Research from a text message: market scans, deck builds, fact checks

A meaningful chunk of knowledge work is research that ends in a deliverable: a deck, a memo, a competitive scan, an investment thesis, a hiring panel brief, a board update. The shape is consistent: gather sources, synthesize, produce an artifact. Most knowledge workers spend 4-12 hours a week on this kind of work; for some roles it is half the job.

The agent compresses the gather-and-synthesize part. Not by replacing thinking, but by handling the mechanical labour: pulling the sources, cross-checking the facts, turning a 30-minute conversation into a 6-page deck. This post walks through what that looks like in practice, where the wins are, and what the agent does not do.

A representative research session

A founder asks the agent to scan the personal-AI-agent category before a board meeting. The session is one chat thread, ~12 minutes total.

You: research the personal AI agent category. who's
     out there, what are they raising, what's their
     positioning. board meeting tomorrow morning.
 
Ajeris: On it. Want me to focus on consumer (ChatGPT,
        Pi, Replika style) or business agents (voice
        receptionists, customer-service automation)
        or both?
 
You: both, but consumer leading.
 
Ajeris: Working. ETA 8 minutes.
 
[8 minutes later]
Ajeris: Done. Summary:
        - 14 companies tracked
        - $2.1B raised across the cohort in 2024-2026
        - 4 buckets: pure-chat, ambient/messaging-first,
          voice-receptionist, vertical-AI
        - Inflection raised the most ($1.3B); Pi product
          slowed in 2025, sold technology to Microsoft
        - The breakout name in messaging-first is unclear
 
        Want the full deck or the short version?
 
You: short version, plus the deck for tomorrow.
 
Ajeris: Short version coming as a single message.
        Deck building - I'll send the link when ready.

A few minutes later the deck is in the user's inbox. 14 slides. Cited. Each company has a one-page profile; the cohort summary is up front; the implications for the user's company are at the end. The user spends 20 minutes editing for tone and emphasis, then closes the loop.

Total time: 30 minutes for what would normally be a 4-6 hour research-and-deck-build session.

What the agent does in that 8 minutes

The mechanical work that fills most of those 4-6 hours when a human does it.

The user did none of that. The user described the goal and reviewed the output. The synthesis-and-judgment work is still theirs (which is the work that should be theirs); the rest is delegated.

What the agent reads to do this

Three sources, in order of reliability.

Web search. General-purpose web search with citation tracking. The agent prefers primary sources (company announcements, regulatory filings, vetted press) over aggregator sites.

Specific databases. For funding info, Crunchbase / Pitchbook patterns. For SEC filings, EDGAR. For patent searches, USPTO. Domain-specific databases are more reliable than general web for the specific data type.

Your own data. The agent's killer move: it can read your inbox, your meeting transcripts, your past notes, your existing decks. So "research the personal-AI-agent category" includes "what have I already noted about this from my own conversations." The synthesis is grounded in what the user already knows.

The third capability is what most differentiates from a chat-only research tool. ChatGPT does the web part. The agent adds the personal-context part.

What the agent does not do

Three things the agent will not do, by design.

It does not invent citations. If the agent could not find a primary source, it says so. The deck has a "[citation needed]" or "claim flagged for verification" rather than a fabricated quote. This is the most important honesty rule in research-grade work.

It does not replace expert synthesis. The agent's "implications" section is a draft, not a recommendation. For decisions that require deep domain knowledge, the agent's synthesis is a starting point; the expert's judgment is the deliverable.

It does not write in voice for people you have not given it. If the deck needs to be in your firm's voice, the agent uses your past decks as a style anchor. If you have not given it any past decks, the output is in the agent's house style (clean, mid-formal, no hype words).

Specific use cases that work well

A short list of common patterns.

• Competitive scans. Founders before fundraising. PMs before launches. M&A teams before approaches.
• Investment theses. Drafting an initial thesis for a category before a partner meeting. The agent does the data gathering; the partner adds judgment.
• Hiring panel briefs. Researching a candidate's prior companies, cross-referencing with current employees if any, drafting a structured panel guide.
• Conference prep. Researching the speaker line-up, identifying interesting talks, preparing question banks.
• Annual reviews. Pulling the year's key metrics from real data, drafting a "what we shipped" narrative grounded in the actual changelog.

The pattern: any task that has the shape "gather a lot, synthesize, produce a deliverable" fits. The agent handles the mechanical 70%; the user owns the synthesis 30%.

Try it

Pick the next research task on your plate that has the shape "gather, synthesize, produce." Ask the agent. Time it from "started typing" to "deliverable in inbox."