AI Agent Case Study Writing: What It Can and Can't Draft

A marketing agent can turn a locked set of approved numbers, the ones finance and the client both already signed off on, into a draft case study in minutes, one more task alongside the content work a marketing agent already runs every week. That is the useful half of ai agent case study writing. The dangerous half starts the moment the agent runs out of real numbers and starts filling the gaps itself: a rounder percentage, a cause nobody tested, a quote nobody actually said. Research on how large language models turn structured data into prose finds the numbers themselves usually survive the trip intact. What survives less reliably is the sentence connecting them, the one that turns "revenue up 22%" into "because we automated follow-up," when the underlying data never measured that connection at all. Since October 2024, the FTC has had a specific rule for exactly this kind of overreach, with a penalty that applies per violation. This post draws the actual boundary: what a marketing agent can draft safely from numbers you already trust, and what still needs a person's signature before it goes out under a client's name.

What does it mean for a marketing agent to draft a case study from approved numbers?

Drafting from approved numbers means the agent's job starts after the number is already settled. Someone, usually finance, an account manager, or the client directly, has already pulled the figure from a dashboard, a signed report, or a written email, and confirmed it is real and attributable to the work being described. The agent's task is narrower than most vendor pitches suggest: turn that locked set of figures into a readable narrative with a headline, a structure, a quote format, and a call to action, without adding a single number, date, or cause that was not already in the approved set. That is a drafting job, not a research job, and keeping it that narrow is what makes the draft safe to read once instead of fact-checked line by line.

What's the difference between a case study and a review or testimonial?

A case study makes a specific, falsifiable claim: this client, this metric, this date range, this is what changed. A review or testimonial is a quote about how someone feels about a product, which is harder to check but still has to be real. Both carry the same underlying duty: the words attached to a named person or company have to reflect something that person actually said or that company actually experienced. A drafted case study that hits the right structure but invents the reasoning behind a number is making the same kind of claim a fake review makes, just dressed in a longer format.

Where do AI agents already get the numbers right?

On the number itself, more reliably than most people assume. A 2025 benchmark called StructText scored how large language models turn structured, table-shaped data into prose across 71,539 examples pulled from 49 datasets. Its finding on accuracy was clear: the models preserved numeric and temporal facts faithfully and did not hallucinate new ones out of nothing. If the approved number set says a client's close rate moved from 18% to 24% between March and September, an agent drafting from that set is unlikely to misstate either figure or either month. The number is the part of the job agents are already good at.

Why do the figures survive even when the writing around them doesn't?

Because preserving a fact and explaining a fact are different skills. StructText's other finding was that the same models struggle with narrative coherence once they start writing: the correct numbers end up embedded in prose that is harder to extract cleanly, and sentences drift toward explanations the data never stated. The number stays accurate. The story wrapped around it is where an agent starts improvising, and improvising is exactly what a case study cannot afford.

Where does a drafted case study actually go wrong?

Not usually in the headline number. It goes wrong in the sentence right after it, the one that assigns a cause. An approved number set might say a client's response time dropped from four hours to twenty minutes after a build shipped. A careful draft says exactly that. An unsupervised draft often adds a reason nobody measured: "customers started buying twice as often" or "support tickets fell by half," details that sound like a natural extension of the real number but were never in the data at all. That is an unsupported claim riding in on a supported one, and it is the single easiest thing for a reviewer to miss, because the real number next to it reads as proof the whole sentence checked out.

A specific unsupported claim an agent can add without anyone noticing.

In this illustrative example, the approved number set contains one fact: average response time fell from four hours to twenty minutes over a 90-day window. A drafted paragraph reads: "Faster responses meant the client closed 30% more deals this quarter." The first half is the approved fact. The second half is an invented one, a plausible-sounding outcome the agent generated because case studies usually end that way, not because any number in the approved set said it. A reviewer skimming for the real figure sees it, confirms it, and signs off on the sentence next to it without re-checking whether that sentence was ever actually approved.

Why does this carry real legal exposure now?

Because the FTC's current rule on reviews and testimonials, effective since October 21, 2024, gives a fabricated or undisclosed-AI-written customer claim a specific, named violation instead of a vague reputational risk. The rule lets the agency seek civil penalties up to $51,744 per violation against a business that knowingly publishes a fake review or testimonial, including one generated by AI that the business knew or should have known misrepresented a real experience. A case study with an invented outcome attributed to a real client is not automatically covered word for word by this consumer-facing rule, but the FTC's longstanding Endorsement Guides already set the underlying principle for any business claim built on a named person's experience: an endorsement has to reflect that person's actual opinion, and an endorser "can't talk about your experience with a product if you haven't tried it." An agent-drafted quote attributed to a client who never said those words, or a benefit that client never reported, fails that standard regardless of which specific rule a regulator reaches for first.

What does the FTC's 2024 rule actually prohibit?

Three things that matter here directly: creating or selling a fake or false review or testimonial, including an AI-generated one dressed up as a real customer's words; publishing an insider's review or testimonial, from an employee, an officer, or a close relative, without disclosing that connection; and running a site or page that claims to show independent reviews when the business itself controls what appears on it. A case study drafted by an agent and published without any human check against the approved numbers risks the first category the moment one unsupported sentence slips past review.

