Stay ahead of the curve
Get exclusive insights delivered directly to your inbox. Join thousands of tech leaders.
TL;DR
- For orientation, yes: ChatGPT and Copilot summarise an unfamiliar field in minutes. For a funding decision, no. A fluent answer doesn't show what the search missed, and the assistant has no way to measure which technology improves fastest.
- No AI tool can be trusted not to hallucinate. Trust the ones whose every claim traces back to a document you can open, and whose forecast is not generated text.
- Grounding narrows the error, it doesn't remove it: a 2024 Stanford study found leading retrieval-based AI legal research tools hallucinating 17% to 33% of the time.
- GetFocus splits the job: language models read every patent family in full, and explainable maths on the citation network forecasts the winner, validated on 150+ domains.
For a first look, yes. ChatGPT and Copilot summarise an unfamiliar technology field in minutes and draft a decent long list of options. For a decision, no. A chat assistant cannot show you what its search missed, and it cannot measure which technology is improving fastest, which is the question scouting exists to answer. GetFocus splits the job in two: language models read every relevant patent family in full, and explainable maths on the patent citation network does the forecasting, validated on 150+ technology domains at R² ≈ 0.7. The model reads. It never guesses the winner.
What can ChatGPT and Copilot do well in technology scouting?
A general assistant is a strong first-day analyst. Use it to:
- explain an unfamiliar field and its vocabulary
- list the approaches competing for a job, including some from adjacent industries
- summarise the papers, reports and patents you hand it
- draft a first-pass landscape or a briefing note
Copilot can also draw on documents already inside your Microsoft 365 environment, which helps when the context for a scouting question lives in internal reports. It needs no new software. The trouble starts when the first look becomes the answer.
Where do general AI assistants fall short?
On the things a funding decision rests on.
| What scouting needs | ChatGPT or Copilot | Patent and scouting platforms | GetFocus |
|---|---|---|---|
| Coverage | Web pages and what the model has seen; no proof of completeness | Structured patent, paper and company records | Every patent family in the technology, retrieved by meaning across 180M+ patents, then read in full |
| Evidence | Citations when it browses; you check each one | Records you can open | Every claim traces back to the patent families behind it |
| Ranking the options | A plausible opinion | Counts, maturity scores, landscapes | Improvement rate (TIR) and maturity (TRL) for every competing technology |
| Staying current | Starts from zero in every chat | Alerts and dashboards | Rates recomputed as new patents publish, monthly deltas, trigger alerts |
| Shared record | Findings stay in chat threads | A shared workspace | An evidence table your experts can interrogate |
Coverage is the gap you can't see. A well-written answer reads the same whether the search found twenty relevant patents or two thousand. GetFocus casts a deliberately wide net by meaning, then has a language model read each retrieved family in full (title, abstract, claims and description) and rule it in or out against one written instruction. Nothing is skimmed and nothing is sampled (how the dataset is built).
Ranking is the gap that costs money. An assistant ranks technologies by what it has read about them, and what gets written about is what gets filed and funded. In every GetFocus backtest, filing volume pointed at the losing technology. Hydrogen out-filed every other powertrain for 25 years, while NMC batteries improved at 54% a year against hydrogen's 21% from the first data point in 2001 (mobility backtest).
Which AI tools for technology scouting can you trust not to hallucinate?
None, if trust means never wrong. Language models generate the most probable text, and grounding them in documents narrows the error without removing it. A 2024 Stanford study tested leading AI legal research tools built on retrieval over curated case law: they hallucinated less than a general chatbot, and still between 17% and 33% of the time (Magesh et al., 2024).
So don't ask a vendor whether its AI hallucinates. Ask where the model is allowed to be wrong, and how you would catch it:
- Does every claim link to a document you can open? In GetFocus, every finding traces back to the patent families behind it.
- Is the model judging one document, or answering from memory? GetFocus makes each relevance call on one family's own text against the same instruction, never on the abstract alone.
- Is the forecast itself generated text? In GetFocus it is not. The rate is an explainable regression on two citation metrics.
This applies to anyone using AI output in a funding, kill or timing decision. If you are only getting oriented in a new field, a general assistant's first draft is fine: check the three claims you plan to repeat.
Can AI really tell which technologies will win, or will it make things up?
Ask a language model "which technology will win?" and it writes the most plausible answer to that question. Plausible is not measured. A forecast deserves trust when it rests on a measurable indicator and has been tested against outcomes that already happened.
That indicator exists. MIT researchers showed that a technology's improvement rate can be estimated from the patent citation network (Triulzi, Alstott & Magee, 2020). GetFocus computes it from two metrics, cycle time and knowledge flow, and has validated it on 150+ domains. Then it replays history using only the patents public at each date. In the Kodak case, the rate separated digital imaging from film in 1983, seventeen years before film sales fell off a cliff in 2000 (Kodak backtest).
The method · why the rate decides
Performance is the snapshot. The rate is the signal.
An incumbent improving 5% a year against a challenger improving 50%. Judged on today's performance, the incumbent looks safe for years.
Illustration · How the GetFocus method works
That is a forecast you can audit: the inputs are public, the maths is explainable, and the seven backtests show how it performed when nobody knew the ending. A chat answer offers none of that, however confident it sounds.
Can we connect Copilot to a patent and technology intelligence platform?
Often, yes: through an API, a connector or an MCP server, if the platform offers one and your IT team approves it. Before you wire it up, ask two questions.
Does the assistant receive the evidence, meaning patent families, citations and sources, or only the platform's summaries? A summary of a summary loses the trail back to the document. And where do your questions go? A confidential R&D question typed into an assistant is still a confidential R&D question.
Connecting a chat assistant to better data fixes coverage. It does not give the assistant a forecasting method.
FAQ
Is ChatGPT good for technology scouting?
For orientation and first drafts, yes. For decisions, no: it can't prove its search was complete, and it has no measured way to rank competing technologies.
Can ChatGPT read patents?
It reads the patents you give it and summarises them well. Reading a whole technology field means retrieving every relevant family first and judging each one, which is a pipeline, not a prompt.
Does GetFocus use generative AI?
For reading, yes: language models read every retrieved patent family in full and decide whether it belongs. The forecast is an explainable regression on two citation metrics, built on peer-reviewed MIT research (AI patent analytics).
How do you check an AI scouting result?
Open the sources behind the claims that drive the decision. If the tool can't show them, the claim isn't evidence yet.
What is the difference between technology scouting and forecasting?
Scouting builds the list of options. Forecasting tells you which one will win and when. Our scouting guide covers both.
