Most Companies Use AI to Cut Costs. The Bigger Win Is the Other Strand

Stanford HAI's 2026 AI Index reports that 88% of surveyed organizations used AI in at least one business function in 2025. AI-agent deployment, the kind that changes how work actually gets done, remained in the single digits across nearly all business functions. That gap is the whole story.

Most organizations are experimenting. Very few are building. And of the ones that are building, nearly all are building the same thing: an efficiency play. Reduce headcount, shorten cycle times, cut cost to serve. That is a legitimate goal. It is also the smaller half of what AI makes possible.

Published at AI Double Helix framework, the model describes two strands operating in parallel. The first is Internal Efficiency: doing the same work faster, with fewer people, at lower cost. The second is External Value: delivering something to customers that was previously uneconomic, too slow, or simply impossible. Most organizations are three years into the first strand and have barely touched the second.

The second strand is what this post examines: what it actually looks like, why it is harder to build, and how to know whether your organization is pursuing it or just running efficiency projects with better branding.

Why the First Strand Gets All the Attention

Efficiency is easier to measure and faster to justify. You can point to hours saved, headcount avoided, or processing time reduced. The CFO approves it. The board understands it. You can ship a pilot in six weeks.

Building the second strand requires a different question: What could we offer customers if the constraint of cost or complexity were removed? That question is harder. It does not have an obvious answer. It requires redesigning a customer-facing proposition, not just installing a faster tool behind an existing one.

Many AI initiatives stall at the efficiency layer for exactly this reason. Organizations that announced AI strategies in 2023 and 2024 were largely announcing cost programs. Those pulling ahead now are asking what they can build that they could not build before.

What External Value Actually Looks Like

External value is not about using AI to write more content or answer emails faster. Those are first-strand plays dressed in second-strand language.

A genuine second-strand play means the customer proposition itself changes. Three categories are worth examining:

Faster, more accurate customer answers

In industries where customers ask complex questions, including insurance, financial services, healthcare, professional services, and technical products, the cost of a good answer has historically been high. You needed a trained human, often a senior one, often after a wait. That economic constraint meant most customers got a simplified answer, a delayed answer, or no answer.

AI can retrieve and synthesize across policies, product documents, account history, and service records in seconds. The business case is not just cost reduction. A system that resolves issues accurately on first contact, without escalation, affects retention and repeat revenue. Metrics that matter include resolution time, first-contact resolution rate, escalation rate, and repeat contact rate, not just cost per ticket.

A customer-service system that lowers labor cost but gives unreliable answers is not a growth system. Reliability is not optional.

Faster quotes and proposals

In manufacturing, construction, professional services, and B2B generally, the time between inquiry and proposal is a competitive variable. Deals go to whoever responds first with a credible number, especially for time-sensitive opportunities. Historically, compressing that timeline required either more salespeople or simpler pricing, both of which have costs.

AI can review specifications, check pricing rules, reference historical jobs, and produce a first-pass proposal in a fraction of the previous time. Commercial measures that reveal whether this is working include quote-to-order conversion, win rate on time-sensitive bids, and gross margin per quote. Faster response increases the number of opportunities a team can pursue. That is a revenue expansion story, not a cost story.

Services that were previously too expensive to offer

Among the most underexplored categories are customer segments organizations cannot serve profitably at current cost structures. Smaller accounts, lower-margin products, highly customized requests: these get deprioritized because the economics do not work.

AI changes that calculation. When a task that required significant expert labor can be completed much faster with appropriate human review, the economics of a previously unviable service tier can flip. Worth asking is not how many employees can this replace? but rather what customer need were we leaving unserved because it was too expensive?

That question opens an addressable market conversation rather than a cost-reduction conversation.

The Infrastructure Problem Nobody Talks About

Building the second strand is harder than the first because it requires something most organizations skip: reliable infrastructure underneath the interface.

Patrick Gilbert covers this in Never Always, Never Never through a detailed account of how AdVenture Media built its Sherpa system. Brain AI, the original version of the tool, connected to campaign data and produced articulate, confident strategic analysis. It also occasionally pulled the wrong date, misattributed spend across platforms, and hallucinated numbers that sounded plausible but were wrong. The interface looked like the future. The backend was unreliable in ways that were hard to detect.

Nechama, who leads the agency's strategy team, identified the core problem: the tool was asking a large language model to do work better suited for deterministic code. Data retrieval, calculations, structured transformations are not tasks that benefit from probabilistic reasoning. They require fixed logic that produces the same output every time.

Rebuilding the tool meant separating two layers. Deterministic Python pipelines handled data ingestion, calculation, and storage. An AI layer sat on top of clean, trustworthy data and did what AI actually does well: synthesize, identify patterns, generate natural-language insights. Each layer did the job it was suited for.

This distinction matters enormously for anyone building toward the second strand. Demos look good. Interfaces are impressive. Failure comes when you try to rely on a tool for something real and discover the backend cannot support production-level demands. As the chapter puts it: anyone can ship something that looks good in a demo. Building the infrastructure underneath the interface is where real value gets created.

A BCG/Harvard/MIT/Wharton study of 758 consultants reinforces this from a different direction. On tasks within the model's capability range, GPT-4 users completed 12.2% more subtasks, worked approximately 25% faster, and produced work rated approximately 40% higher in quality. On a business-problem-solving task outside the model's reliable capability range, AI users performed 23% worse than the control group. Building without understanding that boundary produces confident-sounding output on the wrong side of it.

> The second strand is not a race to ship the flashiest AI feature. It is an opportunity to build something that actually works.

