The last mile is where enterprise AI actually dies

McKinsey’s 10,000‑leader survey and HBR’s ‘last mile’ diagnosis show that 30 years of consulting have built organizations that cannot turn AI into real value.

The Last Mile of AI transformation

TL;DR:

HBR’s “Last Mile” problem and McKinsey’s State of Organizations 2026 show that after decades of transformation work:

Most enterprises are complex and unready for change,
Experimenting with AI without significant bottom‑line impact,
Structurally unprepared for day‑to‑day AI.

So the real constraint is not models or infra, but the missing, under‑funded System 2 Production Layer that decides whether enterprise AI lives or dies.

---

---

I read the McKinsey numbers twice to make sure I wasn’t misreading them.

At the first glance, The State of Organizations 2026 is another glossy transformation report. Underneath, it quietly documents that three decades of “change” have produced organizations that are complex, unready and nowhere near able to turn AI into real value.

When you place that report next to HBR’s “last mile” summit notes from Harvard and Microsoft, you see the same system from two angles: one as aggregated charts, one as lived friction inside real firms.

The “last mile” here is the point where AI outputs are supposed to change how the company actually works.

Everywhere I look, companies are rolling out copilots and agents, building hundreds of pilots and process automations. Yet when you ask for firm‑level impact, you get silence, hand‑waving or a headcount plan. The “last mile” is where enterprise AI quietly dies.

The story is about what we chose to build (and fund) over the last 30 years.

---

What McKinsey admitted in print

Their 2026 report is built on a survey of more than 10,000 senior executives across 15 countries and 16 industries. The introduction shows leaders are now obsessed with “sustained productivity and long‑term impact.” They see technology and AI as the core of that push. So far, so good.

Then you hit the numbers.

Not ready for change.

“72 percent of leaders tell us that their organizations are not fully ready to face upcoming changes. Even among leaders who are optimistic, only one‑third feel prepared.”​

Overly complex and inefficient.

“Two‑thirds of leaders think their organizations are overly complex and inefficient,”
and that traditional remedies (structural redesigns, cost cuts, flatter hierarchies) are delivering “diminishing returns.”​

AI adoption without bottom‑line impact.

“Less than 20 percent of companies that have tried to adopt the technology have seen significant tangible impact on their bottom lines.”
at the same time, “88 percent of organizations are deploying AI in at least parts of their organizations,”
yet “just as many report no significant bottom‑line impact.”​

Unprepared for day‑to‑day AI.

“Eighty‑six percent of leaders feel that their organizations are not prepared to adopt AI in day‑to‑day operations.”
One in six organizations have no clear C‑level owner for AI adoption,
and only 14% see leaders “consistently championing AI adoption and experimentation with clear strategies and action.”​

Missing the human engine.

Only 20% of leaders believe nonfinancial rewards can instill performance in employees.
In other words, most leadership teams still do not understand, or do not believe, their own people science.

And then, in the middle of the AI section, you get the buried phrase that explains why this all feels so familiar.

Companies that actually do see EBIT impact from AI note that “capturing this value depends as much on people as on technology investments.”

One executive noted:

“for every $1 spent on technology, $5 should be spent on people.”

If I strip away the consulting language and keep only the data, the picture is scary:

• We have organizations that are structurally complex and inefficient

• Leadership teams that know they aren’t ready for change

• Near‑universal experimentation with AI

• Less than one in five companies seeing significant bottom‑line impact

• Almost nine in ten leaders saying they’re not ready to embed AI into daily operations

• Very few leaders truly championing AI

• And one executive quietly stating the only reasonable ratio: 5:1 people to tech.

This is an industry admitting, in its own numbers, that the transformation medicine did not cure the disease.

---

HBR’s last mile: inside the pilot graveyard

Then I read the HBR piece on The “Last Mile” Problem Slowing AI Transformation, where authors describe a closed-door summit at Harvard Business School with senior leaders from a dozen large organizations that are already enthusiastic AI adopters. They already have hundreds of deployments and almost universal access to tools like M365 Copilot, ChatGPT Enterprise or GitHub Copilot.

