Migration noise
Institutions midway through replacing that platform often show conflicting public signals. A university may publicly announce a move to a new platform months before the legacy environment is fully retired.
Evidence
Can commercial teams trust frontier AI models on their own for intelligence about an institution?
We ran two practical tests to evaluate how reliable frontier artificial-intelligence models are when used on their own for commercial research. In both, the models were allowed multiple passes rather than a single attempt, to reflect how teams actually use them.
Executive summary
The first test focused on identifying the correct LMS for a sample of education accounts. The second focused on identifying the right teaching and learning personas within a smaller set of accounts. In both cases the models were allowed multiple passes rather than a single attempt, to better reflect how teams actually use them in practice.
The conclusion was consistent across both studies: frontier LLMs are useful research assistants, but they are not dependable enough on their own for high-precision GTM intelligence.
Even the best-performing LLM in these tests still produced enough errors to create real execution risk. NEO GTM achieved 98.5% accuracy on LMS identification and 98.0% on persona identification.
Study 1
Identifying the correct LMS by account
The purpose was to test how accurately frontier LLMs can identify a current account-level technology fact at scale. LMS identification is commercially relevant and easy to verify. Getting it wrong distorts account prioritisation, weakens displacement strategy, creates poor personalisation, reduces reply rates and wastes seller time.
We sampled 200 education accounts across 13 countries — Australia, Brazil, Canada, Chile, Malaysia, Mexico, the Philippines, Qatar, Singapore, Spain, the United Arab Emirates, the United Kingdom, the United States — and asked three frontier model workflows to identify the current LMS used by each institution, comparing the outputs against a verified source-of-truth dataset. The models were allowed multiple attempts.
NEO GTM was right on 197 of 200. Gemini, the strongest LLM-only result, was wrong on 52 accounts — too high to treat LLM-only research as dependable operational data at GTM scale.
Learning-platform identification accuracy 200 institutions · 13 countries
What the LMS errors tell us
System versus campus variation
In larger groups or multi-campus systems, the central institution and individual campuses may not use the same platform.
White-labeled platforms
Branded portals can obscure the underlying vendor, making public identification less reliable.
Departmental autonomy
Some sub-units operate differently from the institutional standard, creating misleading web footprints.
Heuristic bias
Models tended to over-apply broad market assumptions rather than what was true of the specific institution — favouring what seemed statistically likely over what was actually correct.
Study 2
Identifying the right teaching and learning personas
The second study tested a harder and arguably more important task: finding the right people, not just the right account facts.
We sampled 21 unique accounts and asked the same models to identify key personas in teaching and learning, comparing against a verified list of 200 personas. The models were allowed four passes.
This is the sharper warning. The models were far less reliable at identifying the right people than the right technology environment — and persona accuracy is often what determines whether outreach reaches the people who actually own the opportunity.
Identifying the right people 200 people · 21 institutions · 4 passes
What the persona errors tell us
Seniority bias
The models often surfaced Provosts, Vice Chancellors, VPs and other senior academic leaders because those roles are more visible on the web. The verified dataset often contained the more relevant operational teaching and learning buyers one or two levels below.
Functional mismatch
The models frequently found legitimate academic leaders, but not the specific teaching and learning function leads expected in the source-of-truth dataset.
Unit-level blindspots
At many universities the actual buyers sit inside named teaching innovation, digital learning or academic development units that are much less visible than the executive leadership team.
Regional title variation
In Latin America and Southeast Asia in particular, function ownership can sit in institution-specific structures that do not map neatly to generic English-language job title assumptions.
Verified record versus web reality
Public web research tends to surface the most visible people, while the verified dataset often captures the practitioners and operational leaders who actually matter in a buying process.
What the two studies show together
Frontier LLMs cannot be trusted by themselves as a reliable source of GTM account intelligence.
They are valuable for speed and useful for synthesis. They can narrow the search space and surface promising directions. But when the task requires current, account-specific, persona-specific truth, they still make too many mistakes to be treated as dependable on their own.
The lesson is not to avoid LLMs. It is to avoid LLM-only workflows for high-precision research. A dependable GTM workflow needs a layer around the model that understands the industry, improves research quality, reduces false confidence, and makes outputs usable in live execution.
In these tests NEO GTM delivered materially higher accuracy — 98.5% on LMS identification and 98.0% on persona identification, or up to eight times more accurate than the LLM-only workflows tested.
Published 9 April 2026. Method and error analysis as stated above.
Accuracy is a product decision
The lesson is not to avoid artificial intelligence. It is to avoid relying on a model alone for research that has to be precise — and to put a layer around the model that understands the industry and reduces false confidence.