Most companies calling themselves "AI-first" have no idea what they mean by it.
This isn't a cynical take. It's what the data shows. According to BCG's 2024 research on AI adoption, 74% of companies say they're struggling to achieve and scale value from AI — despite two years of treating it as a top strategic priority. Only 4% describe themselves as creating substantial value from their AI investments.
Think about that gap. A company can spend $50 million on AI, send its executives to every AI conference, publish an AI strategy deck, and still fall into the 74%. The declaration of AI seriousness and the actual results are almost entirely disconnected.
That disconnect is what "AI-first" produces when used as a strategy. It produces declaration, not outcomes.
What the label is supposed to signal
"AI-first" sounds like strategy. It's supposed to mean a company has reorganized its operations around AI. That decisions are informed by AI models. That products are built with AI from the ground up rather than bolted on afterward. That AI is central to how the company competes.
That's what it's supposed to mean. In practice, it means something much narrower: we're serious about AI.
It's the strategy slide equivalent of a firm handshake. The term carries conviction. It lacks content.
I've noticed a pattern in how companies use it. They announce the label in Q1. They buy enterprise AI licenses in Q2. They publish an internal use policy in Q3. By Q4, they're reporting that employees are "adopting AI at scale" — a claim backed by login data, not by any change in how the business performs.
This is "digital transformation" all over again. A decade ago, every company was on a digital transformation journey. Most of them moved some spreadsheets to the cloud. The word transformation did a lot of rhetorical work. The actual transformation happened much less.
"AI-first" is doing the same job today.
The problem with strategy labels
Here's the core issue. A strategy describes a specific set of choices: where to compete, how to win, what you will and won't do. A real strategy creates trade-offs. It tells you which problems you're solving and which you're ignoring.
"AI-first" creates no trade-offs. It doesn't tell you which processes are changing. It doesn't tell you which decisions are different because of AI. It doesn't define success or tell you how you'd know if you'd achieved it.
An AI-first strategy, in the useful sense of the word, would specify: which processes are being redesigned around AI, what outcomes we expect, and how we'll know if we've succeeded. The phrase as companies actually use it specifies none of this.
This matters because vague strategy produces vague effort. When your AI strategy is "be AI-first," your team's job becomes demonstrating AI activity. They find places to deploy AI tools. They hit adoption metrics. They produce pilots. The pilots succeed in controlled environments. Then the pilots fail to scale. Then they start different pilots.
Forrester's research captures this pattern precisely. More than 60% of AI pilots fail to scale beyond controlled environments. Only 10-15% of pilots successfully expand into production operations. Companies aren't failing because they lack AI tools. They're running pilots that look like progress but don't require them to make the hard choices.
The evidence for this failure pattern is consistent
The data here is worth sitting with, because it comes from multiple independent sources and points in the same direction.
MIT tracked over 300 enterprise AI deployments and found that 95% failed to deliver ROI. According to analysis of the MIT findings published in Fortune in 2025, the failure is primarily organizational, not algorithmic. Companies are choosing the path of least friction — deploying AI where it's easiest to deploy, not where it would create the most value. They're avoiding the workflow redesign that actual value creation requires.
Gartner's Hype Cycle for Artificial Intelligence, 2024, found that despite record-level investment, generative AI implementations are "still in early stages" and "few have achieved business value." The gap between investment and value isn't narrowing; it's a structural feature of how most companies are approaching the problem.
RAND Corporation's 2024 research put the AI project failure rate at over 80%. That's twice the failure rate of ordinary IT projects. If anything, AI implementation is harder to get right than the software projects that companies were already struggling with. Calling yourself "AI-first" doesn't change this math.
The pattern is too consistent across too many independent researchers to be an outlier problem. This is a how-companies-approach-AI problem.
Why the approach fails
Most companies make the same mistake. They start with a technology budget. They decide how much to spend on AI, buy the tools, and then look for problems to apply them to.
