Grounding Sync's Category Taxonomy in How Personal AI Agents Are Actually Used¶

TL;DR¶

Real-world personal-agent usage clusters around task type and life-domain, not "decision-making domains." The single largest signal across the NBER/OpenAI ChatGPT study (Chatterji, Cunningham, Deming, Hitzig, Ong, Shan & Wadman, "How People Use ChatGPT," NBER Working Paper No. 34255, Sept 15, 2025, classifying ~1.1M de-identified messages collected May 2024–July 2025), Anthropic's Economic Index and Clio analysis, Menlo Ventures' 5,031-person Morning Consult survey of U.S. adults (April 2025; published June 26, 2025), the GPT Store, Lindy/Manus templates, and Anthropic's official Skills examples is that people describe and organize agents by what the agent does (write, research, schedule, code, summarize, plan) or what part of life it serves (work, learning, health, money, family logistics, creative), not by abstract decision archetypes.
Of Jonathan's five proposed categories (Money, People, Strategy, Dilemmas, Voice), only two map cleanly to observed usage: Money and Voice. "People" partially maps (relationship/coaching conversations exist but are ~2% of ChatGPT volume per NBER). "Strategy" and "Dilemmas" do not show up as natural user-described categories anywhere in the data — they're a researcher's projection. Recommendation: drop or merge Strategy and Dilemmas, and consider adding Health, Learning, Work/Career, and Logistics if the goal is to match how people actually segment their lives.
Users overwhelmingly want one generalist agent that behaves differently per context, not a fleet of specialized agents. 91% of consumers use a general assistant for almost every task; 60% reach for specialists only when the general tool fails (Menlo, 2025 State of Consumer AI). The implication for Sync: category-level scores are defensible as a profiling lens that modulates one agent's behavior, not as a way to spawn separate agents. And if exposed, the categories must match how users already segment their lives or they'll feel arbitrary.

Key Findings¶

1. Three datasets agree on the dominant categories — and they aren't "decision-making" categories¶

The three most rigorous, large-N studies of consumer AI usage converge on a strikingly similar picture:

OpenAI / NBER "How People Use ChatGPT" (Chatterji et al., NBER WP 34255, Sept 15, 2025; ~1.1M classified conversations May 2024–July 2025): - Practical Guidance — 28.1% - Writing — 28.3% - Seeking Information — 21.3% - Technical Help (incl. coding) — 7.5% - Multimedia — 6% - Self-expression (relationships, reflection, roleplay) — 4.3% - Other — 4.6%

Note: Computer programming and emotional companionship are small categories on the consumer surface (programming runs through APIs/Claude Code instead). Relationships and personal reflection = 1.9% of ChatGPT messages. The authors also propose a cross-cutting frame: Asking (49%) / Doing (40%) / Expressing (11%).

Anthropic Economic Index (Nov 2025 / Feb 2026 reports): - Computer & Mathematical (coding) — ~35% of Claude.ai conversations, ~44% of API traffic - Editing/improving written content — top consumer task - Educational tasks — rising from 9.3% → 12.4% over 2025 - Scientific/analytical tasks — ~7% - Business strategy/operations — ~6% - Per the Anthropic Economic Index (Feb 2026 edition): "Claude estimates that software development requests would take a competent professional approximately 3.3 hours to complete without AI…Personal life management tasks are estimated to be simpler, averaging 1.8 hours."

Anthropic's Clio (2024, analysis of 1M conversations) found top categories: coding/software dev (>10%), educational use (>7%), business strategy/ops (~6%), with a very long tail.

Menlo Ventures 2025 State of Consumer AI (Morning Consult, 5,031 U.S. adults, April 2025; published June 26, 2025): Bottom-up survey of how people spend their time × where they apply AI — five categories emerged: 1. Routine Tasks (highest AI usage) 2. Creative Expression 3. Learning and Development 4. Physical and Mental Health (under-served — per Menlo: "only 21% of the 41% of U.S. adults who sought mental health support in the past six months turned to AI — only 9% of the total population") 5. Connection (relationships, social, dating)

Notably absent from Menlo's bottom-up framing: anything that looks like "Strategy" or "Dilemmas" as a free-standing category. Money/personal-finance shows up under Routine Tasks and is called out as one of the most under-served categories.

