Your Business Isn't Failing at AI. It's Using AI Wrong.
The headlines aren't kind to AI right now.
A survey of almost 6,000 executives in the US, UK, Germany and Australia found that 89% saw no change in productivity from AI over the past three years (The Register, reporting the NBER study). Gartner reports that 97% of organisations can't cut costs through AI productivity measures alone, even though three-quarters of board members expect a return (ESG Dive). And MIT's widely quoted GenAI Divide report says that 95% of organisations are getting "zero return" on generative AI (One Year Ago in AI, on the MIT NANDA report; SuperSkills analysis).
It's easy to read all that and decide AI is overhyped. We think that's the wrong conclusion. The research doesn't say AI doesn't work. It says most businesses are using it wrongly.
The failure is in the organisation, not the model
Look past the headline numbers and the same finding keeps coming up.
- MIT puts the failure down to a "learning gap": most generative AI systems don't retain feedback, adapt to context or improve over time, and people abandon tools that don't fit their workflow (One Year Ago in AI; SuperSkills).
- The Productivity Institute found that 89% of organisations use AI somewhere, but only one in four has changed how work is done to get the most out of it. Fewer than half (46%) have got past pilot projects (The Productivity Institute).
- Harvard Business Review argues that most AI tools speed up work inside one function rather than across functions. The authors' survey found employees spend 37% of their time on coordination work: pulling, cleaning, reconciling and passing information between teams (HBR).
One statistic from the Productivity Institute report sums up the problem. Developers using AI coding tools wrote more than three times as much code, but released only 30% more software, because people still had to check, test and launch it (The Productivity Institute).
The AI made the task faster. Nothing around the task changed, so the business didn't speed up.
What "doing AI wrong" looks like
In most organisations we see one or more of these patterns:
- AI as a chat window. Staff get a licence for a chatbot and are told to "use AI". They draft a few emails faster. Nothing else changes.
- AI bolted onto an old process. One step gets automated, but the approvals, handoffs and spreadsheets on either side stay the same, so the time saved disappears into the queue.
- Pilots with no path to production. A proof of concept impresses a steering group and then sits there, because nobody owns it, budgets for it or measures it.
- No baseline. Nobody measured how long the work took before, so nobody can prove it got faster.
- Time saved, but not used. Gartner found only 21% of the time saved by AI was put into more valuable work (ESG Dive).
What "doing AI right" looks like: the Jensen Huang model
Nvidia's CEO, Jensen Huang, put the alternative very simply on the BG2 podcast:
"Suppose I were to hire a $100,000 employee, and I augmented that $100,000 employee with a $10,000 AI. And that $10,000 AI as a result made the $100,000 employee twice more productive, three times more productive. Would I do it? Heartbeat. I'm doing it across every single person in our company right now." (BG2 transcript, The Singju Post)
He has since gone further. At Nvidia's GTC conference he suggested engineers could get an annual token budget worth roughly half their base pay "so that they could be amplified 10X" (Business Insider).
The important change isn't the size of the budget. It's the mindset. Huang treats AI as capacity you allocate to a person, like headcount, not as software you install and hope people use. He also says Nvidia will one day have "hundreds of thousands of digital employees" working alongside its 42,000 people (CNBC).
Going one step further: humans manage the manager
We agree with Huang, and we'd go one step further.
Giving one person a token budget and a set of AI tools still leaves that person coordinating everything: prompting, copying outputs between tools, checking, chasing. They become the bottleneck, which is exactly the coordination problem HBR describes.
The bigger gains come from organising the AI the way you'd organise a team:
- A human sets the goals, the budget and the rules.
- A lead agent, effectively a head of department, turns those goals into a plan and hands out the work.
- Specialist agents do the work in their own areas and report back.
- The lead agent tracks progress, handles the handoffs and only brings the human in for decisions and approvals.
This isn't science fiction. xAI's Grok Bot, launched in August 2026, lets users run several bots in parallel, "including a chief-of-staff Bot overseeing specialists". The bots message each other, pass work along and pull the user in for judgement calls (xAI). Microsoft's Copilot Studio now supports multi-agent orchestration, with "a single front-door agent routing to the right specialist" (Microsoft Learn).
The human stops being the person who does the work and becomes the person who directs it. That's how one person can deliver the output of a whole team.
Why this fixes the productivity problem
The org-chart approach deals directly with the causes the research identifies:
| Why AI fails (research finding) | How a managed agent team addresses it |
|---|---|
| Gains stay inside individual tasks | The lead agent owns the whole workflow from start to finish |
| Coordination work eats the time saved | Agents hand work to each other, so people don't have to |
| Tools don't learn or keep context | Agents share context, memory and a common data layer |
| Pilots have no owner | Every agent has a named human owner and a KPI |
| Nobody measures the result | Agent output is measured like a team's output |
It only works on good foundations
Two warnings.
First, agents are only as good as the data and systems they can reach. An agent team built on messy, siloed data will make mistakes faster. Data quality, integration and governance come before agents, not after.
Second, a fleet of agents needs managing. Gartner predicts that by 2028 the average Fortune 500 company will have more than 150,000 agents in use, up from fewer than 15 in 2025, and only 13% of organisations think they have the right agent governance in place (Gartner). That's why we think businesses need a new function: Non-Human Resources.
Where to start
- Pick one department, not one task. Marketing is often the easiest place to start, as we explain in the one-human marketing department.
- Measure the baseline: how long the work takes, what it costs and what it produces today.
- Fix the data and integrations the agents will depend on.
- Design the agent org chart: one lead agent, clear specialist roles, human approval points.
- Give it a budget and measure it like a team.
If you're not sure whether your data and systems are ready, talk to us about AI Management and we'll help you answer that question.
Stop testing AI tools. Start building an AI team. Talk to Cyber Samurai.
Keep reading
