Shreyas Nair is founder and founder of artificial intelligence Wordsworth AI.
I recently spoke with a friend at a large company that had just completed another AI pilot evaluation. The pilot went well. The tool summarized documents, compiled next actions, extracted action items, and made the old workflow look painfully slow by comparison.
However, the mood in the room changed when the team talked about a full production run. The data was scattered in too many places. No one was sure which system was the real source of truth. legally wanted an approval step; IT wanted access controls; the business team wanted things to move faster. and the people who knew about the exceptions weren’t even in the room.
To me, this is the real story of business AI right now.
The headlines suggest that AI adoption is already a done deal. Accenture releases Microsoft 365 Copilot in approx 743,000 workers after a long internal pilot. General Mills has used AI to analyze thousands of daily shipments and reportedly saved over $20 million in transportation costs. Gartner, Inc. predicted that agent AI would appear in one third of enterprise software and services by 2028. On paper, AI is now mainstream in business.
Inside companies things are much more confusing. The demos work well and the pilots are real. Employees actually use these tools, and that matters. However, using AI alone is not the same as true transformation. A company can have thousands of people summarizing meetings, composing emails, and writing code with AI, but they still have the same broken workflows underneath.
This is where I think the discussion should go. Businesses no longer have a model problem. they have a readiness problem. Most leadership teams are still asking which AI tool they should buy. This is the wrong question. The better question is whether their company is really ready to use the AI tool.
Once AI goes beyond writing and summarizing and starts taking action, a company’s weaknesses quickly become clear. The policy exception is buried in an email. The customer specific rule is in someone’s mind. The customer relationship manager (CRM) says one thing, but the contract repository says another. A person in finance manages the spreadsheet that actually runs the business. AI fails because the company is hard to understand.
I’ve seen this most clearly in proposal and sales workflows. An AI system can read a 200-page RFP and extract the requirements in minutes. This is helpful, but the real work comes after that. What past projects have been approved for reuse? Which executive relationships matter? What claims have been legally rejected in the past? Which resumes are up to date? What winning themes worked in a similar offer, and which ones sounded good but never convinced the buyer?
This knowledge is usually scattered in old slide decks, folders, Slack threads, emails, and in the memories of people who have been around long enough to know all the details.
This is why so many AI projects quietly shrink after the pilot phase. They start out as big transformational efforts but end up as utility tools. They start out as autonomous agents, but evolve into human-in-the-loop control panels. The CEO may say the company needs to become AI first, but six months later, it has 20 disconnected pilots and no clear operating model.
Now, the focus is shifting to systems that can actually take action across workflows with guardrails, escalation paths and audit trails. This difference is important. Getting answers is easy now, but getting things done is still hard, and execution is a matter of company design.
The companies that do it well are changing the systems around the AI tool. They clean data, assign ownership to each workflow, incrementally grant rights to agents, and measure whether AI is actually improving metrics like cycle time, error rates, cost, resolution time, or revenue per employee.
This last point is important because the use of AI can quickly turn into a vanity metric. A company might proudly say that 60% of employees use AI every week, but what has actually changed? Did customers get answers faster? Did the engineering performance improve? Did finance close the books early? Were contracts legally audited with fewer errors? Did sales forecasts become more accurate? If the answer is not clear, the company may not have an AI strategy. It just has AI activity.
This is the risk I’m most concerned about—not that business AI is failing, but that it’s only half-succeeding. It saves time on the edges, creates enough excitement to get more budget, and gives leaders something to talk about in all-hands meetings, but it never gets to the core of how the business works. Three years later, the company has more tools, pilots and dashboards, but not much more real leverage.
Now, the boring work becomes the important work. Where does relevant data live? Who does it belong to? Which systems are reliable? What can an AI agent do without approval? When should it be escalated? What does the audit trail look like? Who is responsible when something goes wrong? These aren’t glamorous questions, but they separate AI theater from real AI transformation.
The next big divide in business will be between companies that treat AI as just a productivity tool and those that can transform their businesses in ways that will surprise the world. This is the first in a series of essays I plan to write about AI adoption in business — what works, what’s bad, and what leaders need to know as AI moves from experiments to core business.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Am I eligible?