Why do buyers already fact-check a case study before they believe it?

Because they assume AI touched it somewhere in the process, and they have started checking accordingly. A 2026 survey of 1,862 technology buyers and 444 vendors, fielded by TrustRadius in January 2026, found that 63% of buyers used AI during their most recent buying process, and 94% of those buyers said they fact-check what AI gives them at least some of the time. The same survey found 74% of buyers lean on customer reviews to inform a purchase decision. A buyer reading a case study today is not reading it the way a buyer did five years ago: the default assumption is now skepticism first, and a case study that cannot survive a quick cross-check loses more trust than one that was never published at all.

What has to be locked before the agent starts drafting?

One file: the approved number set, and nothing the agent is allowed to add to it. That file holds every metric, date range, and quote that finance, the account owner, and the client have all signed off on, written down in a form the agent reads as its entire universe of facts. No number, percentage, or causal claim exists for the draft unless it exists in that file first. This is the same discipline behind the brand voice guardrails a marketing agent already needs to sound consistent: a file the agent reads before every draft, not a style note buried once in a prompt and hoped to stick.

What does a safe case-study workflow actually look like for a marketing agent?

Three steps, in order. First, the approved number set gets written and signed off before the agent opens a draft, the same locked-file discipline that keeps an agent's landing page drafts from wandering off a brand's actual offer. Second, the agent drafts structure and language only: headline, narrative arc, pull quotes formatted correctly, call to action, with every number and claim traced back to a line in the approved file. Third, a named human reviewer reads the draft against that same file line by line, specifically hunting for the sentence that sounds true but was never approved, the failure mode StructText's research says is the real risk here, not a wrong number. If the draft used AI assistance in a way a reader would reasonably want to know about, disclose it plainly rather than let the writing imply a human wrote every word unaided.

If your case-study pipeline right now is one writer copying numbers from memory into a template, that is the process an agent should take over first. If it is still a Google Doc with no sign-off step before a client's name goes live, the agent is not the risk that needs fixing yet. The U.S. Bureau of Labor Statistics counted 47,940 writers and authors employed nationally in May 2025, at a mean annual wage of $86,090, the closest occupation to the person currently assembling these drafts by hand. An agent can take the drafting load off that role. It cannot take the sign-off step off a named person's plate, and no workflow above tries to.

If your last case study shipped without anyone re-checking its numbers against the client's own sign-off, Get your free Systems Plan and tell us which client story you would want drafted and checked this way first. Our team replies within one business day.

Common questions.

Can an AI agent draft a case study from approved numbers?

Yes, for the structure and the language around the numbers. Research testing how large language models turn structured data into prose found the models preserve numeric and temporal facts accurately and do not invent new ones, but they struggle with narrative coherence, so a drafted sentence can connect a real number to a cause nobody actually measured. The number is safe. The sentence explaining it needs a human check.

What's the biggest risk in letting an agent write a case study?

Not a wrong number, an unsupported claim riding next to a correct one. A draft might report an approved figure accurately and then add a result that sounds like a natural continuation of it, a sales lift or a cost saving that was never in the approved data. A reviewer checking only the headline number can miss the invented sentence sitting right beside it.

Is there legal risk to an AI-drafted case study?

Yes, if a claim or quote in it doesn't reflect something the client actually said or experienced. The FTC's rule on consumer reviews and testimonials, in effect since October 21, 2024, lets the agency seek a civil penalty up to $51,744 per violation for a fake or AI-generated review or testimonial a business knew or should have known was false. The agency's longstanding Endorsement Guides set the broader principle: an endorsement has to reflect the endorser's own actual experience, not a line written for them.

Do buyers actually fact-check case studies now?

Increasingly, yes. A January 2026 survey of 1,862 technology buyers found 63% used AI during their purchase process, and 94% of those buyers said they fact-check what AI tells them at least some of the time. A buyer reading a case study today is more likely to cross-check it than a buyer was a few years ago, which raises the cost of a claim that doesn't hold up.

What should never be left to the agent alone?

The approval of the number set itself, and the final sign-off before publish. The agent should work from a file of numbers and quotes that finance, the account owner, and the client have already confirmed, and a named human should read the finished draft against that same file before it goes out, specifically checking for any claim that reads as true but was never in the approved set.

Where these figures come from.

The factual-accuracy and narrative-coherence findings are from StructText, a 2025 benchmark evaluated across 71,539 examples from 49 datasets, presented at VLDB's TaDA workshop. The FTC rule details and the $51,744 civil penalty figure are from the FTC's own press release announcing the final rule, cross-checked against law-firm summaries of the same filing; the rule took effect October 21, 2024. The endorsement-accuracy principle is from the FTC's Endorsement Guides FAQ. The buyer fact-checking figures are from TrustRadius's 2026 B2B Buying Disconnect Report, a survey of 1,862 technology buyers and 444 vendors fielded in January 2026. The wage and employment figure is the US Bureau of Labor Statistics count of writers and authors, May 2025: 47,940 people employed, mean annual wage $86,090. The response-time and close-rate example above is an illustrative example, not a client result.

Related: Marketing agent, brand voice guardrails, landing page drafts.