Enterprise AI detection: the B2B layer most people never see

Enterprise AI detection is the business-to-business side of the detector market: API-first products from vendors like Sapling, Copyleaks, Writer, and Pangram, sold on volume pricing, compliance certifications, and audit trails rather than free web scans. The tools are often stronger than consumer versions — but their accuracy claims are even harder to verify, because almost nobody tests them at scale in public.

One line of disclosure before anything else: we build a detector and humanizer ourselves. Here's our editorial policy — judge accordingly.

When most people picture an AI detector, they picture a text box. Paste your essay, click a button, get a percentage. That's the consumer layer, and it's the layer that gets reviewed, screenshotted, and argued about on Reddit.

There is a second layer underneath it, and it's where most of the money is. Hiring platforms screening ten thousand cover letters a night. Marketplaces checking product reviews before they publish. Publishers scanning freelancer submissions. Learning-management systems piping every student upload through a detection API before an instructor ever sees it. None of this happens in a text box. It happens through API calls, batch jobs, and webhook callbacks, governed by contracts that the public never reads.

This post maps that layer: who sells it, how the products actually differ from the consumer tools, why the pricing looks the way it does, why enterprise accuracy claims are the least verifiable claims in an already murky market, and — if you're the person running a pilot — exactly what to demand before you sign.

How enterprise needs differ from consumer needs

A student checking one essay and a trust-and-safety team checking two million reviews a month are not buying the same product, even when the underlying classifier is identical. Four differences drive everything else.

API-first, not interface-first. The consumer tool is a destination; the enterprise tool is plumbing. An enterprise buyer needs a REST endpoint with predictable latency, batch submission, sensible rate limits, and machine-readable output — a confidence score per document or per sentence, not a colored highlight designed for human eyes. Sapling, Copyleaks, and Pangram all sell dedicated detection APIs; Writer gates API access to its Team and Enterprise plans (Writer help center, checked August 2026). If a vendor's API documentation is thin or hidden behind a sales call, that tells you something about who they actually serve.

Volume, and the failure math that comes with it. Scale changes what an error rate means. A 1% false positive rate is invisible to an individual user — they'd need a hundred scans to expect one bad flag. Run 10,000 human-written documents a day through the same detector and you're generating roughly 100 false accusations daily, every day, forever. We walk through this arithmetic properly on our accuracy hub, but the enterprise version of the lesson is short: at volume, the false positive rate isn't a caveat. It's the product. It determines how many support tickets, appeals, and wrongly rejected candidates you manufacture per week.

Compliance and security. Consumer users click "agree" and paste. Enterprise buyers send customer data, student work, or job applications to a third party, which drags in a procurement checklist: SOC 2 reports, GDPR data-processing agreements, data residency, retention policies, single sign-on, sometimes self-hosting. Copyleaks advertises SOC 2 and SOC 3, GDPR, PCI DSS, and NIST RMF alignment (copyleaks.com, checked August 2026). Sapling's enterprise tier lists SSO/SCIM, advanced security options, and a self-hosted deployment option (sapling.ai/pricing, checked August 2026). These features have nothing to do with detection accuracy — and they routinely decide which vendor wins the contract anyway.

Audit trails. When a consumer detector is wrong, someone has a bad afternoon. When an enterprise detector is wrong, someone may lose a job offer, a marketplace account, or a grade — and the organization needs to reconstruct why. That means logged scores, versioned models ("which classifier build flagged this document on March 3?"), exportable reports, and admin dashboards. A score you can't reproduce later is a liability, not a feature. Detectors update their models continuously, so yesterday's 92% may be unrecoverable today unless the vendor logs model versions. Ask.

The vendors: who actually sells this

Details below were checked against each vendor's own site in August 2026. Pricing changes; treat the numbers as a snapshot and re-verify before you buy.

Sapling

Sapling started as an AI writing assistant for customer-facing teams, and its detector inherits that B2B DNA. The free web detector truncates at 2,000 characters — roughly 400–500 tokens — while paid access extends single queries to 100,000 characters, and the company asks buyers exceeding 5 million characters a month to contact it directly (sapling.ai, checked August 2026). The detector API is metered and usage-based, documented in a developer portal, with a stated focus on use cases like resume screening. Seat-based plans run $25/month for Pro ($12/month billed annually), with enterprise starting around $15/seat/month at a 10-seat minimum.

