Across the country, schools and districts are buying ambitious, expensive AI faster than they are deciding what to do with it. The pressure is real, and feeling it is not a weakness. Boards ask what you are doing about AI, families ask, the vendor in the lobby has a persuasive demo, and standing still can look like falling behind. So the tool arrives with a launch event and a bold promise, and a few weeks later a quieter question shows up in the building: we own this now, so how do we make it worth the money?
That question is the one almost no one writes about, and it is the only one that helps a principal, executive director, or superintendent once the purchase order is signed. This is a playbook for answering it, whatever you bought.
The clearest current illustration is a San Diego charter network that spent a reported $500,000 on two humanoid robots. I want to be careful about how I use them here. Being early is not a crime; the network serves a genuinely hard student population, and its catch-up model has a real track record of moving students the system had already lost once. I point at their purchase throughout because the specifics are public and instructive, not because they are uniquely at fault. The pattern is everywhere, and any of us could have made a version of the same call under the same pressure. They are simply the sharpest current picture of it.
Here is the honest news for whoever is holding a receipt like that: success is still reachable. The catch is that you will not get there by chasing the win the marketing promised. You get there by redefining winning to something the tool can actually deliver.
Change the scoreboard before anything else
There is a trap here that sinks more of these efforts than anything else, so it is worth naming first.
If the scoreboard is some version of "the robot teaches better than a screen," the effort loses, and not because a team executed poorly. It loses because that is not something the technology can do this year. The strongest recent evidence in education points the other way from the hardware: well-designed AI tutoring can produce real learning gains. But that evidence is about the software, the model that asks good questions and adapts to a wrong answer, delivered on ordinary devices. There is no comparable evidence that a humanoid body wrapped around that software improves a single outcome, and the demonstrations so far suggest the body mostly adds latency.
So the first move is subtraction. Take "the machine improved test scores" off the table as the headline metric. It is the goal you are least equipped to hit and most likely to be judged on. Replace it with something the tool can plausibly move in a single semester, and make it something other than academic achievement.
The strongest candidate has been reached. Does the tool pull in chronically disengaged, credit-recovery students who were not showing up for human help before? If yes, and you measured it against a comparison group so it is a finding and not a story, that is a genuine, defensible win. It also quietly answers the sharpest criticism these purchases attract, the "you should have hired tutors" line, because it shows the tool doing something additive rather than replacing a human who was already reaching those kids.
Note: This article was researched and written by Justice Jones with AI assistance, then reviewed and edited by our team. External studies and sources are credited to their original authors. Examples from our own work reflect our organizational practice.
Repoint it at the jobs it can actually win
A tool like this usually ships configured for several roles at once. The San Diego robots were set up with four roles: a teacher, a college and career planner, a translator, and a wellness coach. Treating those as equals and staking the program on the hardest of them is the common mistake. They are not equally winnable, and the job now is triage.
Translation for families is the strongest bet, and it is not close. Real-time, multilingual communication with families is a real and underserved need for exactly this kind of student population, and it is a task language models handle reliably. It also does something a hallway of tutors was never going to do at scale, which reframes the tool as additive rather than substitutionary. Lead with it.
College and career planning is second because it produces a measurable outcome. A plan either exists or it does not, so you can show progress without commissioning a research study, as long as a counselor reviews what the machine drafts.
Direct academic teaching is the crowded lane with the weakest evidence for the hardware, so demote it from the headline. Let the tool answer content questions if a student wants that, but do not build the success story on it.
None of this is glamorous, and all of it means publicly doing less than the launch implied. That is exactly the trade worth making. A quiet, honest downgrade to two jobs done genuinely well beats a loud commitment to one job the tool cannot do.
<!-- 24/7 Teach interactive element: the repoint planner (post-purchase salvage) -->
<!-- Standalone. No external dependencies. No localStorage. Drop into post body. -->
<div id="repoint" style="max-width:680px;margin:2rem auto;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,Helvetica,Arial,sans-serif;color:#30455c;background:#e6ecf0;border-radius:16px;padding:28px 28px 22px;box-sizing:border-box;">
<div style="font-size:0.78rem;letter-spacing:0.08em;text-transform:uppercase;color:#63697a;font-weight:700;margin-bottom:6px;">You already bought it</div>
<h3 style="margin:0 0 4px;font-size:1.5rem;line-height:1.2;font-weight:800;color:#30455c;">Pick the job your AI can actually win</h3>
<p style="margin:0 0 20px;font-size:0.98rem;line-height:1.5;color:#63697a;">One question. Your answer points the tool at a job it can plausibly do well and hands you a metric you can defend in a board meeting. This is not a substitute for professional judgment. It is a starting point.</p>
<div style="font-size:0.96rem;font-weight:600;margin-bottom:12px;color:#30455c;">What is your most acute, measurable gap right now?</div>
<div id="rp-options"></div>
<div id="rp-result" style="display:none;margin-top:20px;">
<div id="rp-plan" style="padding:22px;border-radius:12px;background:#fff;border-left:6px solid #07c4c8;"></div>
<div style="margin-top:14px;padding:18px 20px;border-radius:12px;background:#30455c;color:#e6ecf0;">
<div style="font-size:0.82rem;letter-spacing:0.06em;text-transform:uppercase;font-weight:700;color:#f59070;margin-bottom:8px;">Two non-negotiables first, whatever you picked</div>
<div style="font-size:0.95rem;line-height:1.55;">
<strong>1. Gate the wellness-coach role.</strong> Do not let an AI persona act as emotional support for at-risk teens until you have a written protocol that hands off to a human on any sign of distress. The evidence on adolescent AI attachment is still thin, which is a reason for caution, not comfort.<br><br>
<strong>2. Get the data claim on paper.</strong> "It does not record and memory is wiped" needs to be documented, independently verified, and shared with families, not stated verbally. For minors on public money, an unverified privacy claim is a liability, not a safeguard.
