in 1987, when he noted that the computer age was visible everywhere except in the productivity statistics. Electrification took decades to appear in the numbers, not because the technology was weak, but because the gains only arrived once factories were physically redesigned around what electricity made possible. The lag was organizational, not technical.

That is the actual work in front of most enterprises, and it is not a model selection problem. It is the unglamorous business of redesigning how work happens so the tool returns something to the person using it. The organizations that get there will stop calling it AI at some point, the way nobody calls a spreadsheet a decision support system anymore.

Everyone else will keep buying tools that watch their employees and wondering why adoption is flat.

If you are rolling AI into an organization and want a second pair of eyes on whether it is delivering leverage or just visibility before it reaches everyone, reach out.

Shubhendu Tripathi is an AI and ERP strategy consultant based in Toronto, and the host of The Integration Layer, a podcast on AI, enterprise systems, and the work of making them fit together. Connect on LinkedIn or reach out at tripathis@qubittron.com.