Sapling claims a 97%+ detection rate on AI content and under 3% false positives on human text for longer documents, and lists coverage of current frontier models including GPT-5, Claude 4.5, and Gemini 2.5. To its credit, the same page says plainly that "no current AI content detector (including Sapling's) should be used as a standalone check" and that false positives and false negatives will occur. That is the right caveat, stated where buyers can see it. Worth noting: Sapling was one of the seven detectors in the Stanford TOEFL bias study, which we take apart in our deep-dive on Liang et al. — a 2023 result, against a much older model, but a reminder that vendor caveats exist for a reason.

Writer

Writer is the odd one out: an enterprise generative-AI platform that happens to ship a detector, rather than a detection company. The web tool is free with a 60-word minimum and a 5,000-word cap; API access requires a Team or Enterprise plan, with Team API usage capped around 500,000 words monthly and higher enterprise limits negotiable (Writer help center, checked August 2026). Writer's framing is refreshingly modest — its own documentation says the detector "won't be 100% accurate" and describes it as an indication of likelihood, noting that human writing which follows machine-typical word patterns can trigger it. Writer publishes no headline accuracy percentage at all, which — given what the rest of this article is about — reads less like an omission than a decision. For an enterprise buyer, Writer's detector is best understood as a checkbox inside a broader content platform, not a standalone screening system.

Copyleaks

Copyleaks is probably the most institutionally entrenched vendor on this list. It sells consumer plans (Personal at $16.99/month for 100 credits, where one credit covers up to 250 words; Pro at $99.99/month for 1,000 credits and 25 seats), but its center of gravity is custom-priced Enterprise and Education contracts, with API access included only at that level (copyleaks.com/pricing, checked August 2026). The LMS integration list — Canvas, Moodle, D2L, Blackboard, Schoology, Sakai, Edsby — is the tell: this is infrastructure for institutions. Add the compliance stack (SOC 2/3, GDPR, PCI DSS, NIST RMF), white-labeled API options, and support for 30+ languages, and you have the closest thing to a Turnitin-shaped competitor outside Turnitin itself.

The accuracy claims are aggressive: "99% accuracy backed by independent third-party studies," a stated false positive rate of 0.03%, and per-language figures such as 99.97% on human English text (vendor site, checked August 2026). Treat those numbers as marketing until you can trace the methodology — the third-party study Copyleaks most often cites is Orenstrakh et al., posted to arXiv on 10 July 2023 and never peer-reviewed — an eternity ago in model generations, on a corpus of 164 documents. A 0.03% false positive rate is an extraordinary claim; at that rate you'd expect three bad flags per 10,000 human documents, which independent academic testing of the detector category has never come close to confirming across realistic, adversarial conditions.

Pangram

Pangram Labs is the newest entrant and the most explicitly research-forward. It markets against the perplexity-based approach the older detectors use, training deep-learning classifiers on large paired corpora of human and AI text instead, and it publishes third-party evaluations — a University of Chicago study of product reviews measured Pangram at a 0.5% false positive rate versus 2.4% for GPTZero and 1.7% for Originality.ai (pangram.com, checked August 2026). Its headline claims run to "99.9%+ accuracy" and a false positive rate of 1 in 10,000.

The commercial structure is unusually transparent for this market: a free tier at 2,000 words/day, individual plans at $20/month for 300,000 words, a professional tier at $65/month including API credit, and a straightforwardly published API price — $0.05 per 100 words for its current-generation model (pangram.com/pricing, checked August 2026). Customers listed include Quora and WikiEducation. Pangram deserves genuine credit for publishing methodology and per-domain error rates rather than a single magic number. It still deserves the same skepticism as everyone else on the claims themselves: 1-in-10,000 false positives is a lab-conditions figure until someone outside the company reproduces it on text the model has never seen.

Pricing models compared

The consumer market sells subscriptions. The enterprise market sells four different shapes of contract, and the shape tells you how the vendor thinks about your risk.

All figures from vendor sites, August 2026. Re-verify before purchase.

Notice the pattern. Per-seat pricing (Sapling) assumes humans review each result. Credit pricing (Copyleaks) meters documents and lets the vendor blend detection with plagiarism scanning in one wallet. Per-word API pricing (Pangram) is the purest infrastructure play — it scales linearly with your pipeline and makes cost modeling trivial. Custom contract pricing (Copyleaks enterprise, everyone's education tier) maximizes vendor deal size and minimizes public comparability. None of these is dishonest. But only one of them lets you compute, in advance, exactly what screening a million documents will cost.

