CLASS 7June 16 · AI-Enhanced Workflows

Class 7: AI-Enhanced Workflows. "Don't replace a human with AI. Double your power instead."

Taught byIrina KellyUX Designer, Salesforce
Leads AI-enhanced workflows at product; part of the Creative AI team

The class in brief

Irina Kelly (Salesforce) taught the AI-Enhanced Workflow Canvas: map a project's steps, then work down through five rows.

The first row is desired outcome.
The second row is barriers.
The third row is AI use case.
The fourth row is tools.
The fifth row is how your own role shifts.

Dom and Tony opened with a debate on who owns creative value once AI speeds everything up. After this page you can map any real workflow onto the canvas, tag each AI use Automate, Enhance, or Advise, and stop starting every chat from a blank page.

Pick two: good, fast, or cheap

Why this matters · 6:21 PM

Clients are compressing time, not quality. Someone has to hold the line.

The night opened with the room's running AI-portrait debate about whether an artist should be paid for AI-assisted revisions, and who owns the vision through them . Dom's reframe: clients used to give three weeks for a creative exercise. Now the request reads the same but assumes AI, and the time shrinks to one.

A classmate read the market the same way from the job side. Postings that once just asked for design skill now add "scale your design work with AI" as a qualifier, a year apart.

"We kind of have to stand up as protectors of curation, and what is good."

Dom, on the creative's job once AI compresses time

of three you get: good, fast, cheap. "Whatever tools we use to make the sausage, that's how long it takes to make the sausage."

year is roughly how long it took job postings to start asking designers to "scale your design work with AI," per the room's own read of the market.

The framework · 6:44 PM

The AI-Enhanced Workflow Canvas

Irina opened her half of the night on the reframe the whole lesson sits on: organizations think "replace one human with AI."

What she wants instead is everybody keeping their own AI partner and getting twice as capable .

The canvas itself has one student per Miro frame. Columns are your project's 4 to 6 high-level steps, and five rows go underneath.

Desired outcome"if I do it well, it will be…"

The adjectives you'd use to describe a step done right. This row is where the human expresses taste; that's the value nobody automates away.

Barriers"but first I must…"

Not a problem statement. The one hard thing you must actually do to reach the outcome, kept separate from what good looks like.

AI use case"it would be great if I had help…"

Where AI could take load off this step. Leaving it blank, "this stays human," is a valid answer.

AI tools"for this, I could use…"

Name the actual tool per step and tag it. One step can carry several tools.

My role"now I've become…, with more time for…"

How the job itself shifts once the row above is filled in. The night's real question, made concrete per project.

The exercise · 7:16 PM

Work out your own workflow

The model answer

Irina's own six UX steps, filled all the way across.

She ran her own workflow (client briefing, brainstorm, research, and so on) through the canvas first. Desired Outcome and Barriers were written out row by row, as the pattern to copy .

Homework: finish the last two rows, AI Tools and My Role, by the next class.

A canvas finished early

A classmate working on an auto-show booth project had a canvas filled end to end well before the room's quiet work block ended. It was a finished canvas, steps mapped from brief to install, each with its own barrier and AI use case named plainly rather than left blank.

Leaving a row blank is an answer

Irina's repeated point through the work block: a step with no AI Use Case entry isn't an unfinished row. It's the canvas telling you the step stays human, and that's worth writing down as deliberately as any tool name.

Automate, enhance, or advise

The craft · 7:55 PM

Three questions to ask of any AI use case before you name a tool.

Post-break, the AI Use Case row got its own vocabulary . Can it Automate a repetitive task and just run (meeting notes, weekly research)? Can it Enhance by working with you (brainstorm, "upload a brief, give me 15 campaign territories with mood and headlines")?

Or can it Advise, surfacing insight you couldn't reach alone, like reading a client's history to predict their objections before the meeting ?

mindsets to tag any AI use case against: Automate, Enhance, Advise. One step can carry more than one.

campaign territories, mood plus headlines, was Irina's own example of what "Enhance" looks like in one prompt.

The tools · 8:18 PM onward

Sorted by what they actually do

Irina's own habit, when a classmate asked how she keeps projects straight across so many tools: "I already have a workspace that knows so much about my project. I'm not starting from a blank state."