2. The platform-side taxonomies tell the same story¶

GPT Store official categories (8): Top Picks, DALL-E, Writing, Productivity, Research & Analysis, Programming, Education, Lifestyle. Over 159,000 public GPTs, 3M+ total created.
Anthropic Skills repo organizes official skills into: Document Skills (PDF/DOCX/PPTX/XLSX), Creative & Design, Development & Technical, Enterprise & Communication.
Lindy templates group by department/task: Sales, Customer Support, Email, Meetings, Recruiting, CRM Updates, Research.
Manus's most-cited use cases: travel planning, financial/stock analysis, market/competitive research, resume screening, slide-deck creation, website generation, data dashboards.

Across every platform, the organizing axis is task type (sometimes blended with department or content modality), not domain-of-judgment.

3. Top 8–12 observed use-case clusters, ordered by apparent frequency¶

Synthesizing the NBER, Anthropic, Menlo, GPT-Store, and Lindy/Manus signals:

Writing & editing (drafting, summarizing, translating, polishing) — the #1 work use case (~40% of work-related ChatGPT messages per NBER); about two-thirds of "Writing" volume is modifying user-supplied text, not generating from scratch.
Practical guidance / how-to — tutoring, customized advice, ideation; ~28% of all ChatGPT volume; this includes a lot of life-management micro-decisions (workouts, recipes, budgeting, parenting tips).
Information seeking / research / fact lookup — direct substitute for web search; ~21% of ChatGPT; deep-research products (Perplexity, ChatGPT Deep Research, Claude Research) are a fast-growing prosumer category.
Coding & technical work — small on consumer ChatGPT (~5–7%) but dominant on API/Claude Code (~35–44%); per Anthropic's Feb 2026 Series G announcement, "Claude Code's run-rate revenue has grown to over $2.5 billion; this figure has more than doubled since the beginning of 2026."
Learning & education — exam prep, tutoring, language learning, course design; 12.4% of Claude.ai usage; one of fastest-growing GPT Store categories.
Document creation (slides, spreadsheets, formatted docs, PDFs) — Anthropic's official Skills repo is heavily oriented here; Excel formulas, PPT decks, Word docs are top "specialized" GPTs.
Creative expression (image generation, video editing, music, design, brand assets) — Menlo: "Tools that enable professional-grade results (think Canva for design) capture nearly half of all spending on specialized AI tools (45%)" — i.e., the creative-tools category broadly captures ~45% of specialized-tool spend; Canva, CapCut, Midjourney, Nano Banana, Veo dominate this surface.
Email, scheduling & inbox management — top Lindy/agent-platform template category; also a leading "I wish my agent could…" unmet need on consumer side.
Personal life logistics (meal planning, travel, shopping, family/childcare coordination) — explicitly top use case for Manus's consumer demos and Menlo's "parent" super-user segment (Menlo: 34% of parents use AI for managing childcare).
Sales, marketing & CRM ops — dominant departmental B2B use; bulk of Lindy/Relevance/Manus templates.
Personal finance / money — flagged by Menlo as one of the most under-served personal categories; only a small share of finance-help seekers currently use AI; appears as a domain in user-built Claude Projects and finance-themed custom GPTs.
Emotional support, coaching, companionship — 1.9% of ChatGPT messages (Relationships and Personal Reflection topic per NBER); small in volume but disproportionately present in headlines; Menlo: only 21% of mental-health seekers used AI.

4. What users say they want but can't yet do (5–8 unmet-need clusters)¶

End-to-end personal logistics — Menlo's headline example (Shawn Carolan, June 2025): "Imagine a parent saying, 'Plan summer for my kids,' and the AI takes care of everything: researching camps, juggling schedules, buying gear, coordinating carpools." The complaint pattern is "AI can advise but can't act."
Trusted financial and health advice — high-frequency, high-friction, high-trust domains where general assistants are felt to be inadequate. Per Amy Wu Martin of Menlo: "The best opportunities are in messy, recurring problems where general tools break down. Think deeply personal responsibilities like managing money, healthcare, or coordinating logistics across the entire family."
Memory and continuity — recurring frustration in user reviews of coding agents and Lindy: "AI coding agents are great at remembering what you just told them… but they're not great at remembering everything you've ever told them" (Smiansh, "The Real Struggle with AI Coding Agents"). Power users explicitly cite memory as a top blocker.
Tool/action capability beyond chat — recurring "I want it to actually do the thing": book the calendar, send the email, file the form. SMS-fronted agents (Poke; horizontal text-first agents like Manus and Genspark) are a direct response to this.
Personalized voice/style matching — J.D. Hodges' custom-instructions template: "I'm a [type] writer. My voice is [conversational, formal, technical]… don't rewrite my voice out of the piece." Common complaint that defaults sound generic.
Judgment / pushback / challenge — small but visible cluster of users explicitly building "Change My Mind"-style agents (a top GPT Store custom GPT "designed to provide concise counter-arguments… challenge the user's opinions analytically, identifying biases, and presenting alternative perspectives") or instructing Claude/Projects to "flag anything that looks wrong, even if I didn't ask" and "challenge my reasoning."
Graceful escalation and reliability — repeated theme in agent failure-mode literature: users bounce off agents that hallucinate confidently, can't say "I don't know," or can't hand off to a human.
Integration with the user's actual stack — Gmail/Calendar/Notion/HubSpot are constantly cited; Olivia Moore of a16z (Top 100 Gen AI Consumer Apps, 6th edition, March 2026): "Context compounds: the more an LLM knows about you, the better results it can provide and the more you use it" — i.e., agents that know you will displace those that don't.