Why enterprise claims are even harder to verify than consumer ones

Consumer detector claims are at least falsifiable by anyone with patience. You can paste a hundred known-human essays into a free tool and count the flags — people do, constantly, and the results circulate. Enterprise claims escape even that weak accountability, for five structural reasons.

First, nobody outside can test at the tier that's sold. The enterprise product may run different thresholds, different model versions, or custom sensitivity settings (Copyleaks explicitly sells adjustable sensitivity) than the free demo. Testing the text box tells you about the text box.

Second, pilots happen under NDA. The organizations best positioned to publish real-world error rates — a hiring platform that screened 500,000 applications — are contractually and reputationally the least likely to. No company wants to announce how many candidates its detector wrongly rejected.

Third, the vendor grades its own homework. Most published "independent" studies in this market are commissioned, or run on corpora the vendor supplied. Even honestly run internal evaluations suffer from training-data contamination: if the benchmark's human text was scraped from the public web, the model may have seen it. The RAID benchmark (ACL 2024) — over 6 million generations, 11 adversarial attacks — found that detectors advertising 99%+ accuracy were "easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models." The gap between brochure and benchmark is the whole story.

Fourth, base rates hide in the marketing. "99% accuracy" blends two numbers a buyer needs to see separately: the catch rate on AI text and the false positive rate on human text. If only a small fraction of your incoming documents are AI-generated, even a good detector produces a flag pool heavily salted with innocent authors. Run the arithmetic before the pilot, not after — our how detectors work explainer covers why the threshold behind these numbers is a business decision, not a law of nature.

Fifth, model drift outpaces contracts. You sign in January against a detector evaluated on last year's models. By June, the text flowing through your pipeline comes from models that didn't exist at evaluation time. Sapling and Pangram both advertise continuous retraining, which is the right response — and also means the product you tested is not the product you're running by renewal time. The historical record here is humbling: OpenAI retired its own AI text classifier in July 2023 after conceding it caught just 26% of AI text while falsely flagging 9% of human text. If the organization that built the generator couldn't reliably detect its output, treat every vendor's frozen-in-time accuracy number as provisional.

The buyer's evaluation checklist: running a pilot that means something

If you're the person responsible for a detection pilot, your job is to convert marketing into measurements. The protocol below is the enterprise version of the open framework we describe in our replication methodology post; the principles are identical.

Build your own corpus, from your own domain. Vendor demo corpora are the vendor's home field. Assemble at least a few hundred documents you know are human — pre-2022 archives from your own systems are gold, because their provenance predates ChatGPT — plus fresh human text written under observation, plus AI text generated by the models your users actually use, at current versions, with prompts resembling real abuse. Test the mixed case too: human drafts lightly edited with AI, and AI drafts heavily edited by humans. Every published study agrees this blended middle is where detectors are weakest; the Washington Post's April 2023 Turnitin test found exactly that, and Turnitin itself acknowledged blended documents are the hard case.

Measure false positives and false negatives separately, at your operating threshold. Get raw scores from the API, not just verdicts. Plot the distribution. Ask where the vendor's default threshold sits and who chose it.

Demand subgroup numbers. The Stanford study found seven detectors falsely flagged non-native English writers' essays 61.22% of the time on average while sailing through native-speaker text. If your pipeline touches international applicants, ESL students, or non-English content, test those populations explicitly. A detector that performs beautifully on average and terribly on a protected subgroup is a lawsuit with a dashboard.

Pin the model version. Require the API to return a model/version identifier with every score, and require contract language covering notification when the model changes. Otherwise your audit trail is a log of numbers no one can ever reproduce.

Price the errors, not the API calls. A vendor charging twice as much per word but producing half the false positives is usually the cheaper product once you cost out appeals, support load, and wrongly rejected humans. Decide before the pilot what a false positive costs your organization in dollars and reputation. That number, multiplied by measured FP rate and volume, is the real price tag.

Decide what the score is allowed to do. The most defensible deployments use detection as one weak signal routed to a human, never as an automated verdict. Turnitin — whose incentives run the other way — displays an asterisk instead of a number for documents under 20% flagged, precisely because low-range scores are unreliable. If the market leader in academic detection won't let a low score stand alone, your procurement policy shouldn't either.

We'll say the quiet part once: our own AI detector gives sentence-level readouts and a free 10,000 detection words a month, and we publish no accuracy percentage for it — because we haven't yet published a methodology that would make such a number honest, and a number without a methodology is marketing. That standard is the same one we're urging you to hold every vendor in this post to.

