Research AI
Enablement Plan

Two parallel tracks — tools & fluency and the Signal Stack — that move research from individual experimentation to shared, trustworthy systems.

Champion: Mike Lytle April 15 – June 30, 2026 Last updated: April 17

Situation Assessment

Every researcher is using AI. Several have pushed into Builder territory. What we lack is standardization, shared systems, and a way for experiments to compound.

94.1%
Design dept adoption rate
32
Active users (ChatGPT + Gemini)
28.1%
Using Projects (persistent context)
50%
Messages from top 5 power users

What the adoption data tells us

Note: This data covers ChatGPT/Gemini only, not Claude Code, Cowork, or skills. An AI Pulse Survey (week of April 27) will give a fuller picture.

Prototype-First Pilot

Ending April 17. Setup friction → collaboration friction → handoff gaps. AI comfort converged to 4/5 by Week 3. Effectiveness tracked to process maturity, not tool fluency.

Signal Stack Session

April 8 team working session. Seven components proposed. Interest gathered, healthy skepticism expressed — especially around synthetic users and governance.

AI Champions Program

Research, product design, and content design developing separate-but-laddered enablement plans. The May 6 design offsite is the shared milestone.

What Thumbtack Expects

AI-first in consequential work. "Match rigor to risk." Stay at the craft frontier. Share openly — individual experiments must become team knowledge.


Strategic Frame

Two parallel tracks that reinforce each other. Track 1 creates the conditions for Track 2. Track 2 gives Track 1 a purpose beyond individual fluency.

Track 1 — Tools & Fluency

Get every researcher to a comfortable Operator baseline, with Builders supported to go deeper. Tool access, skillshares, dedicated build time, and a path to the May 6 offsite.

Track 2 — Signal Stack

Build the shared systems that define how research uses AI at scale. Evaluation frameworks, behavioral systems, synthetic user investigation, knowledge infrastructure, and the operating model.


Track 1: Tools & Fluency

Every researcher confident and capable with AI tools. Shared workflows for common tasks. Nobody feeling pressured — just enabled.

Builders

Give them visibility, surface their work to leadership, connect experiments to the Signal Stack, protect their time.

Operators

Help them standardize one pattern. Granola recipe + Claude Project is a good starting point. Encourage sharing.

Explorers

One real task under deadline pressure. Pair with an Operator buddy. The offsite is a natural forcing function.

ToolPurposeAccess
Claude EnterpriseResearch Projects with standing context, Ask ThumbtackPending enterprise contract (mid-April)
Claude Code / CoworkSkills, automation, behavioral analysis pipelinesIT request template in #research-team-all
Granola + RecipesStructured meeting outputs — "User Interview" and "Research Readout" recipesLaunched; private team folder set up
NotebookLMCross-corpus querying — "Ask Research" prototypeAvailable via Okta
Prototyping PlaygroundResearch stimulus, prototype testing with usersBroader access planned for offsite
ChatGPT EnterpriseGeneral-purpose AI assistant, GPTs, Projects for persistent contextAvailable via Okta
Gemini + GemsCustom Gems for synthetic users, structured workflows, multimodal analysisAvailable via Okta
Gemini AI StudioAdvanced prompting, multimodal analysis, prototype evaluationAvailable via Okta

Skillshare Series

Three cross-functional Wednesday sessions (2–3 PM PT) building toward the offsite. Plus research-specific demos during team meeting time.

April 15

"How I Built This"

Content design GPT + Figma plugin walkthrough. Researchers observe how tools get built; consider what research-specific tools we'd want.

April 22

Skills, GPTs, Gems & MCPs

Customizing your AI toolset. Researchers bring one workflow to automate; leave with a working skill or GPT draft.

April 29

Prototyping Playground

Hands-on build session. Researchers learn to use prototypes as research stimulus; preview offsite format.

Potential Research-Specific Additions

During existing team meeting time or async:


Design Offsite · May 6

Morning session (10:00–12:30, 2.5 hours). Full Cap D org (~30 people). Format: kickoff → show & tell → problem-first sprints → demo share-out.

🔧 Build a Stack Component

Researchers self-select into one Signal Stack area — a research repository skill, a concept evaluation prompt, a behavioral signal prototype — and build a working first draft.

🔀 Cross-Pollinate

Some researchers join cross-functional groups to build tools that bridge research ↔ design ↔ content: a discovery kickoff skill, a prototype feedback tool.

📦 Walk Away with Something

Every researcher leaves with at least one artifact they can use the following Monday — a skill, a prompt template, a research site, or a prototype.

Pre-work (before May 4)

Show & tell candidates: Synthetic user gem + research site · Behavioral conversation audit at scale · Human evaluation framework for AI-generated content


Track 2: Signal Stack

A system of interconnected tools and processes that define how research generates, evaluates, and scales insight in an AI-augmented world.

Note: This is an early snapshot. Components, sequencing, and ownership will evolve as the team works through the strategy and starts building.

Foundation

Evaluation Frameworks

Define how we evaluate outputs and experiences. Shared criteria: clarity, relevance, trust, usefulness.

Low Risk · First Pass

AI Evaluation

Structured feedback on outputs and prototypes — first pass. Not for replacing judgment or defining product direction.

