Research · MSc dissertation
How EdTech startups use AI agents to cut costs and scale
A survey of 45 startup professionals in 10 countries on where AI agents pay off, what gets in the way, and the vendor risk that comes with them.
Key numbers
- already use or are piloting AI agents
- 91%
- average rating for cost reduction
- 3.82/5
- adopted AI agents mainly to cut costs
- 73%
- rely on OpenAI and Anthropic
- 71% · 64%
Research question
How do startups implement AI agents to reduce operational costs and accelerate growth, and what stands in their way?
What I found
- 01
Adoption is close to universal
91.1% of respondents were actively using or piloting AI agents. Customer support (62%), administrative automation (56%) and content generation (53%) were the most common uses.
- 02
Savings come from language-heavy work
The biggest cost savings were in content creation, customer support and marketing: high-volume, repetitive tasks that today's large language models handle well.
- 03
Growth without proportional hiring
The most common experience was scaling without adding staff at the same rate. But one in five respondents found implementation more complex than they had expected.
- 04
Barriers come in clusters
Data privacy (47%), implementation costs (42%), data quality (38%) and the skills gap (31%) were reported at similar rates and tend to reinforce each other. Resistance from teams was rare (9%).
- 05
Vendor concentration is the hidden risk
Most startups build on OpenAI (71%) or Anthropic (64%); only 16% use custom models. A change in pricing or API terms can hit core operations directly.
How it shapes the way I build
- Start where language models are strongest: high-volume support, content and admin workflows, where results show quickly.
- Plan for the integration work, not just the subscription: data preparation, workflow redesign and ongoing maintenance.
- Stay provider-agnostic, so switching between OpenAI and Claude is a configuration change rather than a rebuild. The chat assistant on this site works with either.
- Treat privacy as architecture: GDPR and the EU AI Act shape how data flows through an agent, not just the paperwork around it.