The average company is running 21 AI projects at once, according to a Rackspace Technology survey from mid-2025. MIT’s Project NANDA found, in the same year, that 95% of generative AI pilots produce zero measurable P&L impact. Put those two numbers together and the conclusion isn’t that AI doesn’t work — it’s that almost all of that activity is defense: cost and efficiency plays layered onto processes that already existed. Almost none of it is offense: new AI-native products or revenue lines that couldn’t have existed without it. The tricky part of offense usually isn’t the model, it’s wiring a new AI-native build into the complex, aging core systems most corporates already run on — a problem the current wave of AI startups mostly doesn’t have, since they’re building on a blank slate instead of retrofitting one.
That split matters more than the pilot count. A company can run 21 pilots forever and never once bet on something new.
1. Defense and Offense, Defined
Defense optimizes what a business already does: faster support tickets, automated documentation, cheaper back-office processing. It is low-risk, easy to fund, and easy to kill quietly if it underperforms.
Offense builds something the business could not sell before: a new product, a new revenue line, a capability that only exists because AI made it possible. It is higher-risk, harder to greenlight, and it requires committing real architecture and a real team — not a side project.
Most corporate AI budgets are almost entirely the first category. That is not an accident.
2. Why Everyone Defaults to Defense
Defense fits inside the governance most corporates already run: quantitative metrics, leaning towards a zero-mistakes culture, approval from committees built to protect the core business. A pilot that shaves 10% off a support queue is easy to approve and easy to justify if it doesn’t work out. A pilot that bets on an entirely new AI-native product asks for a different kind of approval altogether.
That’s part of why the pilot count keeps climbing while the outcomes don’t move. More defense pilots make the existing business marginally better, one at a time, but they don’t add up to offense on their own. Each one draws on the same stretched internal team, split across many priorities instead of fully dedicated to one.
3. What Focus Looks Like When It Works
Siemens is a company mid-transition. According to Peter Körte (Siemens CTO, Fortune, 2025), by July 2025 Siemens had 460 distinct AI use cases in production and over 15,000 bots built by its own employees since October 2024 — real scale, but scale that’s hard to govern or improve as one thing. In June 2026, Siemens launched Intelligence Center X to bring production data, models, and workflows onto one managed platform, reporting a 95% drop in manual effort and an 85% faster resolution of production issues on what it covers so far (Siemens press release, 2026). That’s focus paying off, even if it’s still working through only part of the sprawl it built up getting there.
Walmart shows the fuller arc. In 2023, the retailer ran a sprawl of specialized bots across product search, personnel, suppliers, and advertising. By 2025, it had consolidated to four focused AI agents serving customers, employees, developers, and partners (Walmart, July 2025). By 2026, its Sparky assistant reached 50% customer usage and, notably, drove 35% larger shopping baskets (Walmart Q4 earnings call, April 2026, via Yahoo Finance). That last number is the difference. Faster support tickets are defense. Bigger baskets are offense — new revenue that would not have existed without the bet.
4. Building an Actual Offense Case
The old way — tool-first
“We’re building an AI assistant” starts from the technology and works backward, hoping a use case appears. Klarna replaced 700 support agents with an OpenAI-powered chatbot, automated 67% of chats, and projected up to $40 million in savings (Klarna press release, 2024) — then watched customer satisfaction and service quality decline (Medium, 2026).
The GoTeams way — outcome-first
Define the user, the outcome, and the lever before writing the first line of code. What changes for the customer, what number moves because of it, and only then which AI approach gets you there. Skip that order, and even a technically impressive build ends up optimizing the wrong thing.
Why outcome-first isn’t enough on its own
Outcome-first thinking is what keeps a defense project from quietly optimizing the wrong thing. It’s not enough on its own for offense. Offense asks for something further: a team whose entire mandate is the new product, not a shared pool of internal engineers splitting attention across a 22nd pilot. GoTeams builds teams on exactly that principle — one team, one mission, fully dedicated to a single product, with no shared projects and no context-switching — because a genuinely new AI-native product does not get built by people who are also on call for the existing ones.
5. What Offense Actually Looks Like
Helloparts (with HUK-Coburg; automotive data platform; grew from 3 to 34 people over 4 years). Helloparts uses AI to predict the ideal replacement part for a given vehicle — a product that could not have existed at this accuracy before. The result: return rates fell from 20% to under 5%, workshops using the platform saw profitability rise by an average of 14%, and the platform now drives over €150 million in annual parts orders in its second year alone. Nothing about Helloparts optimizes an existing process. It created a new one, and a new revenue line with it.
6. Frequently Asked Questions
Why do most corporate AI pilots fail to show a measurable return?
MIT’s Project NANDA (July 2025) found that 95% of generative AI pilots produce zero measurable P&L impact. Three things separate the pilots that work from the ones that don’t: focus (a small number of high-impact bets instead of dozens running in parallel); outcome-first thinking (defining the user and the result before picking the AI approach); and a team or partner fully dedicated to the outcome, not a shared pool of engineers splitting attention across every other pilot. Most pilots have none of the three, which is why the pilot count keeps climbing while the returns don’t.
Does a pilot automatically lead to a real product?
A pilot exists to test whether something works before committing full resources to it — that’s true whether or not AI is involved. But running one doesn’t guarantee it leads anywhere. What decides that is what the pilot is testing, not how long it runs. An optimization pilot fine-tunes something the business already does — faster tickets, cheaper documentation — and it can run indefinitely without ever becoming more than that. A pilot built around a genuinely new capability or revenue line is testing something different: whether that capability holds up well enough to scale into a real product once it does.
How many AI pilots should a company run at once?
There’s no single right number — it depends on the size and structure of the company. But the instinct to run as many as possible is usually the wrong one. Rackspace Technology’s 2025 survey found companies run 21 AI projects simultaneously on average, and most of that is a wide spread of low-impact initiatives rather than a deliberate portfolio. A handful of clearly defined problems, each with a real outcome attached, beats a large number of pilots competing for the same stretched team, regardless of company size.
Should AI initiatives be outcome-first or tool-first?
Outcome-first. Define the user, the outcome, and the specific lever before selecting the AI approach. Skipping that step is what let Klarna’s chatbot replacement of 700 support agents optimize for cost while customer satisfaction and service quality declined. Defining the outcome first means quality is part of the target from day one, not something that shows up as a surprise afterward — which is the trade-off Klarna made when it optimized for cost alone.
The Bottom Line
If your only goal this year is to trim cost from an existing process, a well-run defense pilot is a reasonable, low-risk place to stay. But if your AI roadmap is entirely defense and you’re wondering why the return never shows up at the P&L level, the fix isn’t a 22nd pilot. It’s committing a dedicated team to one real outcome the way Walmart did with Sparky and Meinungswerk did with AI-driven research interviews. GoTeams builds AI-orchestrated teams fully dedicated to a single product — no shared pilots, no context-switching — so your next AI bet is built to be offense from day one. Activate your GoTeam today.
Leave a Reply