And they are stuck:

“The primary obstacle to progress is rarely model quality or data availability, but rather the ‘last mile’ of transformation where technical capability must meet organizational design.”

HBR’s summit notes are very consistent. A global investment bank has more than 250 LLM-connected apps in production, a food and beverage giant runs pilots across 185 countries, an apparel group has automated over 18,000 finance processes. Another participant, a global payments network, says over 99% of employees now use copilots. And an industrial manufacturer reports double‑digit productivity gains for thousands of engineers.​

And yet, when their finance teams go hunting for impact in headcount or cycle-time numbers, they largely come up empty.

HBR finds the real reason:

The gains sit inside local workflows
because nobody has redesigned roles, budgets, and processes
to harvest the freed time.

AI exposes decades of process debt and overgrown controls. In some companies, human‑in‑the‑loop governance that worked for a few use cases breaks once they start running hundreds of agents.

If you overlay this on McKinsey’s figures, you get a complete picture:

• The macro view by McKinsey says:
  Complex, unready, heavy AI experimentation, little bottom‑line impact, deeply unprepared for day‑to‑day AI,

• The micro view by HBR says:
  Hundreds of pilots, 99% Copilot adoption, double‑digit productivity in pockets, but value trapped because no one owns the last mile into the operating model.

That last mile is not a technical gap. It is a System 2 gap.

---

What dies in the last mile: jobs

When organizations don’t have a credible plan for the last mile, AI’s “value” shows up in exactly one place. The staff list.

Atlassian’s March announcement is a clean yet ugly example. The company is cutting about 10% of its workforce (roughly 1,600 people) as part of a restructuring to “enhance focus on AI and enterprise sales.” It expects restructuring charges between $225 million and $236 million, with around $169-174 million for severance and $56-62 million for office reductions. Mike Cannon‑Brookes didn’t hide the logic: AI changes both the skills and the number of roles Atlassian needs.

WiseTech Global is going even further. In late February, it said it will cut around 2,000 jobs – roughly 29–30% of its approximately 7,000‑person workforce – over the next 18–24 months, as part of a “strategic AI transformation initiative.” The CEO has described internal work that used to take 6–7 months now taking a day, and customs‑expansion work that took up to two years now being done in a fraction of the time with AI.

I’ve been in these rooms and I’ve seen the same decisions.

When there is no serious Production Layer, no redesigned operating model, no agreement on what to do with freed capacity, the spreadsheet will do the only thing it knows. It will move headcount down until the numbers balance.

This is exactly what McKinsey and HBR tell you to expect. McKinsey’s leaders are under pressure to “reestablish high performance” and “break through the productivity ceiling.” They already tell you that traditional productivity levers are exhausted and that AI is the new lever. HBR shows you that saved hours are currently “re‑absorbed into low‑value activities,” not structurally harvested.

If you are a CFO in that environment, and you see AI‑driven efficiency with no credible plan to redeploy people into new, verified work, you have two options. Admit that you don’t have a real AI strategy. Or cut people and call it AI transformation.

There is a third move I see all the time.

Many companies have designed their IT around pushing more of the work into external IT and systems integrators.. They are expensive, but reducible – you can always cut the contract next year.

On paper, that looks like flexibility. In practice, you are outsourcing the one capability you actually need in an AI era: the ability to redesign and re‑wire your own systems.

You talk about moving IT from “cost center” to “business enabler,” then hand that role to a partner whose business model is billable hours, not your margin.

No external SI will genuinely behave like your business enabler unless they also capture a big share of the value. I’ve seen this in Cloud era, and I see this again in AI era.

---

What it looks like when System 2 finally shows up

Disclosure: I have no commercial relationship with Glia. I happily use them here because they are one of the first public examples of what a serious System 2 product looks like in a regulated domain.

On March 11, Glia announced what they call an “industry‑first contractual guarantee” against AI hallucinations and prompt injection for more than 700 banks and credit unions using their Banking AI platform. They promise: no hallucinated content or prompt‑injection output ever reaches a customer or member.

