Halfway through our conversation about AI in schools, Stephen Hodges turns to his own developers.
Suppose he told them to start using a new AI coding tool. Would an instruction from the CEO be enough? In his account, they would still need training and incentives. Giving people access would not automatically change how they worked.
It is a useful moment in a conversation with someone trying to bring AI teaching into education systems. The adoption problem is close enough to find inside his own company.
Hodges, CEO of Efekta Education, wants to make personal tutoring available at a price comparable to a textbook. That ambition depends on software being able to take on more of the work of instruction. But listen to how he describes implementation: district ambassadors, teacher training, support, follow-up and a division of responsibilities with ministries.
The question I took away for software teams is simple: how much of a successful pilot is being delivered by people whose work never appears in the product metrics?
Count the people around the product
In Hodges’s account, Efekta trains ambassadors in school districts. Those people train teachers locally. The company supports them and follows up where adoption is falling behind.
That arrangement tells us something important about what is being deployed. A school receives technology within a system of training and support. The people around it help turn access into use.
For a team evaluating its own AI pilot, I would start by writing down everything those people do.
Who gets a reluctant user through the first session? Who notices that a team has stopped using the tool? Who answers the question the interface does not explain? Who fixes an exception before anyone logs it as a problem?
These are diagnostic questions, rather than claims about Efekta’s implementation. The interview gives us the reason to ask them: Hodges describes local support as part of the rollout itself.
A pilot can establish that a particular combination of software and support works. To expand it responsibly, we need to know which parts of that combination must expand too.
If one helpful person is quietly carrying the process, adding customers adds work for that person. Some of that work may eventually become a product feature. Some may need to remain a service. Either way, it belongs in the estimate.
An escalation needs somewhere to go
Hodges also describes a learner who remains stuck after the system has tried to teach a concept. At that point, the system alerts the teacher.
For an engineering team, that is a more useful starting point than a broad commitment to human oversight. There is a condition, an action and a person expected to respond.
The implementation questions follow quickly. What should the teacher see? How much of the student’s recent work matters? Can the teacher understand the problem without replaying the entire interaction? What happens while the student is waiting?
Those details are my design questions, not a description of features demonstrated in the interview. They are the work required to turn the proposed handover into something a teacher can use.
Hodges makes a similar distinction around suspected cheating. The software can flag unusual behaviour, but the teacher has context the system lacks and decides how to respond.
For other AI products, the relevant person may be an account manager, a support specialist or an engineer. The same design question remains: when the software reaches its limit, does somebody receive enough information to take the next step?
An alert with no clear recipient is unfinished work. So is a queue that nobody has time to review.
Localisation changes the estimate
One of Hodges’s examples involves an Egyptian student working through material about going to a mall and buying a burger. He uses it to explain why learning content needs cultural adaptation.
The learning objective might travel well. The example used to teach it might not.
For a software team, that is a useful distinction to make before promising a launch in another country. Some parts of the product can be reused. Others need someone who understands the setting well enough to judge whether they make sense.
Hodges also discusses curricula, existing systems and the information teachers and ministries require. Those requirements extend the work beyond translated screens.
I would want a rollout estimate to show that work explicitly. Who reviews the material? Who owns the integration? Who agrees on the reports? Which decisions can the supplier make, and which require the customer?
Until those questions have owners, a launch date rests on assumptions about other people’s availability.
Usage leaves the hardest question open
The measurement discussion adds another complication.
Hodges says his team examines the relationship between use of the system and learning results. He also acknowledges how difficult it is to separate the software’s contribution from teacher quality or student motivation.
A motivated student may practise more and achieve better results. A well-supported teacher may encourage more use and improve the classroom experience in other ways. Seeing usage and outcomes rise together does not tell us how much the software caused.
This matters for product teams because usage is relatively easy to report. The outcome a customer cares about can be much harder to establish.
Hodges leaves room for more than one useful outcome: reaching a higher level, or reaching the required level in less time. Those would imply different evaluation plans.
Before expanding a pilot, I would want agreement on the outcome being pursued and the evidence that would change our minds. More sessions may be useful evidence of adoption. They still leave the question of benefit to be answered.
Give the next customer an honest operating plan
The strongest lesson I took from this conversation is about estimating the whole job.
Efekta’s ambition concerns the cost and availability of individual instruction. Hodges’s account of putting it into schools includes people who train, intervene, adapt materials and coordinate responsibilities. Both belong in the same discussion.
For a founder or product lead, that suggests a practical exercise before the next rollout: sit with the people who supported the pilot and reconstruct an ordinary week. Include the calls, reminders, exceptions and explanations. Decide which work the product will absorb, which the supplier will provide and which the customer has agreed to own.
That is the operating plan the next customer needs.
What is the one task a person still has to do for your AI pilot to succeed?
I’d be interested to hear the answer. It may tell us more about the next stage of the product than another demo.
Watch the full EdTech Dots conversation: Stephen Hodges on what schools still need.
Krzysztof Kosman, 1000software. This essay draws on my conversation with Stephen Hodges. Descriptions of Efekta’s practices reflect his account; the product-design questions and conclusions are my interpretation.