Why Pilots Do Not Become Operating Systems

Between experiment and production is where most AI value gets stranded. Failure modes are consistent across industries:

Capability mismatch. Employees apply the model to tasks requiring dependable accuracy or domain judgment beyond what the system reliably provides. BCG's study showed that adding prompt-engineering guidance did not help: accuracy actually fell further, to 60% among participants who received prompting guidance on the task in question, compared to 70.6% among those who simply had access to GPT. Training in prompting did not compensate for using AI on the wrong task.

No business owner. Technology teams run the pilot. The operating function never owns the outcome metric. When the technology team declares success and moves on, nothing changes in the actual workflow.

Measuring the wrong things. Organizations track time saved or licenses deployed. Error rate, rework, customer outcomes, and revenue per transaction go unmeasured. Time saved that is not redirected toward higher-value work is not an economic win.

Missing integration. Tools run as standalone interfaces rather than connecting to the systems, approvals, and exception-handling processes that determine how work actually gets done.

A more detailed treatment of where rollouts fail appears in the post on why AI rollouts fail, which covers the structural reasons most AI programs stall before reaching production.

The Organizational Structure That Makes This Work

A centralized AI team owning everything will not produce the second strand. Distributed structure works better:

• An executive sponsor who sets strategic priorities and resolves trade-offs between functions
• A central enablement group that provides approved tools, security controls, evaluation standards, and reusable components
• Business-function owners who choose use cases, redesign processes, and own performance results
• Domain experts who define quality standards and acceptable error thresholds
• Risk, legal, and data teams who establish controls proportionate to the use case

Business owners should be accountable for a specific outcome: quote conversion, service resolution time, new-service margin. Technology owners should be accountable for reliability, integration, and maintainability. Combining both responsibilities in an IT pilot is a common route to weak business impact and no one who answers for the revenue result.

Training should also be role-specific. The AI training for employees post covers what that actually looks like in practice. BCG's evidence suggests the most important thing to train is not prompt syntax. It is task selection: where the tool is strong, where it is weak, and when not to use it at all.

The Human Side Is Not Optional

Employee resistance to AI is often rational, not irrational. People worry about job displacement, loss of professional judgment, accountability for machine-generated errors, and whether efficiency gains will benefit the organization or justify headcount cuts.

Rollouts succeed when employees can see three things. A concrete benefit to their own work: less repetitive preparation, faster access to information, more time for judgment-intensive tasks. A clear accountability model: humans remain responsible for defined decisions, AI outputs are reviewable. A credible reinvestment plan: saved capacity goes toward customers, quality, or development rather than being treated purely as a reduction target.

Organizations that treat the second strand as a cost program will generate exactly that kind of resistance. Those that treat it as a capability expansion, framing it as what they can now offer that they could not before, tend to get more from the same tools.

Stanford HAI reports that AI agents reached approximately 66% task success on OSWorld, a benchmark involving real computer tasks. That is meaningful progress. It also means agents fail roughly one in three attempts. In production use, the practical boundary remains: use AI autonomously only where errors are detectable, reversible, and low consequence. Everything else requires human review in the loop.

A practical look at where agentic approaches make sense versus where workflow automation is the better fit appears in the post on AI agents vs. workflow automation, worth reading alongside this one.

Audit: Is Your AI Program Pursuing the Second Strand?

Use this audit with your leadership team. It takes about 20 minutes. The goal is to separate efficiency theater from genuine external value creation.

---

SECOND STRAND READINESS AUDIT

Run this in a 20-minute leadership session. Answer honestly.

Section 1: Where Your AI Is Pointed

• [ ] Can you name one customer-facing proposition that AI has changed in the last 12 months? (Not a cost reduction: a capability the customer can now access that they could not before.)
• [ ] Is there a customer segment or service tier you currently cannot serve profitably? Have you asked whether AI changes that calculation?
• [ ] In your last AI initiative review, did anyone present revenue metrics alongside efficiency metrics?

Section 2: Infrastructure Honesty

• [ ] Do your AI-assisted tools pull data from deterministic, auditable sources, or are they asking an LLM to retrieve and calculate?
• [ ] Can you identify a specific task where your current AI deployment has a known accuracy limit? Do frontline users know that limit?
• [ ] Is there a human review step with a defined quality threshold before AI output reaches a customer?

Section 3: Ownership

• [ ] Name the business owner (not the technology owner) of your most significant AI initiative. What outcome metric are they accountable for?
• [ ] Has that initiative changed a core process, or is it still running as a standalone tool alongside the original process?
• [ ] Is saved time being redirected toward more customers, better service, or higher-value work, or is it simply absorbed?

Section 4: Task Selection

• [ ] Have you tested AI on specific tasks before deploying broadly? What was the baseline, and what did the test show?
• [ ] Are there tasks in your current AI deployment that fall outside the model's reliable capability range? How do you know?
• [ ] Do your employees know when not to use the AI tool?

Scoring

• 10–12 checked: You are building toward the second strand. The gaps you found are worth addressing now.
• 6–9 checked: You have efficiency programs in place. Reaching the second strand requires deliberate redirection of investment and ownership.
• Fewer than 6 checked: Your AI program is experimentation. That is a reasonable starting point, but it will not produce revenue growth on its own.

---

Run this quarterly. Answers change as your program matures, and the gaps that appear are more useful than any score.

To see where your organization sits on the capability spectrum, the AI Maturity Ladder post maps out what Dabbler, Practitioner, Architect, and Strategist actually look like in practice.

Start Here

One question. Take it into your next leadership meeting.

What customer need are we currently leaving unserved because it is too expensive or too slow to deliver, and does AI change that calculation?

If your team has an answer, build toward it. If no one has thought about it, that is the starting point. The first strand will keep running on its own momentum. Pursuing the second strand requires a deliberate decision.