The rollout probably went fine. Licenses were purchased, an executive sent the announcement, IT ran a few training sessions, and someone started a channel for sharing prompts. Usage numbers looked healthy in the first month.
Then a quarter or two went by and the CFO asked a simple question. What did we get for this?
If the room goes quiet when that comes up at your company, you have a lot of company.
You are in good company, which is not great news
McKinsey's 2025 global survey found that nearly two thirds of organizations have not started scaling AI across the enterprise. About 39 percent of respondents reported any EBIT impact from AI, and most of those said it was under 5 percent of EBIT. McKinsey's summary is that tools are everywhere but most companies haven't worked them deeply enough into how work gets done. An earlier McKinsey finding, reported by IEEE, put it more bluntly: nearly eight in ten companies use generative AI, and nearly the same share report no measurable bottom-line impact.
BCG looked at about 1,250 firms and found that only 5 percent are getting AI value at scale, while 60 percent report little or no material value despite real spending. IBM asked 2,000 CEOs and heard that only 25 percent of AI initiatives have delivered the expected ROI, and only 16 percent have scaled enterprise wide.
The number that got the most attention came from MIT's NANDA group, which said 95 percent of the organizations it studied saw no measurable return on generative AI pilots. That one deserves a caveat. It rests on a few hundred deployments and interviews from a preliminary research program, so it is better read as a direction than a universal failure rate. But the other studies, run with different methods and different samples, point the same way.
A license is not a change in how work gets done
Part of what makes this hard to see is that it feels like it's working. Engineers are writing more code than ever before. A salesperson who has never opened a code editor is building a small app to keep track of their accounts. Decks look better than they have in years. Walk the floor and everything looks busy and productive, and in a way it is.
But output and outcomes are different things. More code means more to review and maintain. An app built by one rep tends to live on that rep's laptop. A beautiful deck still has to persuade someone. Activity goes up, while the numbers the business actually runs on, like cycle time or win rate, may not move at all.
“When you hand someone a license, you give them a tool. A return shows up when a process changes. Those are different things, and most rollouts only deliver the first one.
MIT's report found that the tools mostly lift individual productivity rather than business performance. That is real, but it is hard to bank. If an analyst saves 20 minutes a day on email, nobody's budget line moves. The time gets absorbed into the day and the CFO never sees it.
Then there is the plain matter of whether people keep using the thing. Firms that audit Microsoft 365 environments report that a meaningful share of assigned Copilot seats see little or no use after the first 90 days. They sell license management, so read that with the appropriate grain of salt, but the cost math is worth doing for your own situation. At Copilot's $30 per user per month list price, 1,000 seats is $360,000 a year. Whichever vendor you picked, multiply it out and ask what you can point to.
At $30 per user per month, 1,000 seats is $360,000 a year. If a meaningful share go unused after 90 days, a big part of that is paying for access nobody opens.
It also helps to know how most companies measure return today. A Wharton survey found that nearly three quarters of companies say they track ROI, but the lead researcher noted that those reports rely on self-assessment more than hard evidence. And a Gallup poll from late 2024 found that only 15 percent of U.S. employees said their workplace had communicated a clear AI strategy. People can't pull in the same direction when nobody has said what the direction is.
The money often lands in the wrong place
One of the more useful findings in the MIT work is about where budgets go. More than half of generative AI budgets went to sales and marketing tools, while the strongest returns showed up in back-office work like cutting outsourcing and agency costs and streamlining operations (Fortune's write-up).
More than half of generative AI budgets went to sales and marketing tools - the visible, exciting use cases.
The strongest returns showed up in back-office work: cutting outsourcing and agency costs and streamlining operations.
Nobody decided that on purpose. It is what happens when tools get picked before problems do. Visible, exciting use cases attract funding, and the dull, valuable ones wait.
BCG sees the mirror image in the companies that do well. Those firms concentrate on core business processes rather than experimenting at the edges, and they avoid the scattershot approach.
An AI strategy starts with discovery
Strategy gets used as a grand word, but at the start it is mostly a set of unglamorous questions. Where are the real problems? Who has them, and what does fixing them be worth? What data exists, and in what shape? What are the security and compliance limits? Then comes the part many teams skip, which is scoring every idea on two things: the value if it works and how hard it is to build and get adopted.