Low Risk · First Pass

Synthetic User System

Predicts how users might respond. Fast, early evaluation before building. Team may decide not to pursue — valid outcome.

Higher Risk · High Value

Behavioral System

Shows what actually happens in use. Paths, drop-offs, hesitation, breakdowns. Ground-truth signal from real behavior.

High Risk · High Value

Research Acceleration

Speeds up synthesis, study design, and pattern identification. Reduces operational overhead. Not for replacing research.

Foundational

Knowledge System

Makes past research accessible, contextualized, reusable. Captures recurring behaviors and connects research to decisions.

Governance

Operating Model

Ensures systems are used and maintained correctly. Confidence levels, escalation paths, when to trust vs. when to verify.

Design Principles

1

Trustworthy systems

Every component must earn the team's belief.

2

Calibration loops

AI predicts → behavior shows → gaps improve the system.

3

Low-risk first

First-pass tools screen and narrow. They don't decide.

4

Transparent confidence

Systems show confidence level and suggest next steps.

5

Governance from day one

Don't build what we won't maintain and calibrate.

6

Shared ownership

No one builds something the team doesn't believe in.

Working Model: Roles & Ownership

Connected roles that drive the Signal Stack forward:


Action Plan

April 15 – June 30, 2026. Concrete actions organized into four phases.

Weeks 1–2: Foundation

April 15–28
Tool access: Confirm Claude Code/Cowork for all researchers via IT request template.
Survey: Send workflow/pain point survey to all Cap D functions by April 16.
Knowledge System: Connect with Research Ops PM on scope and operationalization.
Skillshare #1 (April 15): "How I Built This" — capture research-relevant ideas.
Share team experiments in #research-team-all: "This is how we make experiments compound."
Skillshare #2 (April 22): Skills/GPTs/MCPs — bring one workflow to automate.
Team demo (week of April 21): Behavioral analysis pipeline + research repository prototype.
Evaluation Frameworks session (60 min, week of April 28): Articulate shared criteria.
Research Acceleration: Identify 3 high-value patterns by April 25.

Weeks 3–4: Offsite Prep + Offsite

April 29 – May 9
Skillshare #3 (April 29): Prototyping Playground hands-on session.
Pilot "Ask Research" Knowledge System with one team.
Playground access confirmed for all researchers before May 4.
Curate offsite project list from survey results + AI Champions brainstorms.
Design offsite (May 6): Build first-pass Signal Stack components. Show & tell. Walk away with something usable.
Capture outputs in shared repository. Identify what to keep building.

Weeks 5–7: Signal Stack Phase 1

May 10–30
Pilot playbook: Research sections on user signal, trustworthy output, "match rigor to risk."
Knowledge System debrief: Expand, iterate, or re-scope?
Research Acceleration quick wins: Standardize and share identified patterns.
Behavioral System scope: Connect with DS/Eng on data access. Path from request-based to direct exploration.

Weeks 8–11: Phase 2 + Scale

June 1–30
AI Evaluation pilot: Structured heuristic evaluation for one product area.
Synthetic User investigation: Parallel study — synthetic vs. real research. Team decides together.
Operating Model v1: When to use each system, confidence thresholds, escalation paths.
Functional enablement check-in: What's working, what's blocked, what we need.
Q2 retrospective: What did we build? What do we believe in? What gets formalized? What gets stopped?

Team Activation

How we sustain momentum without adding more meetings or pressure.

📢

Weekly · 5 min

Rotating share in team meeting: tried → happened → learned. No polish.

🛠️

Monthly · 30 min

Show, Don't Tell. One person goes deep on a build. Actual tool, actual output.

📊

Biweekly · 15 min

Signal Stack check-ins. Component leads: where we are, what's next, blockers.

🔍

Quarterly

Use case review. What's worth formalizing? What goes in the playbook? What do we stop?

Responding to Common Concerns

"AI is changing my role."
The researchers who define what "grounded and trustworthy" means for AI-assisted discovery own the most important layer. That question is yours to answer.
"I don't have time for this."
Don't add a learning block. Find work already happening under deadline pressure and try AI there. The work is the learning.
"I don't trust the output."
This is research's core instinct and it's right. "Match rigor to risk." AI for exploration is fine; AI as sole source for high-stakes decisions is not.
"The vibes feel competitive."
Name it. The Signal Stack is explicitly designed to be shared. "This is not about who owns what. It's about what we want to build together."
"Is this safe for participants?"
Don't guess. Complete AI in the Workplace Training. Route PII or behavioral data through IT/Legal. Escalation: Champion → Katie/Nicole → IT/Legal.

Research's Strategic Position

Research isn't just a participant in AI-assisted product development. Research is the function that determines whether it produces trustworthy outputs.

The Prototype-First Pilot's core open questions — "Where does quality degrade? What guardrails are required?" — are Research questions. The Week 3 finding that prototypes expose product gaps that mocks wouldn't have surfaced is a Research finding. The async feedback gap is a Research problem. The "match rigor to risk" framework is becoming the operating principle for the whole org.

The Signal Stack — if developed with the rigor and skepticism this team brings — becomes the continuous learning loop between people and AI systems that the company needs.

Make sure leadership knows what's being built here. Make sure it's in the playbook. And make sure the team believes in it — because systems no one trusts are worse than no systems at all.