Technically, they do three things that matter:

• They separate understanding from response:
  LLMs help understand intent, but final answers are constrained by a proprietary approvals framework and pre‑approved content.​

• They wrap this in bank‑grade controls:
  automated PII redaction, end‑to‑end encryption, malware scanning, continuous audits and no independent sharing of PII.

• They turn that into a contractual guarantee. Not a slide. A promise you can sue them over.

That is a Production Layer product. It treats AI not as a tool, but as a source of liability that has to be controlled with explicit verification and governance.

---

What happens when you skip System 2

On March 9, a federal judge in San Francisco granted Amazon a preliminary injunction against Perplexity’s Comet browser agents. The court accepted Amazon’s view that, even though users had given Comet permission to log into their Amazon accounts, Perplexity did not have authorization from Amazon itself. The order blocks Perplexity from accessing password‑protected sections of the site and requires them to destroy previously collected Amazon data.

Technically, the agent worked. Legally and commercially, it did not.

A minimal Production Layer would have treated “log in to Amazon and buy things for the user” as a high‑risk pattern, not a default action. It would have checked site terms, enforced a policy that agents cannot access password‑protected areas of third‑party platforms without an explicit commercial agreement, and routed those flows into either a human brokered experience or an approved Amazon API.

In other words, the control plane would have blocked or reshaped the behaviour that got Comet sued. System 1 executed the task. System 2 was missing when it came to deciding whether that task should be allowed at all.

The McKinsey Lilli breach is the same failure in a different costume. Recently, a security startup pointed an autonomous agent at McKinsey’s internal LLM platform and, via a decades‑old SQL‑injection bug in an unauthenticated API, gained read‑write access to the production database in under two hours, spending about $20 in tokens.

Independent write‑ups say that vulnerability exposed tens of millions of internal chat logs, hundreds of thousands of files and tens of thousands of user accounts. McKinsey acknowledges the flaw and says it has found no evidence of client‑confidential data being exfiltrated.

Technically, Lilli is a capable internal assistant. Architecturally, it shipped without a Production Layer that forbids unauthenticated paths into sensitive data, segments access by design, and red‑teams the surface with the same kind of agent that eventually hacked it.

System 1 (a powerful internal assistant on top of a huge RAG corpus) was strong. System 2 was simply not present.

When you put these stories next to each other, you start to see a clear direction.

Banks and platforms will not tolerate “pilot‑rich, transformation‑poor” approaches once AI touches real money or real customers.

A functioning Production Layer stops being a governance nice‑to‑have and becomes the-must-have.

If you don’t build it, someone else (like Glia) will sell it to you – or a platform or regulator will simply shut your agents down. Or, as McKinsey just discovered with Lilli, your own AI system will eventually get hacked in ways a serious Production Layer should have prevented.

---

The 5:1 correction

McKinsey’s anonymous executive is right. For every $1 you spend on technology, you should be prepared to spend something like $5 on people.

Not in general. In AI specifically.

For the last three decades we poured money into systems, tools, reorganisations and dashboards and treated the operating model – workflows, roles, verification, governance – as a secondary stream. We built enormous System 1 capacity and almost no System 2.

Let’s revisit the numbers.

McKinsey’s report shows what that buys you:

• Less than 20% of companies that tried to adopt AI seeing significant, tangible bottom‑line impact.​

• 86% of leaders feeling unprepared to adopt AI in day‑to‑day operations.​

• Only 14% seeing leaders consistently championing AI with clear strategies and actions.​

• One in six companies with no C‑level AI owner at all.​

HBR shows how that feels on the ground: 250+ LLM apps, 18,000 automated processes, 99% Copilot usage – and productivity gains that “remain trapped inside individual workflows” because no one has redesigned roles and budgets to capture the reclaimed time.​

The wrong conclusion from this is “AI doesn’t work” or “we need bigger models.”

The right conclusion is:

We funded the wrong part of the system.

If I take the 5:1 line seriously, the corrective move is boring and radical at the same time: put the Production Layer on the budget as a first‑class asset.