One notebook or gem per client, reused every session, not a fresh chat each time. The slides behind it: the Communication category and its category card.

Research tools carried the same discipline further. NotebookLM answers only from what you upload and cites who said what. It says "I don't have that" instead of guessing, because for work that must be right, accuracy outranks creativity.

Perplexity's Model Council runs one question across Google, ChatGPT, and Claude at once . It shows where they agree, disagree, and go their own way .

On the design side, Figma Make turns a prompt into a click-through prototype and Figma Sites publishes it. On presentations, Gamma and Beautiful.ai draft a deck from a dump of thoughts so time goes to the story, not the pixels.

"I don't start from a blank chat. I have a workspace set up that already knows my project."

Irina, answering how she keeps projects straight across tools

Methods and prompts

Five methods to take with you

METHOD 01 · THE CANVAS part of: keep the judgment human

Map the workflow yourself first

Write your own steps, desired outcomes, and barriers before asking AI anything. The canvas only works if the taste (the outcome row) and the real friction (the barrier row) come from you.

Where it came fromTaught by Irina Kelly during the AI-Enhanced Workflow Canvas exercise in class 7. Each student mapped four to six steps of a real project, filled in the desired outcome and barriers, then named an AI use case.Use it whenReach for this when you are about to bring AI into a project and want to name what success looks like and where the real difficulty sits before letting a tool answer for you.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here is one real workflow from my work, 4 to 6 steps beginning to end: [list them]. I'll write the desired outcome and the one barrier for each step first. Then you check me: did I keep "what good looks like" separate from "what it takes to get there"?

For any step I leave without an AI use case, tell me honestly whether "this stays human" is true, or whether I'm just avoiding it.

You will know it worked whenit names a specific step where you likely wrote a barrier instead of an outcome. Or it calls out a 'stays human' step you may be avoiding.

METHOD 02 · THE MINDSETS part of: structure your ask

Tag each task Automate, Enhance or Advise

Every AI use case earns one of three labels before you pick a tool. The label tells you what kind of tool you actually need, not just that you need one.

Where it came fromIntroduced by Irina Kelly right after the workflow canvas exercise in class 7, as the vocabulary for the AI Use Case row: label each AI use as Automate, Enhance, or Advise before choosing a tool.Use it whenUse this once you have listed possible AI uses for a workflow and need to figure out what kind of tool each one actually calls for.

Same idea elsewhereGive it two keyframes instead of one description (class 2) · Speak the language of the internet (class 3) · Write it as a GOLD block (class 4)

Working prompt

Here's my workflow with an AI use case noted for each step: [paste]. Tag each one Automate (runs on its own, set once), Enhance (works with me, brainstorm-style), or Advise (surfaces insight I couldn't get alone). One step can carry more than one tag. If a step doesn't actually need AI, say so plainly.

You will know it worked wheneach step in your workflow carries one of the three tags, and it says plainly if a step doesn't actually need AI at all.

METHOD 03 · THE WORKSPACE HABIT part of: context beats prompts

Prime a workspace instead of starting blank

Irina's own answer to "how do you stay consistent across tools": one notebook or gem per client, fed once, reused every session instead of re-explaining the project each time.

Where it came fromIrina Kelly's answer in class 7 when a student asked how she keeps projects consistent across different AI tools: she keeps one notebook or gem per client, loaded once and reused every session.Use it whenSet this up when you are starting repeated work on the same project and are tired of re-explaining the brief in every new chat.

Same idea elsewhereControl the input and build gradually (class 2) · Show it your voice and not just the output (class 3) · Evolve the reference instead of restarting (class 5)

Working prompt

Before we start: here is my project [upload the brief, brand material, past work, whatever you actually have]. Study it and hold it as context for this entire conversation. Don't make me re-explain the project every time I ask something new.

You will know it worked whenlater questions in the same conversation get answered using the project material you uploaded, without you having to repeat it.

METHOD 04 · RESEARCH part of: go wide then narrow

Run one question through a model council

Perplexity's Model Council pattern, portable to any set of models you have access to: same question, multiple models, and the value is in where they disagree.

Where it came fromCovered in class 7 during Irina Kelly's research tools segment, describing Perplexity's Model Council feature, which runs one question across multiple models at once and shows where they agree or disagree.Use it whenUse this when a research question matters enough that you want to see where different AI models disagree, not just take one answer.

