McKinsey put a number on it this month. More than 80% of companies say they are not yet seeing bottom-line impact from their AI investment (McKinsey, 2026). The headlines read that as a model problem, a training problem, or a change-management problem. None of those is the real bottleneck.
The bottleneck is operating model. The work has not been re-imagined, the rituals around it have not shifted, and the people being asked to make the shift are the same people doing the day job.
Alexis Krivkovich said it cleanly on the McKinsey podcast (link above), that "the day-to-day workflows and the rituals around ways of working will need to fundamentally change". 75% of roles need fundamental reshaping right now. That's not in two to three years, that's right now.
The thing is, the explanation that is not in the McKinsey paper is that three tracks have to run in parallel (business model, workforce, data), and inside the workforce track, two tiers of opportunity have to run at once. The $1k to $10k wins that ship this quarter, and the $500k to $1M Board paper wins that the small wins fund. Skip any track or either tier, and the paradox continues.
The paradox in plain English
The typical business has done three things in the last eighteen months. Bought licenses, run training, and told staff to experiment. Each feels like progress, and each shifts accountability onto the individual. Give someone a Copilot licence and a workshop on prompting, and the assumption is that the next step is theirs, that they will figure it out, build an agent, and redesign the job while still doing the job.
The domain experts have workflow knowledge and often no Claude or ChatGPT capability.
The technically literate cohort have capability and no domain knowledge.
Nobody in-house has the mandate, the time, or the engineering depth to join the two.
The organisation is being asked to transform itself while nobody is resourced to do the transforming.
What the training-first playbook actually delivers
Training as the focal point fails for a reason. AI is treated as a skill to acquire, not infrastructure to build. Staff leave the room with enthusiasm, open Copilot, Claude, Gemini or ChatGPT on Monday, and instantly hit a wall because building basic workflows is easy but anything needle moving takes more experience (and often data + technology capability that most don't have access to).
Claude has no context on their work, their clients, or their process.
Gemini's output reads generic, so editing takes longer than writing would have.
Nobody around them is using ChatGPT or Copilot the same way, so there is no pattern to copy.
The above aren't meant as individual digs at any one platform, those same issues apply across them all with the baseline toolsets most are enabled on.
To be super clear, that is not a failure of the staff member. It is a failure of the operating design.
Where the lighthouse team comes in
One path through is a lighthouse team. Not an innovation lab, not a centre of excellence with slides. A team whose job is to ship production infrastructure inside real workflows and hand the operating pattern to the business, one workflow at a time. Its success metric is the second team using its infrastructure without asking for help.
Three resources separate a lighthouse team from the typical "AI working group".
Dedicated time, full-time, not a Friday-afternoon side hustle.
Engineering capability, somebody who can build context files, write skills, connect tools, and ship infrastructure.
Domain pairing, engineers sitting alongside the experts who know the workflow, the edge cases, and the judgement calls.
Transformation happens by concentrating capability, proving the new model, then propagating it.
The Ramp proof
Ramp published the clearest case study so far. Seb Goddijn's "We Built Every Employee at Ramp Their Own AI Coworker" (Ramp, 2026-04-10) is the blueprint. Ramp hit 99% AI adoption. Not through training, but through a small team building infrastructure that auto-configures on install, surfaces tools inside the workflow, and ships pre-built skills (fast becoming the defacto packaged AI workflow capability across all platforms) through an internal marketplace. Their receipts are worth reading:
Over 350 skills shared company-wide, written once and reused by everyone, appearing organically in the same chat interface through natural language exposing the skill automatically.
One engineer ships a skill and sixty reps level up overnight.
Someone who had never opened a terminal now runs scheduled automations that would have needed an engineer six months ago.
The people who got the most value were not the ones who attended training, they were the ones who installed a skill on day one and got a result.
That last line is the load-bearing one. The product did the enablement, not the training.
What a skill actually is
A skill is a pre-packaged workflow with context built in. Not a prompt, not an agent, not a chatbot trick. When someone types "draft an SOW for this client" into Claude, a pre-built skill fires. The skill knows how SOWs are structured, what clauses belong, which systems to pull pricing from, and what checks to run before handing the draft back. The operator does the thing they wanted to do, speaking in natural language, and the infrastructure surfaces automatically.
The harder truth is that the difference between "we have licenses" and "we have an agentic operating model" is measured in infrastructure, not in seats. A 500-person business with fifty skills wired into daily workflows is operating agentically. A 5,000-person business with 5,000 Copilot seats and no skills is operating the same way it was last year, with a bigger SaaS bill. They've likely pumped out thousands of Agents that workers have to remember the Agent's name, navigate to find it etc. = destined to peter out.