This feels right. It's how companies buy software. Budget first, implementation second.
But it's wrong for AI, for a specific reason.
Software adds value by automating a defined task. AI adds value by changing how a process works. These are different kinds of projects. Changing how a process works requires understanding why the process is the way it is, redesigning it around new capabilities, and then managing the transition as people learn to operate differently. None of that is captured in a technology budget.
Think about what "starting with the budget" actually produces. You have money for AI tools. Your team finds use cases. The use cases are evaluated on whether AI can do them, not on whether doing them would materially improve the business. The easiest-to-demo use cases win: document summarization, meeting transcription, internal chatbots. These are real AI applications. They are almost never the applications that would move a business metric.
MIT's finding — that companies avoid friction — is the same observation from a different angle. Friction, in this context, means workflow redesign. It means asking employees to change how they work, not just add a new tool to how they already work. It means redefining processes, not augmenting them. Companies avoid this because it's harder, slower, and less impressive in a quarterly update.
The result is exactly what the data shows: widespread adoption, minimal impact.
What separates the 4%
BCG's research on AI implementations that actually create substantial value found a consistent pattern across the companies that succeed. Three things separate them from the 96%.
First: they start with a specific broken process, not a technology budget.
The question isn't "how should we use AI?" It's "what is broken, and would AI fix it?"
These questions look similar. They produce completely different projects.
Starting from a broken process forces specificity. You have to name the process. You have to describe exactly what's wrong with it. You have to define what "fixed" looks like — and that definition gives you something to measure. It also forces an honest answer to whether AI is actually the right tool, versus a process redesign, a data quality fix, or a hiring decision.
Starting from a technology budget produces pilots looking for problems. The pilot succeeds when the AI demonstrates capability, not when the underlying process improves. The success criteria are "AI worked in the demo" rather than "the problem is smaller." When you measure AI activity instead of process outcomes, you can have a fully successful AI program that makes no difference to how the business runs.
Second: they measure outcomes, not tool adoption.
Adoption metrics are appealing. They're easy to collect. They provide visible evidence of progress. X% of employees have activated their AI assistant. Y thousand prompts submitted this week. Z hours of AI-assisted work completed.
None of these tell you whether the business is better.
The companies getting real value from AI measure downstream: error rates in a specific workflow, time-to-close on a specific process, customer resolution rates in a specific support queue. These metrics are harder to attribute cleanly to any single tool. They require understanding the process well enough to know what a good outcome looks like.
This is exactly why most companies avoid them. Adoption metrics tell a clean story. Outcome metrics tell a messier, more honest one.
Consider what a real outcome metric requires. You need a baseline. You need a process that was slow or error-prone in a specific, documented way. You need a measurement period long enough to see genuine change. You need to be willing to find out that the AI made no difference — or made things worse. That's not comfortable. But it's the only kind of measurement that produces information you can act on.
If you can't name a specific process that measurably improved because of your AI investment, you have activity, not results. The AI automation guide framing is useful here: automation creates value when it reduces a measurable problem, not when it generates proof of use.
Third: they give people time to learn, instead of targets to hit.
This one runs counter to how most organizations operate.
The instinct is to set adoption targets: every department must have X% of employees actively using AI tools by Q3. This makes the initiative legible. It creates accountability. It produces a clean number to report.
It also backfires. When people have adoption targets, they hit the adoption targets by doing the minimum that qualifies as adoption. They log in. They submit prompts. They don't actually change how they work. The metric climbs. The process doesn't change.
BCG's 10-20-70 framework describes where successful AI implementations actually put their effort: roughly 10% on the technology, 20% on processes and data, and 70% on people and change management. This surprises most executives. It shouldn't.
AI changes how people work. That's the entire point. Changing how people work takes time, practice, and enough psychological safety to try things that might fail. You can't mandate that with a quarterly target. The companies succeeding at this are the ones treating AI change management as a first-class problem, not an afterthought to a technology rollout.