Same idea elsewhereThe lateral reframe (class 2) · Start a fresh chat instead of one long thread (class 3) · Force the creative partner deeper (class 4)

Working prompt

Research [topic], and before giving me one answer, show me where different models would agree, where they'd disagree, and what unique angle each brings. Then give me your synthesis and flag anything you're least confident about.

You will know it worked whenit names at least one specific point where models would disagree, not just a single averaged answer, and flags its own weakest point.

METHOD 05 · RESEARCH part of: know where it fails

Ground the answer in your own sources

NotebookLM's core discipline, usable in any tool that supports uploaded context: the model answers only from what you gave it, and says so when it can't.

Where it came fromCovered in class 7 during Irina Kelly's research tools segment, explaining NotebookLM's discipline of answering only from documents you upload and saying so plainly when it cannot.Use it whenUse this when accuracy matters more than creativity and you need an answer that will not fill gaps with the model's outside guesses.

Working prompt

Answer only from the sources I've given you: [upload]. If something isn't in these documents, tell me you don't have it instead of filling the gap from general knowledge. Cite which source each claim comes from.

You will know it worked whenevery claim it makes is tied to one of your uploaded sources, and it tells you plainly when something isn't in them at all.

The close · 9:03 PM

The homework is the last two rows

The closing slide left the room with a fully-filled example canvas, tool logos named per step. It was the target for the two rows Irina skipped for time .

The AI Tools row asks you to name the tool per step and tag it. The My Role row asks you to finish the sentence "now I've become…, and I have more time for…." Both rows are due by the next class.

Try this prompt

Quiz me on the five rows of the AI-Enhanced Workflow Canvas and the three mindsets. Then hand me a small real workflow of mine and make me fill the canvas out loud, one row at a time.

You will know it worked whenit quizzes you on the five rows of the workflow canvas and the three mindsets first. Then it makes you fill the canvas out loud on a real workflow, one row at a time.

The shelf

Tools and references

Tools that night

The AI-Enhanced Workflow Canvas, one per student on Miro
Meeting-capture tools: Otter.ai, Fireflies, Teams, Zoom AI.
Source-grounded knowledge tools: NotebookLM, Copilot Notebook.
Perplexity + Deep Research + Model Council; Synthetic Users
Midjourney, Ideogram, Nano Banana; Figma Make, Figma Sites
Gamma, Pop.AI, Beautiful.ai, Figma Slides, Copilot PowerPoint

Named in the room

Good, fast, cheap: pick two, the room's framing for compressed timelines
An ethics session and an exponential-thinking class, both named as coming later in the semester
The Canvas homework: finish the AI Tools and My Role rows by the next class

Every frame from this class

35 captures, in order. Click any one to see it full size.

Pick two: good, fast, or cheap

A slide titled Before Class: Explore and Prepare, Hopes and Fears of AI in My Work, with a three step homework task and a task deadline. — Before the AI Enhanced Workflows class, taught by Irina Kelly of Salesforce, students were asked to post on a shared Miro board a favorite or least favorite part of their work. They also posted their hopes and fears about AI in that part of the job.

The Workflow Canvas

A slide reading 'A symbiotic, learning relationship for both the Human and the AI' with two overlapping shapes labeled Human and AI. — Irina Kelly's opening slide framed the class thesis: AI enhanced workflows should be a symbiotic, learning relationship between the human and the AI, not one replacing the other.

A slide titled Some Workflows at an Ad Agency listing nineteen numbered steps from client brief intake through wrap and learnings. — As a reference for the AI Enhanced Workflow Canvas exercise, this slide lists nineteen typical steps in an ad agency project, from client brief intake through wrap and learnings, for students to group into their own workflow stages.

A slide titled UX Design Workflow showing a horizontal flowchart of four steps: Briefing, Brainstorm, Research, Design. — Irina Kelly used her own UX design process, Briefing, Brainstorm, Research, and Design, as a simplified model workflow for students to reference while mapping their own projects.

A slide titled Some Workflows at an Experiential Agency listing twenty numbered steps from client brief intake through post event debrief. — A second reference workflow, from an experiential events agency, gave students a twenty step example, from client brief intake through post event debrief, to compare against their own process.