Do this with the people who own the processes, not just the technology team. A ranking built by one group tends to get quietly ignored by everyone else.
A short list in a defensible order, plus an honest answer to whether your data can support the first few items. Without it, you end up funding whatever the loudest person in the room pitched last.
Split the portfolio: some quick wins, mostly bigger bets
Once the opportunities are scored, we suggest a rough split of 40 percent of effort on quick wins and 60 percent on larger efforts. Treat that as a starting point, not a law of nature. The right mix depends on your data, your appetite for risk, and how much patience your finance team has.
A starting point, not a law of nature - the right mix depends on your data, risk appetite, and your finance team's patience.
Quick wins are the high value, low difficulty items, where the data is in decent shape and the path to production is short. They matter for a few reasons. They show results in weeks, they earn credibility with finance, and their savings can help pay for what comes next.
The larger efforts are where enterprise level change lives. McKinsey notes that meaningful bottom-line impact from AI is still rare, but its results suggest that thinking big can pay off. A portfolio made only of small wins tends to produce a lot of activity that never adds up to much.
The reverse has its own problem. Gartner expects that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. Big bets with no early proof points are exactly the ones that get cut. Mixing the two keeps the program alive long enough to finish the hard parts.
If you are in this spot, the AI Primer is built for it
This is the situation we designed the AI Primer for. It is a three week engagement for organizations that know AI matters but need to decide where to start, or that have a stalled backlog of ideas nobody can agree how to rank.
Your goals and KPIs, the problems that matter, your data and cloud platform, and your security constraints.
A one day workshop with your stakeholders, where every candidate is scored together on business value and implementation difficulty.
A prioritized roadmap, a data readiness assessment, and build-ready briefs for the top three, each a thin slice to production in 6 to 12 weeks.
Every candidate gets an effort and ROI estimate attached. The deliverables are yours whether or not you work with us afterward. You can also find the details on the AWS Marketplace listing.
Your ChatGPT or Claude licenses don't go to waste in this picture. They become one tool among several, pointed at the places where a map says they will matter.
Write down three workflows where AI changed something you can measure. If that is easy, good. If it isn't, you have learned something useful, and you know where to start.
Turn a stalled rollout into a roadmap
The AI Primer is a three week engagement: discovery, a stakeholder scoring workshop, and a prioritized roadmap with build-ready briefs for your top three opportunities. The deliverables are yours whether or not you continue with us.
References
- McKinsey & Company. "The state of AI in 2025: Agents, innovation, and transformation." mckinsey.com
- IEEE Transmitter. "Generative AI 2026: Companies Looking for Business Value." transmitter.ieee.org
- Boston Consulting Group. "The Widening AI Value Gap," October 2025. media-publications.bcg.com
- Boston Consulting Group. "Strategies to Tackle the AI Skills Gap," 2025. bcg.com
- IBM Newsroom. "IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles," May 6, 2025. newsroom.ibm.com
- Fortune. "MIT report: 95% of generative AI pilots at companies are failing," August 18, 2025. fortune.com
- Let's Data Science. "MIT NANDA's 2025 Report Says 95% of Organizations Saw No GenAI Return." letsdatascience.com
- Think Digital. Summary of "The GenAI Divide: State of AI in Business 2025" (MIT NANDA), July 2025. thinkdigital.travel
- Rencore. "The hidden cost of unused Copilot licenses: what we keep finding." hub.rencore.com
- AOL Finance. "Why generative AI went from risk to business imperative at U.S. companies" (Wharton research). aol.com
- Fortune via Yahoo Finance. "CEOs say that just a fraction of AI initiatives are actually delivering the return on investment they expected" (cites Gallup, late 2024). finance.yahoo.com
- South China Morning Post. "Over 40% of agentic AI projects forecast to be scrapped by 2027 due to lack of value" (Gartner). scmp.com
- AWS Marketplace. "StronglyAI AI Primer: AI Strategy and Roadmap Engagement." aws.amazon.com