What it looks like when it actually works
The companies getting real value from AI don't spend much time talking about being "AI-first." They describe what they fixed.
A logistics team that reduced manual data-entry errors by 60% in invoice processing — a specific process, a specific number, a specific before-and-after. A financial services firm that cut analyst prep time for a defined class of client presentations from 8 hours to 90 minutes. A retail planning team that stopped spending half the week pulling inventory reports because those reports now surface automatically.
In each case: one process, one outcome, one metric. No declarations about AI identity.
Notice what these companies did not do. They didn't start by asking "how do we become AI-first?" They asked: "what's the thing that's slowest and most painful right now, and is AI the right tool to fix it?"
The question sounds small. The ambition looks limited compared to a company that's declared AI transformation of its entire operations. But the specificity is exactly what makes them succeed where companies with much larger AI ambitions don't.
According to HBR research on companies succeeding with AI, published in 2025, a consistent finding is that successful organizations "start narrow and go deep" rather than broad and shallow. One process fully transformed creates more value — and more organizational learning — than ten processes slightly augmented.
The deeper problem with "AI-first"
There's a logical confusion at the center of the "AI-first" label.
The companies calling themselves AI-first have made AI the priority. They're investing in AI, measuring AI adoption, building AI strategy. The companies actually winning with AI have made their processes the priority. They're investing in specific improvements, measuring specific outcomes, and using AI where it helps.
These are different orientations. They produce different decisions.
When AI is the priority, you buy AI tools and find them things to do. When your processes are the priority, you identify what's broken and find the best tool to fix it. Sometimes that's AI. Sometimes it's a better process design. Sometimes it's both.
I think this is the deepest problem with the "AI-first" label. It reverses the causality of what successful companies are doing. Successful companies don't win because they prioritized AI. They prioritize AI for specific things because it helps them win at those things. The tool serves the goal. The label makes the tool the goal.
When the tool becomes the goal, you end up optimizing for AI presence rather than business improvement. Your metrics measure AI activity. Your investments flow toward visible adoption. Your energy goes toward demonstrating that AI is central to operations rather than toward making operations actually better.
This is how you end up in the 74%.
What to actually do
The useful questions look different from the standard AI strategy questions.
Not "how do we become AI-first?" but: which three processes in your operations are most broken? Not "could be enhanced by AI" — broken. Slow, error-prone, inconsistent, requiring constant manual intervention.
For each one: what would "fixed" look like? What specific metric would tell you the process improved?
For each one: is AI actually the right tool, or is the underlying problem a data quality issue, a process design issue, or a staffing issue?
If AI does help: what does the person doing this job need to know? How much time do they need to actually change how they work? What support do they need to do that without a quarterly adoption target looming over them?
A company that can answer those questions for three processes has a real AI strategy. It doesn't need the label. And it probably won't use it, because the label is for people who haven't yet done the work of asking those questions.
There's one more thing worth naming. The companies that succeed tend to resist the pressure to show AI everywhere at once. They resist the temptation to announce AI transformation of their entire operations before they've transformed anything. They start with something small enough to measure and important enough to matter. Then they do it again. The organizational knowledge compounds. The second process they fix is easier than the first, because they've learned how to do it. By the third, they actually know what they're doing.
This is the opposite of "AI-first." It's "specific problem first, and AI where it helps."
The 4% of companies creating substantial value from AI are doing something that's available to every company. They're not smarter. They don't have better AI tools. They don't have bigger budgets.
They started with a different question. Instead of "how do we become AI-first?", they asked "what specifically needs to improve, and can AI help?"
The question is smaller. The results aren't.
According to BCG's 2024 AI adoption research, the companies at the front of AI value creation consistently apply a 10-20-70 principle: roughly 10% of implementation effort on the technology, 20% on data and process design, and 70% on the human side — training, change management, and giving people time to learn. Most organizations invert this ratio and spend the most on the technology that matters least.
Originally published on Superdots.
Top comments (0)