A slide titled Work Out Your Workflow showing a filled in table of Barriers and AI Use Cases across six UX process steps. — An example canvas showed how to fill in the Barriers and AI Use Case rows for a UX workflow, based on designing a commerce platform experience, modeling what students were asked to do with their own projects.

A blank Miro template titled AI Enhanced Workflow Canvas with rows for Desired Outcome, Barriers, AI Use Case, AI Tools, and My Role. — This blank Miro template was the core exercise of the class: students filled in their own project's steps. Then they filled in rows for Desired Outcome, Barriers, AI Use Case, AI Tools, and My Role, to map where AI fits into their actual work.

A slide titled Desired Outcome for UX Workflow with a table defining desired outcomes for Client Briefing, Brainstorm, and Research steps. — A model answer showed how to write a Desired Outcome statement, starting with 'if I do it well, it will be...', for the Client Briefing, Brainstorm, and Research steps of a UX workflow.

Work out your workflow

A slide titled Barriers for UX Workflow showing a table of 'but first I must' barrier statements for Client Briefing, Brainstorm, and Research steps. — A model answer showed how to phrase a Barrier statement, starting with 'but first I must...', for the Client Briefing, Brainstorm, and Research steps of a UX workflow.

Automate, enhance, or advise

A slide titled Turn Obstacles into Opportunities with three columns, Automate, Enhance, and Advise, each listing example AI prompts. — Irina Kelly's three part framework for using AI, Automate repetitive tasks, Enhance creative work, and Advise on decisions, was shown fully worked with example prompts for each, the lens students then used to fill their own AI Tools row.

A slide titled AI Use Case, 10 mins, showing a filled in workflow table across six UX process steps. — This example showed how students should complete the AI Use Case row of their canvas, phrased as 'I would be great if I had help...', across all six steps of a UX workflow.

Sorting the tools

A title slide reading 'Elevate Communication with AI' listing Research, Design, and Planning and Execution. — Irina Kelly organized the AI tools portion of the class into four categories, Communication, Research, Design, and Planning and Execution, starting with tools that improve how teams communicate.

A slide showing the six panel cartoon of how a swing project was explained, understood, designed, documented, installed, and billed. — A classic cartoon of a tree swing showed how a project idea gets distorted between what the customer explained, what different teams understood, and what finally got built and billed, setting up the class discussion on why communication tools matter.

A slide comparing meeting transcription AI tools Otter.ai, Fireflies, Teams, and Zoom AI, with a search example asking 'What did I promise Charlotte?' — Meeting transcription tools like Otter.ai, Fireflies, Teams, and Zoom AI were shown automatically capturing key points and action items from a meeting, so nothing discussed gets forgotten or missed.

A slide titled Microsoft Copilot showing a Teams meeting screen with automatic notes and a meeting recap card. — Microsoft Copilot was shown documenting a meeting automatically and filling in missing details from related emails and chats, so no critical context gets lost.

A slide titled NotebookLM, Copilot Notebook showing a Copilot notebook interface next to NotebookLM screens for a research paper example. — NotebookLM and Copilot Notebook were shown turning a folder of uploaded briefs, specs, notes, and recaps into one searchable project brain that the whole team can query.

A slide titled Text to Visual showing pasted performance data converted into a bar chart dashboard with an executive summary. — Microsoft Copilot or Google Gemini can turn pasted client feedback or report data into a clean visual summary, like a chart with an executive summary, without manual formatting.

A slide titled Brainstorm with an LLM showing example screenshots of Claude, ChatGPT, and Gemini generating premium membership perk ideas. — Irina Kelly's tips for brainstorming with ChatGPT, Gemini, or Claude were to start most projects with an open conversation, use natural language, be humble, ramble freely, and ask for action items at the end.

A slide titled Large Files and Data showing a NotebookLM workspace called Velmora Nuit Doree Campaign Strategy and Launch Blueprint. — NotebookLM was shown synthesizing a large set of uploaded campaign documents into one searchable notebook, so a team could ask it questions like what the strongest creative directions and open questions are.

A screenshot of the NotebookLM homepage showing featured notebooks including Sherlock Holmes, a Google brain research paper, and an Economist globalization report. — A look at NotebookLM's own homepage showed the range of source material it can organize, from a Sherlock Holmes collection to a Google research paper and an Economist report.

