Are You Making the Most of AI? Most Businesses Are Barely Scratching the Surface

I’m not a neuroscientist. Give me an LLM and I could produce a fascinating explanation involving neurochemistry, quantum physics and the extraordinary computational machinery between our ears. It might even sound convincing. But the old claim that people use only 10% of their brains? MIT says that is a myth.

The business version of that myth is more interesting: how much of AI’s practical capacity is your organization using? There is no honest percentage to put on the full potential of a technology that keeps changing. Yet plenty of companies have given employees a chatbot and called it an AI strategy. That is a little like buying electricity to charge a flashlight while leaving the factory running on steam.

Short answer: Most businesses can get more value from AI by moving beyond individual prompts and embedding it into measurable workflows, software, data and decisions. The goal is not maximum AI usage. It is faster cycle times, better customer outcomes, lower operating friction, and new capabilities that would otherwise be uneconomical.

AI adoption is widespread. AI advantage is not.

The Stanford 2026 AI Index reports that 88% of surveyed organizations used AI in at least one business function in 2025, up from 55% in 2023. But AI agent deployment was still in the single digits across nearly every function. And a McKinsey global survey found that roughly one-third of respondents’ organizations had begun scaling AI across the enterprise.

Now zoom out from surveys of organizational leaders to the wider U.S. business population. The U.S. Census Bureau’s Business Trends and Outlook Survey found that 17%–20% of firms reported using AI in the previous two weeks during late 2025 through spring 2026. In May 2026, use was reported by 37% of firms with at least 250 employees, compared with fewer than 20% of firms with four or fewer employees. Information and finance businesses reported higher usage rates than retail.

These figures are not measuring the same thing. One is a survey of organizations that report AI adoption; the other samples U.S. businesses and asks about recent use. Read together, they reveal a real gap between having access, using a tool, and changing the business. That gap is where your strategy belongs.

Early adoption helps only when you change the way work gets done

Technology history is full of companies that purchased the latest system and changed absolutely nothing about their operating model. The internet did not reward every business with a website. Cloud computing did not make every migration cheaper. And being first to buy a new AI license will not automatically improve margins.

What creates a more durable advantage is redesigning work around a capability competitors cannot simply switch on overnight. In a 2026 McKinsey analysis, companies attributing at least 5% of EBIT to AI were three times more likely to pursue broad operating-model redesign and twice as likely to redesign workflows before choosing tools. This is a correlation among surveyed companies, not proof that redesign alone causes those financial outcomes. But it points toward a sensible management question: what can the organization now do differently?

A faster email is nice. A sales process that qualifies, researches, routes and follows up without dropping a lead is a business system.

What does using AI beyond ChatGPT actually look like?

Think in four layers. Each layer builds on the previous one, and none requires that every task become autonomous.

1. Assistance: employees draft, analyze, translate, research and summarize with a model. This is useful, but value often remains at the individual level.
2. Augmented workflows: an AI step extracts contracts, classifies requests, enriches CRM records, drafts proposals or reviews software changes, with deterministic validation and a human checkpoint.
3. Connected systems: the workflow reads and writes approved data through APIs, queries internal knowledge, calls existing machine-learning models, and routes outcomes to Salesforce, ERP, ticketing or custom applications.
4. Delegated operations: tightly scoped agents run on an event or schedule, use several tools, preserve an audit trail, and escalate uncertain or consequential decisions.

An LLM does not replace every other form of software. Use deterministic code for arithmetic, identity checks, eligibility rules and irreversible transactions. Use predictive ML when a calibrated statistical forecast is the right tool. Use language models for unstructured information, explanation, drafting and tool orchestration. This combination is often safer, easier to test and more economical than asking one model to improvise everything.

For the engineering architecture behind that distinction, see our guide to the AI agent harness.

Where the value lives depends on your business model

Industry matters, but so do size, gross margin, repeatability and how the business makes money. Here are examples to evaluate, not promises of automatic returns.

Business model	High-value AI workflow	Measure the result
Professional services	Research → client intake → scoped proposal → deliverable quality review	Time to proposal, win rate, rework, billable capacity
B2B sales organization	Lead enrichment → qualification → CRM hygiene → tailored follow-up	Qualified meetings, pipeline velocity, data completeness
Subscription software	Support classification → reproducible issue → code/test suggestion → customer update	Resolution time, escalations, retention, defect escape rate
Manufacturing or field operations	Document/SOP search → exception detection → work-order draft → supervisor approval	Cycle time, throughput, downtime, safety exceptions
Marketplace or transaction platform	Listing enrichment → matching → anomaly flag → reconciliation	Conversion, fraud loss, manual review cost, fulfillment quality
Regulated business	Evidence collection → document reconciliation → policy checks → human sign-off	Audit preparation time, accuracy, exceptions and traceability

For small firms, the first win may be removing an expensive handoff between three systems. For a large enterprise, the bigger challenge may be identity, permissions, governance and cross-department orchestration. Neither should begin with a 50-agent org chart.

An AI agent doesn’t need a desk, or a permanent server

One overlooked opportunity is the ephemeral agent: an AI-powered process that starts when work appears, receives limited permissions, produces a useful artifact, and exits. You pay for the execution instead of operating an always-on bot, though model usage and infrastructure can still carry costs.

Imagine a repository that needs a weekly maintenance assessment or a pull request that needs an additional review. GitHub Actions can provision a temporary runner, check out a repository and invoke Claude Code. The agent inspects the relevant code, returns findings, then the job terminates.