</div>
</div>
<button id="rp-reset" style="margin-top:16px;background:none;border:1px solid #a7b1c2;color:#63697a;border-radius:8px;padding:9px 16px;font-size:0.88rem;font-weight:600;cursor:pointer;">Start over</button>
</div>
</div>
<script>
(function(){
var OPTS = [
{
label: "Families struggle to communicate with the school",
lead: "Lead with the translator role",
why: "This is the strongest bet you have. Real-time, multilingual communication with families is a genuine unmet need for this population, and it is a task large language models handle reliably. It also answers the loudest criticism directly, because it does something a hallway of tutors was never going to do at scale.",
metric: "Families reached in their home language, documents translated, and family-meeting attendance rate, measured against last year.",
caveat: "Have a human verify any high-stakes translation (legal, IEP, disciplinary) before it goes out. AI drafts, a person signs off."
},
{
label: "Students lack college and career plans",
lead: "Lead with the college-and-career-planner role",
why: "Second-strongest bet, because it produces a countable output. A plan either exists or it does not, so success is legible without a research study.",
metric: "Completed, counselor-reviewed plans per student, and downstream steps like FAFSA completions or applications submitted.",
caveat: "A counselor reviews every plan. The AI drafts and organizes, a human owns the advice a student actually acts on."
},
{
label: "Disengaged students are not showing up for help",
lead: "Use the robot as a draw, and measure the counterfactual on purpose",
why: "If a talking machine pulls a chronically disengaged, credit-recovery student through the door on a day nothing else would, that is a real, defensible win. It proves the tool is additive to human support, not a substitute for it.",
metric: "Resource-center visits by previously-disengaged students, compared against a matched group that did not get the robot.",
caveat: "Novelty spikes then fades. It only counts if the bump holds past roughly six weeks. Track sustained visits, not opening-week traffic."
},
{
label: "Students are behind academically and need remediation",
lead: "This is the robot's weakest lane. Do not lead here.",
why: "The rigorous evidence for AI helping learning is about the tutoring software, which runs on the Chromebooks you already own. The humanoid body adds latency and cost, not learning. Put remediation on the screen and reserve the robot for the jobs above.",
metric: "If you must measure the robot here, use mastery on one specific skill against a control group, but expect the screen, not the robot, to do the real work.",
caveat: "Staking the program's success on 'the robot teaches better than a screen' is the one bet you are most likely to lose. Avoid it."
}
];
var wrap = document.getElementById('rp-options');
OPTS.forEach(function(o, i){
var b = document.createElement('button');
b.setAttribute('data-i', i);
b.className = 'rp-opt';
b.style.cssText = 'display:block;width:100%;text-align:left;background:#fff;border:1.5px solid #a7b1c2;border-radius:10px;padding:14px 16px;margin-bottom:10px;font-size:0.96rem;line-height:1.4;color:#30455c;cursor:pointer;font-weight:600;';
b.textContent = o.label;
wrap.appendChild(b);
});
wrap.addEventListener('click', function(e){
var b = e.target.closest('.rp-opt');
if(!b) return;
var o = OPTS[parseInt(b.getAttribute('data-i'))];
Array.prototype.forEach.call(wrap.children, function(x){
x.style.background = '#fff'; x.style.borderColor = '#a7b1c2'; x.style.color = '#30455c';
});
b.style.background = '#07c4c8'; b.style.borderColor = '#07c4c8'; b.style.color = '#fff';
document.getElementById('rp-plan').innerHTML =
'<div style="font-size:1.2rem;font-weight:800;color:#30455c;margin-bottom:10px;">'+o.lead+'</div>'+
'<div style="font-size:0.96rem;line-height:1.55;color:#63697a;margin-bottom:14px;">'+o.why+'</div>'+
'<div style="font-size:0.82rem;letter-spacing:0.05em;text-transform:uppercase;font-weight:700;color:#07c4c8;margin-bottom:4px;">The metric to defend</div>'+
'<div style="font-size:0.96rem;line-height:1.5;color:#30455c;margin-bottom:12px;">'+o.metric+'</div>'+
'<div style="font-size:0.82rem;letter-spacing:0.05em;text-transform:uppercase;font-weight:700;color:#f59070;margin-bottom:4px;">The trap to avoid</div>'+
'<div style="font-size:0.96rem;line-height:1.5;color:#30455c;">'+o.caveat+'</div>';
document.getElementById('rp-result').style.display = 'block';
document.getElementById('rp-result').scrollIntoView({behavior:'smooth', block:'nearest'});
});
document.getElementById('rp-reset').addEventListener('click', function(){
document.getElementById('rp-result').style.display = 'none';
Array.prototype.forEach.call(wrap.children, function(x){
x.style.background = '#fff'; x.style.borderColor = '#a7b1c2'; x.style.color = '#30455c';
});
document.getElementById('repoint').scrollIntoView({behavior:'smooth', block:'start'});
});
})();
</script>
Handle the liability before you chase the upside
The wellness-coach role is the piece that can end a program on its own, and it deserves attention before anything else on this list, because it puts a school's credibility and a child's safety on the line, not just a metric.