That means:

1. Owning verification.
   Treat every AI output that touches a customer, a decision, or a ledger as a hypothesis, not a fact.
   Build workflows that test those hypotheses against trusted sources before they act.

2. Owning agents.
   Build “agentic control planes” like the ones HBR describes: dashboards and policies that define:
   who can create agents,
   what actions they can take,
   how they’re monitored, a
   nd how they’re retired.​

3. Owning roles.
   Redesign jobs so that humans become orchestrators and verifiers of AI‑driven work, not random human‑in‑the‑loop friction.
   Make “agent orchestration” and “AI process architect” real roles, not slideware.

4. Owning time.
   Decide, explicitly, what happens to the reclaimed hours.
   If you do not, they will disappear into the organisation’s noise, and the only visible “value” will be layoffs.

I am tired of drama around AI statistics. Last year’s 95% failure headlines were a perfect example of how to miss the point.

The story is not that AI is disappointing. It is that we have spent 30 years optimising for the wrong layer of the system, and we are still trying to fix a System 2 problem with more System 1.

---

The 90‑day collapse points

If I’m right – that enterprise AI dies in the last mile because we under‑funded the Production Layer – you should see the pressure show up in four places over the next 90 days.

1. AI‑labelled restructurings.

By the end of Q2 2026, at least three more large software or services firms (public, roughly >5,000 employees) should announce restructurings or layoffs where AI is explicitly named as the driver – language like “AI‑driven efficiency,” “funding AI investment,” or “pivot to AI” in their filings or earnings calls.
If instead those announcements lean toward role redesign and redeployment without net cuts, it would suggest some organizations are starting to build a real Production Layer, and this Signal weakens.

2. Banking‑grade guarantees.

By June 30, 2026, watch whether at least five additional vendors selling into banking, insurance, or healthcare publicly offer contractual guarantees around AI behaviour (for example, hallucination‑free outputs, prompt‑injection defenses, or similar), not just “responsible AI” marketing language on their websites and press releases.
If Glia remains an outlier and no one else is willing to underwrite their AI in contracts, it means the Production Layer still isn’t being treated as a standalone market.

3. Budget splits in plain sight.

In Q2 earnings calls, investor days, or annual reports, look for at least two major enterprises that disclose AI or “digital” investment plans where 40% or more of new spend is explicitly earmarked for people, process redesign, and governance – beyond licences, infra and tools.
If every disclosed budget keeps 80–90% of the money on technology, the 5:1 correction McKinsey hints at has not started.​

4. Consulting offers that lead with the last mile.

Between now and the end of the quarter, see whether any of the big firms (e.g. McKinsey, BCG) reposition their flagship AI offers so that “last‑mile operating‑model redesign,” “agent control planes,” or “Production Layer build‑outs” are front‑and‑center rather than buried as change‑management streams.
If they keep selling models, infra and “double transformations” as the main act, expect the last mile to keep killing enterprise AI.

If all four tests fail, this Signal either collapses or becomes a slower‑burn story.
If they hold, then we are simply watching the same pattern from multiple angles: McKinsey’s charts, HBR’s summit notes, AI‑linked layoffs, banking guarantees, and legal injunctions all pointing at the same missing layer.

---

If you want more of this kind of work

I write these Signals between working with teams who are trying to make AI stop dying in the last mile. Each one takes me 6–8 hours to research, cross‑check and write.

If this helped you see your own organization more clearly, the best way to support it is simple:

and

---

Key Sources

McKinsey – The State of Organizations 2026, HBR – “The ‘Last Mile’ Problem Slowing AI Transformation”, Reuters / CNBC / TechCrunch – Atlassian 10% workforce cut to self‑fund AI, The Straits Times / ABC – WiseTech 2,000 job reduction in AI shift, Glia press release – banking AI hallucination and prompt‑injection guarantee, GeekWire / Search Engine Journal – Amazon preliminary injunction against Perplexity’s Comet agent, Andrei Savine – “Stop AI drama – System 2”