5. How users describe specialization — the critical product question¶

From a targeted pass on Reddit / Twitter / GPT Store names / Substack posts (n=~16 distinct user-shared agent configurations):

Axis users describe specialization on	Approx. frequency	Examples
Task type	~50%	"Email GPT," "Pitch Panda" coach, "Write For Me," "Scholar GPT," "Resume GPT," "research agent," "summarizer," "scheduling agent"
Domain	~19%	Cat-CKD vet-advice project, vegetarian-diabetic recipes, "finance/money," stock-analyst GPT
Persona/style	~19%	"Be my coach" (Dan Shipper, Every), Socratic tutor (Ethan Mollick), "Change My Mind," "blunt counter-arguer," "match my voice"
Tool integration	~6% (rarely headline)	Almost always paired with task: "agent that monitors my Gmail and notifies me on Slack"
Mixed / multi-axis	~30% of power-user setups	Task + Persona + values is the dominant pattern among Dan Shipper, Ethan Mollick, J.D. Hodges

Pattern: Names skew Task. Personas/style sit inside the prompt, not in the name. Domain framing is more common in personal/life contexts (health, money, pets, recipes; e.g., Megan Ellis in XDA describes her Claude Projects: "a project to get advice on what to ask my vet with regards to my cat's chronic kidney disease, a project for vegetarian recipes that will also be suitable for a person with diabetes, and my daily schedule project"). Power users describe multi-dimensionally; casual GPT Store creators stay single-axis (Task).

Decision-making / judgment-style agents do exist but are described in persona-of-an-advisor terms ("coach," "advisor," "counter-arguer," "challenge my reasoning") much more than in domain terms. Dan Shipper's example: "I have a similar GPT that I created to be my coach. I give a long prompt, including my goals and current situation, and then ask for advice. It's a great listener — arguably more patient than my spouse." This is an important signal for Sync — see Hypothesis Test below.

6. Generalist vs. specialist preference¶

The data is unambiguous: - 91% of consumers use a general AI assistant for almost every task (Menlo, 2025 State of Consumer AI). - 60% have tried specialized tools — but typically only when the general tool fails. - People stick with general assistants because of convenience and familiarity — specialized tools "add friction" (sign-up, learning new interfaces). - Specialists win where general assistants fall short: creative work (creative tools capture 45% of consumer specialized-tool spend per Menlo), high-trust personal domains (money, health), and high-skill professional tasks (coding via Claude Code, Cursor).

Implication for Sync: Users do not want a fleet of specialized agents to manage. They want one agent that gets smarter about them. Category-level scores should therefore be framed as facets of a single profile, not as separate agents — and exposed only if doing so adds explanatory value the user actually wants to see.

Natural category framings that emerge from the data (not from any hypothesis)¶

These are the four taxonomies that actually emerge from observed usage, in descending order of how well they explain the variation:

Framing A — Task type (the dominant frame across all platforms)¶

Categories: Write · Research · Plan/Schedule · Code · Create (image/video/design) · Summarize/Analyze · Communicate (email/messages) · Decide/Advise Rationale: This is the axis the GPT Store, Anthropic Skills, Lindy, and Manus all default to, and the axis users most frequently use when naming their own custom GPTs/Projects. It maps cleanly onto NBER's classifier output. Weakness for Sync: most of these are execution categories with no obvious "decision profile" overlay.

Framing B — Life-domain (Menlo's bottom-up frame)¶

Categories: Routine Tasks · Work/Career · Learning · Health · Money · Connection/Relationships · Creative Expression Rationale: Closest to how users actually segment their lives. Strongest fit for a product whose value proposition is "represent me in delegated decisions." Money, Health, and Connection are exactly the high-friction, high-trust categories where general AI fails users today.