An AI persona offering encouragement and coping strategies to teenagers, many from high-stress environments, is the highest-liability element in a purchase like this. I want to be careful rather than alarmist. The research on adolescents forming unhealthy attachments to responsive AI is still thin, and I do not have a verified study to cite for it, so treat this as a plausible risk rather than a settled finding. But thin evidence points toward caution, not comfort, when the users are vulnerable minors, and the downside is a child in distress getting a machine instead of a person. One bad interaction, with no offsetting results anywhere else, can define the whole story.
So either retire that role, or gate it hard: a written protocol that hands off to a trained human on any sign of distress, an explicit rule that the persona is never positioned as emotional or mental-health support, and a counselor in the loop by design. Safety here is not a feature of success. It is the precondition for being allowed to pursue it.
The same "get it in writing" discipline applies to data. Officials in the San Diego case have said the robots do not record and that memory is wiped after each interaction. Good, if true and documented. But for minors, on public money, a verbal assurance is not a safeguard. Get any such claim documented, independently verified, and shared with families in plain language.
Make the research real, or stop calling it research
When a purchase like this is announced as groundbreaking research, that framing invites scrutiny, it often cannot survive. In the San Diego case, the program was described as being the first of its kind to research physical AI as a teaching partner, and as of the reporting, no one has been able to identify a research partner, a study design, or a measurement plan. That gap draws the harshest criticism, and it is almost always self-inflicted.
There are two honest paths, and it is worth picking one on purpose. Either make it real research, which means partnering with an actual university school of education, securing institutional review, defining a comparison condition, and committing to a question before you know the answer. Or drop the word "research" and call it a pilot. Both are respectable. What keeps drawing fire is the claim of research without the apparatus of research. Close that gap or rename it, and do not paper over it by naming a study partner that does not exist.
Fix what is visibly broken, and publish either way
Two smaller things, both fixable, both worth catching early.
First, interaction quality. In the San Diego case, a reporter watched one robot deliver a lesson as a historical figure in a session full of stops and starts, speaking too fast, repeating its introduction several times while students scrambled to keep up, and the school's own dean reportedly called it clunky. Most of that is tuning, speech rate, and turn-taking, and it is fixable in configuration. But if the flagship demonstration is awkward, that awkwardness becomes the public story regardless of what else is true. Get the experience tight before the next observer walks in.
Second, commit early to publishing your results, whichever way they fall. If the reach metric moves, you are vindicated publicly, with data. If it does not, you become the rare school honest enough to say so, which is its own credibility and buys enormous trust for the next decision. The only outcome that reads badly is silence, because silence looks like hiding.
What this means for any leader who has already made the bet
Strip away the robot, and this is a general playbook because the underlying problem is general. The purchase gets made before the work that tells you what to measure, what to prioritize, and what to guard against. You cannot un-buy the tool, but you can do that work now, after the fact, and let it steer what comes next.
The sequence, even retroactively, is the same. Define the one problem you are actually solving, in a sentence, with a number. Point the tool at the narrow jobs it can win rather than the broad one it was sold for. Close the safety and privacy gaps before chasing any upside. Set a metric you can defend in a board meeting and a date to check it. Then decide, honestly, whether to expand, hold, or walk away.
At 24/7 Teach, we have supported more than 50 organizations through decisions shaped like this one, and the readiness-and-governance work that ideally comes before a purchase can almost always be retrofitted after one. It is less comfortable in that order, and it still works. We build in this space ourselves, through our sister company's K-8 platform Naomi-AI, which is why we hold a firm line on the difference between a tool pointed at a real job and an impressive object in search of a purpose.
Notice the thread running through every winning move here: the AI drafts, reaches, and assists, while a human decides, reviews, and teaches. That is the principle we build everything on, human-led and AI-facilitated. It is a narrower victory than the press release promised, and it is one you can actually reach.
About the author
I'm Justice (Justice Jones), co-founder and Chief Strategy & AI Officer, and a former K-12 principal, with over 20 years of experience in education leadership and Learning & Development. Today, I lead AI integration and build agentic AI systems in production for education and training companies, where "mostly works" simply doesn't cut it.
That work includes the seven-agent architecture and curriculum-design skills behind Naomi-AI at 24/7 Teach. This is where I write about what it actually takes to ship AI across K-12 schools, universities, and training organizations: the wins, the governance tradeoffs, and the failures included.