FAQ

What is enterprise AI detection? It's AI-text detection sold to organizations rather than individuals: API access, batch processing, volume pricing, compliance certifications (SOC 2, GDPR), admin dashboards, and audit logging. Vendors include Sapling, Copyleaks, Pangram, and Writer. The classifier may resemble the consumer tool; the contract, controls, and failure economics do not.

Which enterprise AI detector is the most accurate? Nobody outside the vendors can currently say. Copyleaks claims 99% accuracy with 0.03% false positives; Pangram claims 99.9%+ with 1-in-10,000 false positives; Sapling claims 97%+ with under 3% false positives — each with fine print. Independent academic testing (the RAID benchmark, ACL 2024) found detectors fall well short of advertised numbers under adversarial conditions. Run your own pilot on your own documents.

Why do enterprises pay for detection when free tools exist? Free tools cap scan length, offer no API, no SLA, no compliance paperwork, and no audit trail. An organization screening thousands of documents daily needs machine-readable scores, logged model versions, and a data-processing agreement — none of which a free text box provides.

How much does enterprise AI detection cost? As of August 2026: Sapling's enterprise seats start around $15/month with usage-based API pricing; Copyleaks enterprise is custom-quoted; Pangram publishes $0.05 per 100 words for API detection with custom enterprise plans above that. Expect real enterprise contracts to be negotiated, not listed.

Can an enterprise detector's false positives create legal risk? Potentially, and especially in hiring and education, where a wrong flag affects a person's livelihood or record and where disparate impact on non-native English speakers is documented in peer-reviewed research. This isn't legal advice — but any deployment that auto-rejects humans on a detector score alone should go past your counsel first.

Should detection scores ever trigger automatic action? The evidence says no. Every serious vendor caveat (Sapling's "not a standalone check," Writer's "won't be 100% accurate," Turnitin's asterisk policy) points the same direction: detection is a screening signal for human review, not a verdict. Automating rejection on a probabilistic score at scale guarantees a steady stream of wronged humans.

How often should we re-evaluate a deployed detector? Quarterly, minimum. New generator models ship constantly, vendors retrain silently, and an evaluation older than a couple of model generations describes a product that no longer exists. Build re-testing into the contract.

Key facts

• Sapling claims a 97%+ detection rate and <3% false positives "on longer texts," free scans truncate at 2,000 characters, and enterprise volume starts above 5M characters/month (sapling.ai, August 2026).
• Copyleaks claims 99% accuracy and a 0.03% false positive rate, holds SOC 2/3, GDPR, PCI DSS, and NIST RMF alignment, and gates API access to custom enterprise contracts (copyleaks.com, August 2026).
• Pangram publishes API pricing at $0.05 per 100 words and claims a 1-in-10,000 false positive rate; a University of Chicago evaluation measured it at 0.5% FP on product reviews vs 2.4% for GPTZero (pangram.com, August 2026).
• Writer's free detector spans 60–5,000 words; API access requires Team (~500K words/month) or Enterprise plans; Writer publishes no accuracy percentage (Writer help center, August 2026).
• The RAID benchmark (ACL 2024) — 6M+ generations, 11 generators, 11 adversarial attacks — found detectors claiming 99%+ accuracy were "easily fooled" by attacks and unseen models.
• OpenAI retired its own AI text classifier in July 2023 at a 26% catch rate and 9% false positive rate.
• Liang et al. (Patterns, 2023) measured a 61.22% average false positive rate on non-native English speakers' TOEFL essays across seven detectors, Sapling among them.

Sources

1. Sapling — AI Content Detector page and pricing (sapling.ai), fetched August 2026.
2. Copyleaks — AI Content Detector page and pricing (copyleaks.com), fetched August 2026.
3. Pangram Labs — homepage and pricing (pangram.com), fetched August 2026.
4. Writer — "AI content detector," Writer Help Center (support.writer.com), fetched August 2026.
5. Dugan, L., et al. "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors," ACL 2024 (arXiv:2405.07940).
6. Liang, W., Yuksekgonul, M., Mackey, L., Wu, E., Zou, J. "GPT detectors are biased against non-native English writers," Patterns (Cell Press), 2023.
7. OpenAI — "New AI classifier for indicating AI-written text" (update announcing retirement, July 2023).
8. Turnitin — AI writing detection FAQ (asterisk policy, >20% conditions).
9. Fowler, G. "We tested a new ChatGPT-detector for teachers. It flagged an innocent student." The Washington Post, April 2023.