Using AI in Instructional Design Without Shipping Worse Courses

Picture Wednesday afternoon. The AI-drafted module looks finished: clean screens, a tidy quiz, a tone nobody will object to. On Friday the SME sends it back with a comment on nearly every screen, and the launch date you promised hasn't moved.

If you have opened that Friday email, you know the quiet math: the draft saved you two days, and the review just spent three. Nobody did anything wrong. The tool did what it is good at, and your SME did what she is there for.

AI in instructional design pays off when the draft is the cheap part and a named person checks the tasks where AI is weak. In my view, the one decision you never hand to a tool is which behavior the course exists to change. What follows is a task-by-task map, a five-check review to run before your SME sees anything, and a way to tell whether you're actually saving time.

Key takeaways

• Let AI draft where a wrong answer is easy to spot, and keep a person on the target behavior, the scenario's consequences and every factual claim in a regulated topic.
• Give each AI-assisted step one named checker and one yes-or-no question, so SME review shrinks to a short, predictable routine.
• Count SME review rounds on your next two AI-assisted modules before you promise the business a throughput gain.
• Write the one-sentence behavior before the first prompt, because AI will produce a fluent course about any topic, including the wrong one.

Why does an AI draft that looks finished come back with a comment on every screen?

AI writes plausible content well. It is weak at the judgments that make content teach: whether it fits the behavior, the learner and the facts. A polished draft hides that gap until an SME reads closely, and the time saved on drafting turns up again in review.

The tools are already in use. According to an ATD blog post by Tessa Guiv on ATD Research's July 2025 report, 80% of the 232 instructional designers surveyed currently use AI tools while designing courses. That figure is self-reported use and says nothing about quality either way.

In a January 2026 article in ACM's eLearn magazine, Natalia Echeverry argues that AI-powered authoring tools have focused on generating content to cut production time, with too little thought for the quality of the learning experience.

In my view, a draft that looks finished is the most expensive kind. Everyone who skims it is tempted to approve it, and the slow reading falls to the one person whose calendar is already full.

Where does AI in instructional design save time, and where must a senior person decide?

Let AI draft where a wrong answer is cheap to spot, and keep people on the decisions that settle whether the course works. This map is my recommendation and no study has tested it, so move items into the middle group when the stakes are higher.

• Draft and check: Outlines, first drafts of question banks, tone rewrites and first-pass summaries of source material. Your designer can spot the mistakes on the page.
• Draft and verify: Any sentence about what a policy says or how a system works. AI drafts it, and the SME or policy owner signs it off before a learner sees it.
• Keep human: The target behavior, the obstacles, the scenario's consequences and what counts as success. These decide whether the course teaches anything at all.

In a May 2026 post on a custom e-learning company's blog, Jennifer Bell separates AI-generated content from AI-assisted design, with a sharp example: “An AI can write a beautiful scenario about customer service that doesn’t develop any of the specific decision-making skills the training is supposed to build.”

Our own line is on the Euvouria site: “AI helps us move faster through drafting, structuring, and production. It does not replace learning strategy, behavior mapping, scenario design, or quality judgment.” I'd skip AI entirely for the behavior sentence and the scenario's consequences, because a fluent wrong answer there looks exactly like a right one.

What does a human check look like at each step?

Give each AI-assisted step one named checker and one yes-or-no question they can answer in minutes, and run the checks before the module goes to the SME.

Picture a compliance training lead at a regulated multinational, working with one designer and one SME from information security. On a Tuesday call they agree on the request: a module on what to do when a colleague wants to send a client file to a personal email address. Within the hour, the AI tool returns a tidy draft.

Sent straight to the SME, that draft opens with a policy summary and ends with a quiz where the right answer is obvious. She comments on every screen (one policy line is slightly off, the scenario has no pressure, the quiz tests recall), and the hours the tool saved go into rewriting.

The five-check review, filled in for the personal-email module

Every detail here is hypothetical, so swap in your own behavior, roles and policy.

1. Behavior check (you, before the first prompt): Write one sentence about what people will do differently on the job: “When a deadline tempts them to send a client file to a personal address, they use the approved route and tell their manager.” If the draft doesn't build toward it, stop and re-prompt.
2. Scenario check (your designer): Is the wrong choice tempting, and does it have a visible consequence? Here, a senior colleague with a deadline does the asking, and the learner sees where the file ends up. Ask AI for a scenario and, in my view, it tends to hand you a right-or-wrong quiz in costume, which is why branching scenarios that teach decisions need a person's hand.
3. Fact check (SME or policy owner): Mark every sentence that states what a policy says or how a system works, and send only those. Here, that comes to three flagged statements.
4. Alignment check (your designer): For each quiz question, ask which part of the behavior sentence it tests, and rewrite any question that tests recall of the policy wording. If the objective itself can't be seen on the job, rewrite it so someone can see it first.
5. Learner check (you or a manager): Ask one person who does this job to read the module and say where they wouldn't do it that way. Log the answer with the date and the module name.

This time the SME opens a draft with three flagged statements and a scenario where a senior colleague is pushing. Her comments land on the consequences and the lines only she can verify. And the lead can say what learners will do differently and who checked each part.

How much time does AI really save once review is counted?

Designers report savings, but the surveys ask about their own time and leave out the hours spent reviewing and fixing. Until you've counted those hours on your own modules, treat any saving as unproven.

In a February 2024 conference paper, Laura McNeill of the University of Alabama surveyed 144 instructional designers: 67% reported moderate-to-significant time savings from generative AI, and efficiency ranked as the top benefit. Those answers are self-reported, and they say nothing about course quality or how long review took.

“How do I know it's actually saving time once the SME has been through it?” Measure it on your own work.

• Log: On your next two AI-assisted modules, record SME review hours and the number of comment rounds.
• Compare: Set those next to the last two modules your team built the usual way.
• Decide: Only then tell the business what throughput you can promise.

The number I'd watch is SME rounds per module. (Drafting hours get quoted more, mostly because they're the easiest to count.) If rounds go up while drafting time goes down, you've moved the work onto the busiest person in the building.

What if the business just wants more courses, faster?

Answer with scope: let AI speed up the safe parts, and spend the time it frees on the behavior sentence and the SME's review window.

“Fine, but my VP wants twice the courses this year and I have the same two designers.” Fair enough. I'd ask for one working week before the first draft, to settle the behavior sentence and name the single SME who approves each module.

• Instead of: Promising twice the courses and letting the SME absorb the extra review.
• Try: Agreeing on the behavior and the approver first, then letting AI draft inside that boundary.

SMEs with no time to spare get the same answer: one accountable approver, one consolidated review and only the flagged statements to verify. A one-page training brief keeps the behavior sentence and the approver's name where your VP can see them.

When does the human decision happen in our method?

Before the first prompt. In our Behavior Bridge method, that decision belongs to Define the Behavior, where a Behavior Workshop turns the request into one behavior and produces the Behavioral Brief, a one-page document with a measurable behavior goal.

Think back to that Friday email. Next time, the SME opens a draft with the facts already flagged and the scenario checked against the behavior, so her comments go to the few places only she can judge. This week, pick the next AI-drafted module and run the five checks before it reaches her.

If you already have an AI-drafted module waiting on an SME, send that one module to us for a free diagnostic, and we'll send back the three highest-impact improvements within one to two working days.