What you see when you dissect
51 successes only

0Before you believe it — this report's limits

Let's flag the conclusion first. This report gathered only success cases, leans on interview self-reports, and keeps company names anonymous.

So reading it as evidence for the probability that "do this and you'll succeed" is a mistake.

The right reading is "structures repeatedly observed in the places that succeeded." You have to read it leaving open the possibility that the places that failed may have done the same things.

Even laying down that premise, the six findings were worth keeping.

177% of the hardest part was outside the technology

Of the points practitioners named as hardest, 77% were change management, data quality, and process redesign — invisible costs.

It was not model performance.

There's one more number in the same vein. 61% of the successful projects had gone through at least one earlier failure. The cost of those failures doesn't show up in the final results calculation.

The practical translation is simple — allocate budget and time to embedding rather than to tools. And don't fold at the first failure. Most of the places that succeeded were on their second attempt.

2Where the same task takes weeks, and where it takes years

The most famous contrast in the report is this. On the same customer-support redesign task, a tech company shipped in 6 months, while a large bank answered "just standing it up takes years." Same model, same task.

The frequency of the factors that decided the speed is counted too.

All three are matters of the organization, not the technology. It overlaps exactly with finding 1.

3Exception review over item-by-item approval — a two-fold gap

The source of the number AXHub cites most often is here.

The escalation approach — where AI autonomously handles 80%+ and a person reviews only the ambiguous cases — had a median productivity gain of 71%. The approach where a person approves every item was 30%.

It doesn't mean remove the safeguard. It means move the safeguard's position — from every item to the exceptions.

The design question narrows to one, too. "What has to come before a person."

4Agentic is powerful, but still a minority

Agentic implementations, which run several steps on their own once given a goal, had the largest effect with a median productivity gain of 71% (simple high-automation was 40%).

Yet of the 51 cases, agentic was only 20%.

The report's phrasing is precise — "agentic is not a new UI but a redefinition of the roles of humans and machines."

The reason it's effective and the reason it's rare sit in the same sentence. Redefining roles is far heavier than buying a tool.

5Resistance comes not from the front line but from support functions

When we counted the sources of adoption resistance, the top one was unexpected. Legal, HR, risk, and compliance came in at 35%, higher than front-line users (23%).

That said, these groups often turned into staunch allies when brought in early. Meet them late and they're a wall; meet them first and they're allies.

There are two more numbers that overturn conventional wisdom. Headcount reduction was the largest outcome in only 45% of cases — the other 55% were hiring avoidance, reassignment, or no reduction. And "we can't do it because the data is messy" was untrue in 88% of cases. LLMs actually unlocked the value of unstructured data like voice records, scanned documents, and legacy code.

6How we used it

We also lay out transparently how these six findings were folded into AXHub content.

• 77% and 61% → cards No.1 (getting started) and No.3 (failure patterns), the "keeping records" chapter of lecture L2.
• 6 months vs years → the closing of card No.1, card No.4 (time-reduction cases).
• 71% vs 30% → cards No.9 and No.12, the "human check" design in lectures L1·L3·L5, glossary "escalation."
• Agentic 20% → the reason card No.12 and lecture L5 start from "five boxes before you build."
• Resistance 35% → the "meet the blocking department first" of card No.8 (governance).

If one figure recurs across several pieces of content, its source converges entirely on this one report — which is why it matters to also know the limits (No.0).

7If you read the original

The report is freely available. You don't need to read all of it.

Recommended reading: look first at the summary and the number tables up front, then pick two or so cases close to your own industry and read those. The number tables are worth more than the case narratives.

And remember No.0 the whole way through — this is the common structure of 51 places that succeeded, not a guarantee of success.