AI Builds Fast Lanes Inside Company Red Lights

Open Google Maps before a drive and you learn something every business should understand.

The route determines the trip. You can fly down one empty segment, overtake three cars, and still arrive at the same time because the next traffic light catches everyone. The average speed from A to B is governed by lights, turns, queues, merge points, roadworks, and the order in which they appear.

That is enterprise AI right now.

Companies are building fast lanes inside red lights. They buy Copilot seats, install chatbot wrappers, run prompt workshops, and celebrate that a person can draft an email, summarise a PDF, or write a first version of a proposal faster than before. That speed is real. Business value arrives when the route changes.

The market has priced AI as organisational transformation. Most companies are deploying it as personal productivity.

The route says otherwise.

The $600B question is where the traffic starts

Sequoia’s David Cahn framed the market problem cleanly in 2024: the AI infrastructure buildout implied a revenue requirement that had grown from a $200B question into a $600B question. His math was simple enough to be useful. Start with Nvidia run-rate revenue, double it for the full data centre cost, then double it again for the margin the compute seller needs.

That is the supply-side wager. The industry is building as if there is a vast pool of profitable AI demand waiting to absorb the cost.

But at the enterprise level, the evidence is much messier. McKinsey’s 2025 survey found AI usage spreading widely, while enterprise-level EBIT impact remained far less common. Gartner warned that many GenAI projects would be abandoned after proof of concept because of poor data quality, weak risk controls, cost, and unclear business value. MIT Project NANDA’s 2025 report drew the same line in harsher language: high adoption, low transformation.

This is the mismatch. The money is being spent like AI has already changed the route. Inside most businesses, it has mostly changed one lane.

“AI can make a worker faster at a task. Company speed comes from a shorter route to cash.”

The task gains are real. That is what makes the trap dangerous

The serious pro-AI argument starts by admitting task gains are real.

Brynjolfsson, Li, and Raymond studied a generative AI assistant in customer support and found a real productivity lift, especially for newer and lower-skilled workers. The HBS and BCG “jagged frontier” study showed the same basic pattern in knowledge work: inside the frontier, AI can lift speed and quality; outside it, people can get worse while feeling more confident.

S&P Global’s 2026 labour report makes the modern version sharper. AI adoption is broad, and the common use cases are exactly the task-level ones you would expect: summarisation, translation, data management, process efficiency, employee productivity. But S&P also found that only a minority of initiatives were live and delivering value, and fewer than half were on track to produce positive ROI within twelve months.

Many red lights are organisational. A review step may exist for compliance, but it may also exist because a team’s authority depends on being needed in the loop. A report may be automated, but the meeting around the report may survive because it gives someone visibility. AI threatens tasks and the organisational rituals people use to prove they matter.

So the issue is the workflow limit. AI can move a local metric while review queues, approval rights, missing data, legacy systems, compliance gates, customer trust, and pre-AI incentives keep the business route slow.

If a proposal used to take eight hours to draft and two weeks to approve, cutting the draft to one hour creates a faster car at the same red light.

Workflow route simulator

Task speed is not route speed.

Speed up the draft and almost nothing changes. Remove an approval queue, connect the system, or move judgement earlier, and the business finally gets faster.

AI spend Business outcome

1. Input Request arrives 4h
2. Draft AI lane 8h
3. Review Judgement queue 24h
4. System CRM and ERP 20h
5. Approval Authority stop 28h
6. Delivery Value lands 4h

There are three levels of AI value

The first is task productivity: one person writes, searches, summarises, analyses, or drafts faster.

The second is workflow productivity: the path from request to resolution gets shorter. The handoffs change. The queues shrink. Judgement moves earlier. Systems update themselves from the same facts.

The third is business model productivity: the company can serve customers, price products, deliver work, or scale revenue in a way that was previously impossible.

Most AI adoption is still stuck at level one while the market is pricing level three.

Grove already gave managers the test

Andy Grove’s High Output Management is suddenly more useful in the AI era.

Grove’s core move was to define management by output over activity. A manager’s output comes from the organisation under their influence: decisions made, bottlenecks removed, and customer outcomes improved.

That is the cleanest test for AI adoption.

AI exposes managers who confuse activity with output. A team that produces more drafts, more summaries, and more internal updates may still produce no more revenue, no faster decisions, and no better customer outcomes.

If AI lets one analyst draft faster, that is the analyst’s local productivity. If it changes how sales qualifies demand, how legal handles standard risk, how finance approves pricing, and how managers decide what deserves escalation, then it starts becoming organisational output.

This is why High Output Management feels more relevant in 2026 than another prompt guide. AI raises the price of weak management.

Throughput is a system property

Operations people already know the answer. Little’s Law connects work in progress, throughput, and flow time. If work sits in queues, the system is slow even when individual stations are efficient. AI makes that constraint more visible.

This is why so many AI rollouts feel impressive in demos and disappointing in P&L. The demo shows the fast lane. The P&L measures arrival time.

The Productivity J-Curve gives the economic version of the same point. General-purpose technologies need new processes, new skills, and new operating models before the payoff appears. AI may be the most powerful version of that history.

The forward-deployed engineer boom is the market admitting the truth

Look at the role everyone suddenly wants: the forward-deployed engineer.

Palantir described the original version as an engineer embedded with customers. A16Z now frames forward-deployed teams as the way AI companies operationalise models into real-world solutions. First Round says founders are using the role to push AI products into legacy workflows.

That is the tell.

If AI were just a model-access problem, the winning business would be pure software distribution. Ship the model. Charge the seat. Watch the value appear.

Instead, the market is rediscovering implementation. Startups are hiring people who can sit inside a customer, understand the messy workflow, connect the systems, negotiate the handoffs, and make the thing actually run.

The title is the signal: the most valuable AI work is moving back into the customer’s messy reality.

There are only two honest paths

The first path is greenfield. Build a new company from zero around an AI-native route. Pick API-first systems. Put shared data at the centre. Decide where human judgement enters before the workflow scales.

That is how a new entrant can attack incumbents. It can skip five departments of old handoffs and design a clean route from the start.

The second path is legacy transformation. This is slower, uglier, and more political. Someone has to enter the business and change one process, one integration, one incentive, one approval right, and one relationship at a time.

Most companies want the economics of the first path while living inside the second.

That is why the mismatch persists. They buy AI as if they are rebuilding the route. Then they deploy it as if they are making one lane faster.

The scarce skill is authority

The scarce person is the one who can see the route and has enough authority to change it. They understand the model, but they also understand the CRM field nobody trusts, the approval queue nobody owns, the KPI that rewards delay, the compliance rule that blocks automation, the customer relationship that resists ticket-shaped simplification, and the manager who will kill the project if it threatens headcount.

That is why AI raises the bar for consultants and workflow architects. The weak ones will sell prompt theatre. The useful ones will redraw the route.

Johnny’s verdict

The better question for an AI project is: “which route gets shorter?”

If the answer is only drafting, summarising, searching, or writing internal updates, you probably have a useful tool and an unchanged business. If the answer changes the path from trigger to cash, request to resolution, or insight to decision, then AI has a chance to become enterprise value.

Even if the AI curve compounds, the question for a company is still whether that compounding enters the route or piles up before the next approval gate.

Stop counting fast lanes. Map the red lights.

Then either build a new route from zero, or put someone inside the old one with enough technical skill, political judgement, and authority to change it.