The three parallel tracks
Re-imagining a business for an agentic world is not a single initiative. At a high level it is three tracks running at once, and most organisations are running one of them at best.
Track one: re-imagine the business model. Assume near-zero marginal cost of delivery, customers arriving with their own agents, and the frictions that defined moats for twenty years no longer holding. McKinsey's example: an individual moving money frictionlessly across banks breaks the deposit moat. The question is not "how does AI help us do today's work faster". The question is what the business looks like when the structural assumptions underneath collapse. That is executive work, not a pilot team's.
Track two: take the workforce on the journey. Every person needs to be encountering AI inside their actual work, embedded and organic, pre-configured by someone who understood how to build skills. When the new business model lands, the workforce has been operating with the tools for months, one skill at a time.
Track three: fix the data intelligence underneath. This is the track every business hopes they can skip, and they cannot. Agents pulling from stale CRM records produce confident nonsense. Skills querying an ungoverned warehouse produce numbers that do not reconcile. Platforms that enables skills as capabilities are not the failure point, the substrate is. The trap is treating data as an enterprise-wide overhaul whereas the discipline is to target substrate work at the next $500k to $1M opportunity and ship only what that opportunity needs, and building on-top of that foundation of change.
Small wins fund the big ones
The trap is treating data as a gate, the kind of gate that says "we will start building agents once the data is clean". Every quarter in that posture is a quarter of competitors compounding and the technology frontier moving further away from the oranisation. The way through is to work two tiers at the same time.
The $10k ideas. High-volume workflows where the data is already clean enough. Drafting, summarising, scheduling, routing, internal comms, first-pass analysis. These ship in weeks, and they pay for the infrastructure that underpins everything else.
The $500k to $1M opportunities. Cross-functional workflows where the business model changes, the margin shifts, or the customer experience is rebuilt end-to-end. These require the lighthouse team to have proven the pattern on the smaller wins first.
The small wins are not random time-savers picked because they are easy. They live inside the workflows where the big opportunities sit, and they use the same skill infrastructure the big wins will use. A skill that drafts the first page of a quote today is the same architecture that prices the whole quote next quarter. Small wins that do not point at a big win are productivity theatre, not compounding.
The thing is, this is a sequencing problem disguised as a budget problem.
What the lighthouse model looks like in practice
For the average business, this does not mean building Glass. A 400-person NZ wholesaler does not need Glass. They need three skills shipped on Claude this month, built by one engineer and two domain experts who know what a sales quote actually contains. Anthropic's platform provides the infrastructure, and soon will the others (Copilot, ChatGPT, Gemini), at scale. The missing piece is the team and the mandate.
A practical lighthouse is:
A single team, numbers differ depending on Org size, running full-time for three to six months minimum.
Ideally a ratio of one engineer paired with three to four domain experts on rotation.
The remit is to re-imagine two to three workflows end-to-end and document the pattern so the rest of the business can apply it.
Executive air cover is a single sponsor whose quarterly goals depend on shipping.
What ships in a few months is specific, such as two to three production skills inside real workflows. The authoring pattern documented well enough that a second team can use itm and mapped towards a $500k to $1M opportunity scoped, with at least one engineer inside the business authoring skills unassisted. The skill is the output, and the authoring pattern plus the internal engineer is the compounding asset.
This looks expensive on paper, until the comparison is calculated. License spend, workshop spend, consultant spend, and the opportunity cost of staff building agents in the margins of their day job already adds up. The lighthouse team is not incremental cost, it is reallocation of money already being spent on the approach that is not working.
The harder truth
Claude and ChatGPT are capable enough, and the workforce is smart enough. But nobody has been resourced to build the infrastructure that lets the workforce use AI without configuring it themselves. The paradox McKinsey named is not a mystery, it is a predictable outcome of asking individuals to do what a specialist team can do, and not expecting others who, quite frankly, probably couldn't give a rats about building with AI. Reality is that it's not everyone's cup of tea, the same way not everyone is a Spreadsheet whizz or in to pulling together Powerpoint slides.
The shape of the first step is clear:
Build a lighthouse team that ships infrastructure inside real workflows.
Fix a contained segment of the data substrate underneath while the small wins fund the big ones.
That is not the end state of an agentic organisation, it is the first step toward one, and not many businesses are taking it.
Passionate about all things AI, emerging tech and start-ups, Mike is the Founder of The AI Corner.
Subscribe to The AI Corner
