Stakeholder/landscape intelligence

Strategy leads and program designers run this before committing a budget line. The useful output is a table of positions. Each row names one organization, states what it appears to be prioritizing, cites the document that shows it, and gives the date of that document. A row without a date is worthless, because the whole point is where the field is now.

The agent must not produce a synthesis of what the field believes. That sentence is the thing everyone wants and it cannot be supported. Twenty organizations do not hold a view. They hold twenty positions, some contradictory, several stale, and a few that exist only in a press release.

This is a speed job. An analyst could assemble the same map by hand in two weeks. The agent does it in an afternoon and gets some of it wrong. At the scope stage that trade is usually worth taking.

It also has a second use. A position that nobody in the peer set is taking is a candidate for the round's own contribution. That reading only works if the coverage is honest about which peers were not found.

An unbounded sweep returns whoever writes the most. The organizations with the largest communications teams dominate the map, and the map becomes a survey of publishing habits. Name the peers explicitly. Fifteen named organizations produce a better map than an open instruction to find relevant players. This is not the same job as semantic search over a research corpus, which looks for evidence rather than for positions.

Then rank the source types by how much they cost the organization to produce. A funded call with a closing date and a budget is expensive and hard to fake. A strategy document is cheaper. A conference keynote is cheaper still. A blog post costs an afternoon. Tell the agent to collect all four but to label which is which, because the interpretation differs completely.

Set a date floor. Anything older than the last funding cycle is history, not landscape, and it will drag the map toward whatever was fashionable three years ago. Set a language scope as well, and say so in the output, because an English-only sweep of an international field is a map with known holes.

If the peer set cannot be named, the question is not ready. Go back to the brief. Peers are defined by who competes for the same researchers, the same attention, or the same problem, and that definition is a judgment somebody has to make first.

Every organization's public language is aspirational. The gap between the strategy document and the awarded grants is where the real position sits. Instruct the agent to pull both, and to put them in adjacent columns so the gap is visible rather than resolved into a single summary sentence.

Three fields do most of the work on each row. What the organization says it is prioritizing, quoted. What it has recently committed money or headcount to, cited. And when each was published. A row where the first two disagree is more informative than a row where they match, and it should be flagged rather than smoothed.

Record absence explicitly. If the sweep found nothing recent for a named peer, the cell says nothing found, with the sources checked. That is different from the peer having no position. Deep-research agents fill quiet cells with something plausible unless told not to, and a fabricated position on a peer is the error most likely to be repeated in a meeting.

Run the same sweep through more than one engine if the answer will steer real money. Agreement across engines is weak evidence, but a claim that appears in one and not the others is worth opening. That comparison is its own step, described in running one query through several research engines.

What follows is illustrative. A national research funder is deciding whether to open a call on machine learning for materials discovery, and how to position it. The strategy lead names fourteen peers: seven public funders in other countries, three philanthropic funders, two corporate research arms, and two large consortia.

The sweep returns forty-one sources across those fourteen: open calls, closed call texts with award lists, two strategy refreshes, several conference program listings, and a handful of blog posts. Six of the fourteen have a funded instrument in this area. Four have language but no instrument. Four return nothing within the date floor.

The pattern in the six funded instruments is the finding. Five of them fund model development. One funds the synthesis and characterization capacity that turns a predicted material into a tested one. Every one of the five names experimental validation as a priority in its strategy text, and none of them funds it as a line item.

That gap between stated priority and funded instrument is exactly what the adjacent-column format exists to show. The strategy lead now has a specific position available: fund the validation bottleneck that everybody says matters and nobody pays for. Whether the funder should take it is a separate conversation with different inputs.

The four peers with nothing found stay on the table as blanks. One of them is later discovered, through a colleague, to have a program in preparation that has not been announced. The blank is what made that conversation happen.

Stale sources are the main hazard. A strategy document from four years ago still ranks highly and still reads as current. If the agent cannot establish a publication date, the row does not go in the map. Undated material is the fastest route to a landscape that describes the field as it was.

Echo is the second hazard. One organization publishes a framing, three others quote it, a trade outlet writes it up, and the sweep returns five sources that look like independent agreement. Trace each claim to its origin. Count organizations, never documents, when describing how widely a position is held.

The third is silence read as absence. Organizations that do not publish in English, do not run public calls, or work under commercial confidentiality will be missing from the map regardless of what they are doing. The corporate research arms in a peer set are almost always underrepresented, and the map should say so.

The fourth is drift into competitor analysis. A landscape map at the scope stage answers one question: what are peers scoping toward. It is not an assessment of whether any of them is doing it well. That needs a different corpus, and folding it in here produces confident judgments from press material. Practitioner and registry material can deepen the picture through a targeted grey-literature sweep.

A gap in the map is not automatically an opportunity. Nobody funding a thing is sometimes a signal that everyone has already tried it and stopped. The lead's job is to work out which of those two the gap is, and the map cannot answer that. It can only make the question askable.

Verify anything that will be quoted. Before a peer's position appears on a slide in front of a board, someone opens the cited document and reads the sentence. Attributing a strategy to a named organization from an agent's paraphrase is a risk with no upside.

Date the map itself and treat it as perishable. Six months on, half the open calls have closed and two peers have published new strategies. Re-running the sweep is cheap. Quoting a nine-month-old map as current is the failure this method is supposed to prevent.

Keep the map separate from the decision it informs. It goes in the scope file with its date, its peer list, and its blanks, so that a later reader can see what was known when the call was designed. Filed that way, it stays findable in the repository that holds artifacts across rounds. What the lead concluded from it is a different document with a signature on it.