When a new technology arrives, most companies reach for the same first move: train people.
It’s a sensible instinct. Nobody wants managers pasting questions into a chatbot with no idea how it fails. But training answers one question, whether people know how to use the tool, and skips the one that decides the outcome: is the work set up so the tool can be used safely?
You can raise literacy for everyone and still have a workflow that isn’t ready.
Those are different questions. Treating them as one is how a team can complete the course and still hand an AI assistant a process nobody owns.
Two Words, Two Problems
Here are the plain-language versions.
AI literacy is what a person knows: what the tool is good at, how it gets things wrong, and how to ask it for something useful.
AI readiness is a property of a specific piece of work. It asks whether the procedure is written down, someone owns it, someone checks the output before it reaches a customer, and there is a place for the cases the tool shouldn’t handle.
One lives in a person’s head. The other lives in the operating system. You can raise literacy for everyone and still have a workflow that isn’t ready.
What The Evidence Does And Does Not Say
Start with the finding that surprised researchers. In two Aalto University studies, 246 and then 452 participants used AI to solve logical-reasoning problems from the LSAT. Their scores were 3 points higher than a comparison population, but they overestimated their own performance by 4 points. Higher AI literacy was correlated with less accurate judgment of their own results: more technical knowledge went with more confidence and less precision. Aalto University
Read that carefully. It’s a correlation, not proof that learning about AI makes people overconfident. It covers one kind of task, logic puzzles. It is worth taking seriously because it points at a gap: knowing more about the tool didn’t make people better at noticing when its answer was wrong.
A second study shows the gap in a work setting. Harvard researchers and Boston Consulting Group randomly assigned 758 consultants to no AI, GPT-4, or GPT-4 plus a brief prompt-engineering overview. On a creative product task, AI users completed 12.2% more subtasks, worked about 25% faster and received quality ratings about 32% higher. On a brand-strategy case that required spotting a subtle clue in interview notes, accuracy was 84.5% without AI, 70.6% with it and 60% with the added overview. Harvard Business School
The study’s own explanation was over-reliance: people trusted the analysis and followed it to the wrong conclusion. Harvard Business School It also found that professionals had “no obvious signal” telling them which tasks sat inside AI’s strengths and which outside. Harvard Business School
That last point is the readiness argument. If individuals cannot reliably tell when the tool is out of its depth, you cannot depend on each person’s judgment in the moment. The check has to be built into the work.
The limits matter here too. It was one study, two assignments, consultants rather than store managers, and the extra instruction was a short overview, not a serious training program. The published summary also doesn’t say whether the gap between the two AI groups is statistically reliable. The safe reading is that a little instruction did not protect people on the hard task. It does not show that training is useless.
Your People May Not Be The Bottleneck
The other half of the evidence is almost an opposite warning. McKinsey’s 2025 workplace study concluded that employees were more ready than leaders assumed: it found three times as many employees were using generative AI for a third or more of their work as leaders imagined. McKinsey
If that holds in your business, many of your managers are already using AI. A training session isn’t the starting point for a skill gap; it is catching up with behavior that exists. The decision that remains is where AI use is allowed to touch customer promises, lease dates, candidate answers or site decisions.
Readiness answers that. Literacy doesn’t.
Five Questions For One Workflow
Pick one workflow, not the whole company. For example, answering field questions from the operations manual, tracking opening checklist items, summarizing a lease, or responding to prospective franchisee questions.
- Source: Is there one current, approved version of the procedure or information the tool will use?
- Owner: Can you name the person who corrects the source when it changes?
- Check: Who reviews the output before it affects a customer, landlord, candidate or opening date, and how quickly?
- Boundary: What must the tool not answer, and where does that case go?
- Baseline: Do you know how long the work takes and how often it goes wrong today?
If you can’t answer one or two, the workflow isn’t ready. That is not a reason to ban AI. It is the work to finish before you scale it.
| Workflow | Literacy question | Readiness question |
|---|---|---|
| Field SOP assistant | Do managers know it can sound confident and be wrong? | Which version of the procedure does it use, and who updates it? |
| Lease summary | Do users understand that extracted dates can be missed? | Who verifies the critical dates before they go on the calendar? |
| Franchise candidate Q&A | Does the team know not to improvise claims? | Is it limited to approved answers with a human route for anything else? |
| Opening checklist | Can staff write a clear request? | Is each handoff owned, and what does “done” mean? |
Where The Other Side Is Right
Literacy matters, and the case for it is stronger than a simple contrast makes it seem.
A person who doesn’t understand how these tools fail will not design a good check. Someone has to know why a polished answer can be wrong, and that is a form of literacy. A workflow can look ready on paper and still fail if the people running it rubber-stamp the output.
The evidence on training is thinner than the evidence on overconfidence. The Aalto result is an association, not a test of training (Aalto University), and the Harvard study tested a brief overview, not a full program (Harvard Business School).
The tools are also changing. The research group METR found in 2025 that experienced developers took 19% longer with AI, with a confidence interval of +2% to +39%. In February 2026 it said developers were likely more sped up by early 2026 than that estimate suggested, although its newer data was too unreliable to size the effect. METR Studies of how people misjudge AI can age quickly.
Some regulators have also decided literacy is worth requiring. The EU AI Act, as amended in 2026, tells providers and deployers of AI systems to take measures to support their staff’s AI literacy, while saying this does not require guaranteeing any specific level for an individual. EU AI Act National authorities began supervising Article 4 on August 2, 2026. European Commission That matters only if you operate in the EU. It does show that literacy is easy to require and measure, which may be one reason companies reach for it first.
The practical position is both: train people, and don’t mistake the training record for readiness.