In-house or outsourced technology intelligence: keep the decision, buy the forecast

Do both, but split the work by layer, not by headcount. Keep the strategic question, the thresholds that would change your mind and the investment decision in-house. Buy what a team with general AI tools cannot do well: reading every patent in the domain every month, and measuring which competing technology is improving fastest. GetFocus computes that rate, the Technology Improvement Rate, from the patent citation network and has validated it on 150+ domains at R² ≈ 0.7. Building it yourself means implementing and validating a peer-reviewed MIT method before you get a forecast you can trust.

What should stay in-house, and what should you buy?

The standard advice is a hybrid model, and it is right. What it rarely says is where the line goes. Draw it by what each layer needs: your context, or reading capacity and a validated method.

The reading load is the part teams underestimate. The GetFocus battery radar holds 31,644 patent families (August 2026). A team can sample that. It cannot read every family in full each month and keep the reading consistent enough to compare one route with another.

The judgement layer never moves outside. "The actual decision to invest remains with the head of R&D," says Christophe Perthuisot, former SVP of R&D at Moët Hennessy (read his story).

Why can't an in-house team with AI tools do the forecasting?

General AI tools read and summarise, and your team should use them for first passes over papers and patents. What they do not produce is a rate: a comparable number for how fast each competing technology improves, checked against history.

That number took years of research to establish. MIT researchers analysed performance data for 28 technological domains and found that every one improves exponentially, each at its own rate. Triulzi, Alstott and Magee (2020) then showed the rates can be estimated from patent data. GetFocus automated the method and validated it on 150+ domains: the forecast explains about 70% of the variation in measured improvement rates.

The patent set matters as much as the maths. Get it wrong and every number downstream is wrong. GetFocus retrieves candidates by meaning, then has a language model read every family in full, from title to description, and rule it in or out. The agents do the reading; the maths does the forecasting.

And the output has to beat the obvious signal. In the GetFocus mobility backtest, hydrogen out-filed every other technology for 25 years, while NMC batteries measured 54%/yr against hydrogen's 21% from the first data point in 2001. A summary of the filings pointed the wrong way for a quarter of a century.

Should you buy a tool and run it yourself, or pay analysts?

The usual rule says buy a tool for recurring work you can staff, and pay analysts for high-stakes questions that come up now and then. For technology bets, both halves break.

A tool you run yourself still needs someone to run it, and R&D leaders get 3% of their time for the bets that decide the next five years. Analysts bring judgement, but a study is a snapshot: six months and six figures, stale the day it lands, while the patent record changes every month.

The third option is the one the either/or leaves out. In a GetFocus domain, the agents scout and read, the rate is recomputed as new patents publish, trigger alerts fire when a threshold you set is crossed, and every quarter an analyst walks your team through what changed. Your people keep the questions and the calls. The AI patent analytics page sets out how forecasting differs from search and landscape work.

What does each option cost beyond the fee?

Compare total cost, not salaries against an invoice. In-house means hiring, training, tool and data licences, management time, and the coverage gaps a small team leaves. Outside means the fee, the briefing time, and the risk that findings never reach a decision.

Then add the line most comparisons leave out: the cost of backing the wrong route. The GetFocus backtests price it at $20B+ for the hydrogen powertrain, and at $19B for the storage incumbent that paid to catch up on flash in 2015, 14 years after the rate flagged it.

The input on the GetFocus side is small: one strategic question in plain language, a sign-off on the list of competing routes, and the quarterly review.

Should technology intelligence sit with a central team or every R&D engineer?

Both, with different jobs. A small central owner per domain sets the questions, the thresholds and the review rhythm. Access to the evidence stays open to the engineers who know each technology best.

In a GetFocus domain that means five platform seats for your experts and an evidence table they can open: every question a column, every patent family a row, every claim traceable to the families behind it. Experts question the agents between reviews instead of waiting for the next report.

Open access matters most when the data contradicts the house view. Perthuisot counts that as the point: "You want the system to challenge existing thinking and the prevailing paradigms within the company." Start with the experts who want to try it. When his teams tested the system, "They said it saved them months of work."

How do you test the choice before you commit?

Test on a decision you already face, not a demo dataset. Agree up front what counts as useful: a route your team had not considered, a verdict that contradicts the house view, a trigger worth watching.

The GetFocus Decision Brief is built for that test: one live technology decision, a first-pass radar, an improvement rate for each competing technology, the monitoring triggers we would set and a 30-minute readout with an analyst, in 48 hours and free. You keep the brief either way. If the decision is too sensitive to share, we walk through an example built on your business context instead.

Or test the method on history first. Pick a disruption that undeniably happened in your industry and check how early the rate called it: the seven backtests use only patents public at each date. For why those rates hold, see why technology improvement is predictable.

This split fits any R&D organisation with a technology bet of real consequence on its roadmap. For prior-art searches, freedom-to-operate work or a one-off literature review, an IP firm or a patent search tool is the better buy.

FAQ

Can our team do technology forecasting with ChatGPT or Copilot?

They are good at reading and summarising, so use them for first passes. They do not measure how fast each competing technology improves, and their summaries need checking against the patents. A forecast needs a validated method and a complete patent set.

How many people does an in-house technology intelligence function need?

Fewer than most plans assume once the reading and the rate are bought. What you cannot outsource is one accountable owner per decision domain: the person who sets the questions, agrees the thresholds and acts on the verdicts.

When is an outside service the better choice?

When the decision is high-stakes, when you need an independent challenge to the house view, or when it cannot wait for a team to be hired. If the decision stays open for years, buy the monitoring too: a one-off answer goes stale as new patents publish.

What do we have to share with an outside provider?

With GetFocus, one strategic question in plain language and a sign-off on the list of competing routes. The rates come from public patent data. For a sensitive decision, we work from an example built on your business context.