Here is a deliberately narrow, manually triggered example using the official Claude Code GitHub Action. Add this as .github/workflows/ai-repository-review.yml and configure the appropriate Anthropic credential as a GitHub secret. Authentication and repository permissions must be set up for your environment before use.

name: AI repository review
on:
  workflow_dispatch:

permissions:
  contents: read
  id-token: write

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          prompt: |
            Read this repository. Identify three specific
            opportunities to reduce maintenance effort,
            with file references and test suggestions.
            Do not modify files, publish code, or open a PR.
          claude_args: '--max-turns 6'

This example requests repository read access and instructs the model not to make changes. Treat that prompt as a behavioral instruction, not a security boundary: enforce the necessary tool restrictions and token permissions separately. A production workflow should pin trusted actions, cap runtime and cost, review untrusted inputs, and decide where findings will be saved. Anthropic’s action security guidance is worth reading before enabling unattended write access.

Once the pattern is established, the same architecture can process document inboxes, summarize weekly sales variance, verify product catalog changes, assemble compliance evidence or prepare an operations report. A job starts, uses tools, leaves an artifact, and disappears.

Can AI replace licensed SaaS, or should it just connect to it?

Sometimes the best AI strategy is to integrate existing software. Sometimes it is to replace a narrow, expensive workflow with a custom tool. The distinction is commercial, not ideological.

• Keep the platform when it provides essential compliance, network effects, contractual guarantees, deep domain features or a robust system of record.
• Automate around it when the platform is sound but its handoffs, search, data entry and reporting create excessive labor.
• Replace a specific capability when a costly license supports a small, predictable set of workflows you can maintain more cheaply and safely.
• Consolidate when several applications duplicate the same capability and the migration economics are favorable.

Before canceling a subscription, include implementation, hosting, maintenance, support, migration, audit requirements, vendor exit costs and the opportunity cost of maintaining custom software. Our application rationalization guide covers how to inventory and score a software portfolio before making those calls.

Does AI really improve productivity? Measure it, don’t assume it.

It can, substantially in some contexts. In research on more than 5,000 customer-support agents, AI assistance increased issues resolved per hour by about 14% on average, with larger gains among less-experienced workers. But the effects differed sharply across people and tasks.

There is also counterevidence to indiscriminate deployment. A 2025 randomized study of experienced open-source developers found that AI tools increased task-completion time by 19% in that specific setting. The lesson is not that AI is bad at coding; it is that a demo, a benchmark or an employee’s perception of speed is not a substitute for measuring the end-to-end job.

At the workflow level, track elapsed time, error rate, cost per completed outcome, human review minutes and customer impact. Record the baseline before adding an agent. Count rework and verification. Recovered hours only become financial value if the business uses them for more output, better service, lower spend or avoided hiring, not simply because a spreadsheet declares them saved.

A practical 30-day plan for an AI-ready business

1. Week 1: Map the friction. Interview process owners. Identify three repetitive, high-frequency workflows with clean inputs, a clear owner and measurable outcomes. Establish a baseline.
2. Week 2: Build one controlled pilot. Connect a model to the real data and necessary tools. Prefer deterministic steps around the model; define approvals, access and failure handling.
3. Week 3: Test ugly cases. Use actual examples, including missing documents, conflicting CRM records, confusing customer language, and bad API responses. Measure quality, elapsed time and human effort.
4. Week 4: Decide based on evidence. Expand, revise, or stop. If it works, give it ownership, monitoring, a security review and a support plan; then move to the next workflow.

Avoid treating “number of AI seats purchased” as a success metric. Prefer cost per resolved issue, time from inquiry to proposal, revenue per qualified lead, days to close the books, or another metric directly related to your economic model.

Frequently asked questions

How do I know whether my company is underusing AI?

Inventory where AI touches a process today. If its use ends when a person copies an answer from a chat window, while the surrounding data gathering, approvals and handoffs remain manual, there may be an opportunity to go further. Prioritize based on workflow friction and expected value, not a hypothetical percentage of “AI capacity.”

Do we need to train our own AI model?

Usually not as a first step. Many companies can make progress with a hosted or open model connected to their proprietary data, deterministic rules and existing applications. Fine-tuning or specialized models become worth considering when repeated evaluations reveal a measurable limitation that simpler approaches cannot solve.

Can an AI agent use Salesforce, ERP systems and machine-learning models together?

Yes, provided those systems expose suitable interfaces and permissions. The agent can interpret a request, call an existing ML forecast, validate results with rules, read approved ERP data, and propose a CRM update. Sensitive writes and irreversible actions should use scoped credentials, auditing and human approval when appropriate.

Should we deploy autonomous agents across every department?

No. Start with bounded processes where completion and correctness can be evaluated. Use a fixed workflow when the steps are known; give an agent more discretion only when unexpected paths genuinely require it. See our enterprise AI adoption work and operations and automation services for the implementation areas involved.

The next advantage isn’t a better prompt

The next advantage is an organization that can turn a new capability into a dependable process before its competitors do. Sometimes that means a smart assistant. Sometimes it means a custom application, an ephemeral agent or an overhaul of the workflow itself. The answer depends on your customers, your systems and your economics.

Are you making the most of AI? The useful question is simpler: What does your team still do by hand that a well-designed system could handle, verify or accelerate? That’s a conversation worth having.

Launchpad Technology Group helps teams find and deliver those opportunities through enterprise AI adoption, custom software development, technology strategy and systems automation. Bring us a workflow, not a buzzword.

Research and technical references

• MIT McGovern Institute: the 10% brain myth
• Stanford HAI: 2026 AI Index, economy and adoption
• U.S. Census Bureau: AI use by business size and sector
• McKinsey: State of AI 2025 survey
• McKinsey: operating model advantage (2026)
• NBER: Generative AI at Work
• METR: experienced developer productivity study
• Anthropic: Claude Code GitHub Actions documentation