Nathan Weill here.
I'm the CEO @ Flow Digital, where we help companies unleash their full potential by strategically automating every inch of their workflows. Each week, I share hot AI and automation tips to help you move your business into the future successfully.
Claude Design
Create Landing Pages That Actually Match Your Website
Marketing teams lose days every time they launch a new campaign page, fighting to match brand colors, fonts, and layout patterns from the main site. The result is a graveyard of one-off landing pages that look almost-but-not-quite like the rest of the brand.
Claude Design, launched by Anthropic Labs in April, fixes this by reading your existing website once and applying your brand to every page it generates afterward.
Key Features:
Web capture tool: Grab elements directly from your live site so new prototypes look like the real product, not a generic template.
Persistent design system: Claude reads your codebase, Figma files, or screenshots during onboarding and extracts colors, typography, spacing, and components. Every project after that inherits them automatically.
Conversational refinement: Edit text inline, comment on specific elements, or use adjustment controls to tweak spacing and color without touching code.
Business Impact:
Brilliant's product team reported going from rough idea to working prototype before anyone leaves the room, with output staying true to brand guidelines. One designer testing the tool shipped a full landing page in 3 to 5 hours instead of 2 to 3 days, keeping every generation on-brand because the system was locked once and reused.
The Claude Code to Figma bridge: When a design is ready to build, Claude Design packages it into a handoff bundle for Claude Code. From there, Figma's Code to Canvas feature (launched February 2026) converts the running UI back into editable Figma frames for collaborative polish. The full loop closes: capture your site, generate on-brand pages, ship to code, push to Figma for team review.
Claude Design is available in research preview for Pro, Max, Team, and Enterprise subscribers.
Zapier AI Agents
Stop Babysitting Your AI Workflows
As your list of AI workflows grows, a visibility gap becomes a real bottleneck: Teams have to dig through scattered chat windows to figure out which agent stalled, which finished, and which needs a human nudge.
Zapier's recent overhaul to Zapier Agents tackles this directly. Agents with multiple behaviors are now automatically split into single-purpose agents and grouped into Pods, while a new All Activity view consolidates every run across every agent into one screen.
Now You Can:
Cut review time by replacing agent-by-agent checks with a single dashboard
Catch stalled agents faster with clear "Needs action" status flags
Improve agent accuracy by reviewing run details and refining prompts based on real behavior
Scale agent deployments without losing operational visibility
Key functionality:
Pods: Group related agents together so an inbox triage agent, calendar agent, and task-routing agent live under one parent pod for easier oversight.
All Activity view: See every agent run in one feed, with filters for completed tasks, in-progress work, and runs flagged as "Needs action."
Run details panel: Drill into any individual run to inspect what data the agent referenced and how it interpreted instructions.
For teams already running two or more agents, switching to Pods is a fast way to turn a sprawling agent setup into a manageable system.
Set Trust Levels for AI, Not Free Passes
Cartoon by Tom Fishburne
A lot of organizations treat AI oversight as binary: either the model runs unchecked, or every output gets reviewed. The first option creates risk. The second buries your team in approvals and erases the time savings AI was supposed to deliver.
The smarter approach is tiered trust. Match the level of human oversight to the actual stakes of the task, so low-risk work moves automatically and high-risk decisions still get a human in the loop with the right context already assembled.
How to structure your trust tiers:
Auto-run (low risk): Drafting internal summaries, tagging tickets, enriching CRM contact fields, categorizing inbound leads. The AI acts and logs the action.
Flag on anomaly (medium risk): Updating deal stages, sending external emails from templates, processing refunds under a set dollar threshold. The AI acts but escalates when something looks unusual: a sudden 10x deal size, an unfamiliar vendor, sentiment that doesn't match the request.
Human approval required (high risk): Approving refunds above threshold, sending client-facing summaries, changing pricing, closing or reopening tickets tied to enterprise accounts. The AI prepares the recommendation and pulls every supporting detail into one view so the human isn't digging through five systems.
Next: How to Define Your Stop Criteria
Define Stop Criteria Before You Deploy
Tiers only work if you've decided in advance what would trigger an escalation, a pause, or a full stop. Jeff Gothelf makes this case for AI design work in a recent piece, arguing that without explicit stop criteria, "iteration becomes its own justification." The same logic applies to AI in operations.
Write down, before you turn anything on:
The specific anomaly thresholds that move a task from Tier 1 to Tier 2 (deal size jumps, sentiment shifts, dollar thresholds, volume spikes)
The error rate or false-positive rate that would pause the automation entirely for review
The review cadence for each tier (weekly spot checks for Tier 1, daily anomaly review for Tier 2, real-time approvals for Tier 3)
The owner for each tier, so escalations have a clear path
The goal is to move work down the trust ladder as evidence accumulates, not to lock everything in Tier 3 forever. Tool fluency is becoming table stakes. Knowing when to act, when to flag, and when to stop is what separates teams that compound an advantage from teams that drown in unchecked output.
A New Way to Choose Your AI Model
Measure Team Frustration & Get Our Pilot Checklist
Created by Base44
Most of us choose AI model based on benchmark scores, then wonder why employees keep complaining. A new data point from Base44 founder Maor Shlomo helps explain the gap. His team built a "frustration meter" that analyzes millions of real builder sessions daily, tracking bug loops and repeated requests to score how users actually feel after using each model.
A practical framework for your team:
Use Claude for long-document analysis, contract and policy review, brand-aligned writing, and coding projects. Its 1M token context window and instruction fidelity hold up across multi-step work, which is why roughly 70% of developers now prefer it for code.
Use ChatGPT for image generation, voice workflows, real-time web search, and tasks that need its broader integration ecosystem with Microsoft 365 and custom GPTs.
Use Gemini cautiously for now. Despite its 1M token context window and tight integration with Google Workspace, the frustration data suggests its rapid version changes are hurting day-to-day reliability. Worth piloting if you live in Google Docs, but verify outputs more carefully.
Use a mix when one team handles compliance and content while another handles creative or multimodal output. Most teams that rely on AI heavily end up paying for two tools at roughly $40 per user per month.
Get Our AI Pilot Checklist
AI Pilot Checklist
Setup (Day 0)
Pick 3 to 5 recurring tasks your team runs weekly (e.g., draft client emails, summarize meeting notes, review contracts, write code, build reports)
Assign 2 to 4 team members to run the same tasks in both Claude and ChatGPT
Create a shared tracking sheet (Google Sheets or Airtable works fine)
Daily Tracking (per task, per model)
Task name and date
Which model was used
Time to first usable draft (in minutes)
Number of follow-up prompts needed to get it right
Did you have to redo or heavily rewrite the output? (Yes/No)
Quick frustration rating from 1 to 5 (1 = smooth, 5 = painful)
Quality Spot-Checks (Weekly)
Pull 3 random outputs from each model and review for accuracy
Flag any hallucinations, fabricated sources, or factual errors
Note any tone or brand voice issues that needed editing
End-of-Pilot Review (Day 14)
Average frustration score per model
Average rework rate (% of outputs that needed major edits)
Total time saved vs. doing the task manually
Which task types favored which model
Final recommendation: Claude only, ChatGPT only, Gemini, or a mix
Decision Rule If one model wins on 70% or more of your team's most frequent tasks, standardize on it. If the wins are split across task types, budget for both and document which model handles which workflow.
SMART WORDS OF THE WEEK:
— Daniel J. Boorstin
Real progress comes from building the visibility, oversight, and tiered judgment that turn confident output into verified results.
 |
Stay in touch with our team from anywhere.