Framing C — Interaction mode (NBER's Asking/Doing/Expressing; Anthropic's augmentation/automation)¶

Categories: Asking (49%; advice, learning, decision support) / Doing (40%; task execution) / Expressing (11%; reflection, play, social) Rationale: Cuts across topic; reveals that ~half of consumer AI use is decision-support and advice — a real finding for Sync. But it's too coarse to specialize on.

Framing D — Augmentation vs. automation (collaboration mode)¶

Categories: Directive (just do it) / Task iteration (work with me) / Learning (teach me) Rationale: Anthropic's primary primitive in the Economic Index. Useful as a behavioral-style overlay for Sync (does the user want the agent to decide or to consult?) but not a category taxonomy in the topical sense.

Hypothesis test: How do Sync's proposed five categories hold up?¶

The proposed taxonomy was: Money, People, Strategy, Dilemmas (high-stakes/ethical), Voice (style/tone).

Category	Appears in observed usage?	How frequent?	Verdict
Money	Yes — clearly. Menlo flags personal finance as a top under-served domain. Custom GPTs like GiPiTi finance ("your personal financial advisor"), stock-analyst GPTs, and Claude Projects for budgeting are real and recurring.	Modest in volume today (most finance help is generic), but high in latent demand.	Keep. Strong product-market fit signal; this is exactly the kind of high-friction, high-trust domain where users would value a personalized profile.
People	Partially. Relationships/personal-reflection = 1.9% of ChatGPT (NBER), ~2% of Claude affective use. Coaching, counseling, and interpersonal-advice clusters exist but are small.	Small share of volume; outsized share of sensitivity.	Keep but rename. "People" is ambiguous (does it mean managing relationships, doing people-related work like hiring/team management, or social calendaring?). Consider "Relationships & Coaching" or split into "Relationships" (personal) and "People-management" (work).
Strategy	No — does not emerge as a natural category. It appears as a task ("help me think through a strategy") inside Practical Guidance or as a style ("be analytical"), but no platform, study, or user-built GPT names "strategy" as a top-level category. Closest analogs are "Business strategy & operations" (~6% of Claude) and "Practical Guidance" generally.	Effectively zero as a user-described category.	Cut, or merge into a broader "Work/Career" or "Decisions & Advice" category. The notion of a "strategy profile" feels like a researcher's projection, not a user's mental model.
Dilemmas (high-stakes/ethical)	No — does not appear at all as a category. Anthropic's "Values in the Wild" research touches on ethical reasoning, and users do bring Claude/ChatGPT hard personal questions, but no observed taxonomy treats ethical dilemmas as a distinct category. They appear as edge-cases inside Practical Guidance, Coaching, or Counseling clusters.	Effectively zero as a user-described category.	Cut. This is the weakest category in the proposed set. Users don't think "I need my dilemma-agent." If Sync wants to capture this signal, it's a style overlay ("how does this person weigh tradeoffs?") rather than a category.
Voice (style/tone)	Yes — but it is uniformly described as a style overlay, not a category. Power users (Hodges, Shipper, Mollick) all explicitly tune voice/persona inside their custom instructions. It's the #1 cross-cutting customization users actually apply.	Very high frequency as a modifier; ~zero as a topical category.	Keep — but reframe as a cross-cutting style dimension, not a peer of Money/People. Voice is to a decision profile what tone-of-voice is to a brand — it modulates everything. Mixing it as a peer category will confuse users.

What the data suggests Sync should do¶

Kill: Strategy and Dilemmas as user-facing categories. They don't match how anyone — neither big-study data nor power-user custom prompts — talks about their AI use.

Reframe: Voice as a cross-cutting style/persona layer that applies across all topical categories, not as one of five peers.

Keep and sharpen: Money (strong fit). People (rename and disambiguate — Relationships vs. People-management).

Strongly consider adding (if domain-level categorization is the goal): - Work / Career — by far the largest single bucket of high-stakes decisions in the daily-life-of-a-knowledge-worker frame. - Health — Menlo's most under-served high-friction domain; a natural fit for a profile-driven assistant that knows your constraints. - Learning — explicitly growing in Anthropic's data (9.3% → 12.4%); high engagement. - Logistics / Family / Routine — Menlo's #1 actual usage category; "Plan summer for my kids" is the canonical aspirational use case.

A defensible 5-category replacement informed by the data might be: Money · Health · Work/Career · Relationships · Logistics & Routine — with Voice/Persona as an orthogonal style dimension that overlays all five. This taxonomy maps directly onto Menlo's bottom-up consumer survey and onto the way Claude Projects and custom GPTs are actually configured by real users.

The deeper challenge: do users want this exposed at all?¶

The strongest finding in the data — and the one most important for Sync's product question — is this: users do not currently describe wanting category-level decision-making scores. They describe wanting: 1. An agent that remembers them (memory/continuity). 2. An agent that acts (capability/integration). 3. An agent that sounds like them (voice/persona). 4. An agent they can trust on high-stakes personal stuff (money, health).

None of those is naturally framed as a "category score." The risk for Sync is exposing axes that feel like a quiz output to a user who actually just wants an agent that gets them. A better mental model from the data: the categories are how the profile is built and audited (you played money scenarios, here's what we learned about your risk tolerance); the agent is one entity that behaves accordingly.

Recommendations (staged, with thresholds)¶

Stage 1 — Before shipping any category-level scores publicly: 1. Cut Strategy and Dilemmas. They are not in the data. If Jonathan needs to capture "high-stakes thinking" signal, treat it as a complexity/stakes overlay on every category, not a category. 2. Move Voice out of the category set and into a separate "Style" or "Persona" layer that modulates outputs across categories. This matches how every power user actually configures their agents. 3. Replace the cut categories with at least 2 of: Work/Career, Health, Logistics, Learning. Match Menlo's life-domain frame. Threshold for picking which ones to keep: do the scenarios in Sync's library generate meaningfully different behavioral signal across these domains, or are they redundant? Run an internal validation pass.

Stage 2 — Test the framing with users: 4. A/B the framing. Show one cohort a single "decision profile score" and another cohort the category-level scores. Measure: which cohort says the agent "feels like me" after a week? Which cohort actually uses the agent on a category they specialized in? Threshold to ship category scores: ≥10pp lift in personalization perception or ≥20% lift in delegated-task acceptance. If category scores don't move those metrics, kill them and ship the single score. 5. Frame categories as audit/explanation, not as agent-specialization. "Here's what we learned about how you handle money" reads very differently from "Here is your Money Agent." The first is grounded; the second projects a mental model users haven't adopted.

Stage 3 — If category scores survive testing: 6. Anchor categories to life-domain (Framing B), not to task-type or judgment-mode. Sync's value prop is decision-support and delegation. Life-domain is where users actually compartmentalize trust. 7. Add a cross-cutting style/voice dimension (formality, bluntness, risk tolerance, deliberation speed). This is what power users tune for, and it's orthogonal to category. 8. Watch for the memory/integration gap. Even a great profile is worthless if the agent can't act on it. Make sure the profile feeds an agent that has tool access where users are bouncing off today (email, calendar, finance apps).

Caveats and confidence notes¶

The NBER ChatGPT study is the single highest-confidence dataset (~1.1M classified conversations, OpenAI's own data, peer-reviewed pipeline using Anthropic's Clio-style methodology). Anthropic's Economic Index and Clio are close seconds. Menlo's consumer survey (5,031 U.S. adults, Morning Consult, April 2025) is the strongest life-domain signal but is U.S.-only and self-reported.
Sampling bias to flag: GPT Store top lists skew toward novelty and SEO-discoverable names; Twitter/Reddit signal skews tech-savvy early adopters; Anthropic's data skews developer-heavy; ChatGPT data is the most representative of mainstream consumer use.
The "agent platform" data (Lindy, Manus, Devin, Replit Agent) is heavily B2B-flavored. Their template galleries tell you what sells in workflow automation, not necessarily what an individual consumer wants from a personal decision assistant. Sync should weight Menlo and NBER more heavily for its consumer thesis.
Inference points: The frequency tallies in the "how users describe specialization" section come from a curated sample of ~16 user-shared agent configurations, not a representative sample. The directional pattern (Task ≫ Domain ≈ Persona ≫ Tool) is robust enough to act on, but the exact percentages are illustrative.
What the data does not tell us: Whether users would value a decision-profile-driven agent if presented well. No platform has shipped this yet at scale, so usage data can only tell us what categories users currently think in — not what categories they would find useful if introduced. This is exactly why Stage 2's A/B test is the critical gate.
One tailwind worth flagging: the "Asking" / "Doing" / "Expressing" frame from NBER suggests that ~49% of consumer AI use is already decision/advice-flavored ("Asking"). This is a real and underappreciated tailwind for Sync's core thesis — but it shows up in the data as a cross-cutting mode, not as a domain. The strategic move is to position Sync as the "Asking-mode personalization layer" for whatever life-domain a user is operating in, rather than as a set of distinct decision-agents.