A slide titled Perplexity Labs, Research describing project based AI research, with a screenshot of the Perplexity Labs interface. — Perplexity Labs was introduced as a project based research tool that tracks queries and organizes findings over time into a living dashboard of detailed reports, spreadsheets, and interactive dashboards.

A screenshot of an AI search answer to 'what are top 5 most popular ai tools now' listing ChatGPT, Canva, Gemini, Grok, and DeepSeek in a ranked table. — A live demo asked Perplexity for the top five most popular AI tools right now, returning a cited, ranked table with ChatGPT at the top, followed by Canva, Gemini, Grok, and DeepSeek.

A slide titled Perplexity Deep Research showing the Perplexity Pro interface with a Deep Research option menu open. — Perplexity's Deep Research mode was shown breaking a broad question into smaller sub-questions, searching many sources, and cross checking them into one cited report.

A slide titled Research Deeper, with Sources comparing deep research tools, showing a ChatGPT Researcher interface with footwear market research prompts. — Deep research agents like ChatGPT's Researcher were shown autonomously analyzing multiple sources, summarizing trends, and citing every reference, useful for tasks like competitive analysis.

A slide titled Creating User Personas showing an AI generated persona named Oliver with a personality radar chart and demographic details. — Synthetic Users software was shown generating AI driven personas, like the example persona Oliver, so designers can interview a realistic stand in when access to real users is limited.

A slide titled Human in the Loop showing three sticky notes about workplace software habits above an embedded post from Sam Altman about GPT-4o's sycophantic personality update. — This slide paired sticky note complaints about workplace AI tools with a real post from OpenAI's Sam Altman. He admitted a GPT-4o update had made the AI too sycophantic and people pleasing. Together they made the case for keeping a human in the loop.

A slide titled Visual Exploration describing Ideogram and Midjourney as tools for early concept and mood board generation, with screenshots of both apps. — Ideogram and Midjourney were recommended for exploring visual style, mood, and unexpected aesthetic directions early in a project, though the instructor noted they are less reliable for exact, repeatable design specs.

A slide titled Figma Sites describing team website publishing directly in Figma, with an interface showing a flower shop website design. — Figma Sites lets a team design, build, and publish a real website together inside Figma, which the instructor contrasted with the more solo first Figma Make.

A slide titled Nano Banana describing Google's Gemini image editing model, showing a banana render and example product mockup image pairs. — Nano Banana, Google's Gemini based image tool, was shown generating and editing consistent product mockups and stylized visuals from text prompts, positioned as an easy entry point for beginners.

AI is changing how design feedback gets collected, with heat maps becoming more accessible, LLMs giving structured usability comments, and synthetic user personas like Oliver simulating realistic reactions before real users are available.

A slide titled Drafting Presentations comparing Gamma, Pop.AI, Beautiful.ai, and Copilot PowerPoint, showing sample presentation template grids. — Tools like Gamma, Pop.AI, and Beautiful.ai were shown turning dumped thoughts, documents, and assets into a draft presentation, while Copilot PowerPoint offered a lower learning curve starting point.

A slide titled Drafting Presentations describing Google Slides' 'Enhance Slide' AI feature, with a screenshot of a Gemini panel generating a bar chart slide. — Google Slides' Enhance Slide feature was shown generating a presentation draft from an existing template and brand deck, so the first draft already reads on brand.

A slide titled Add AI Your Workflow Homework showing a completed canvas with rows of AI use cases and small tool logos across five steps. — The homework for Irina Kelly's AI Enhanced Workflows class is to finish the AI Enhanced Workflow Canvas. Fill in the AI Tools row, tagged by Automate, Enhance, or Advise, and the My Role row for every step, shown here with a fully worked example.

Class 7: AI-Enhanced Workflows. "Don't replace a human with AI. Double your power instead."

Pick two: good, fast, or cheap

The AI-Enhanced Workflow Canvas

Desired outcome"if I do it well, it will be…"

Barriers"but first I must…"

AI use case"it would be great if I had help…"

AI tools"for this, I could use…"

My role"now I've become…, with more time for…"

Work out your own workflow

Irina's own six UX steps, filled all the way across.

A canvas finished early

Leaving a row blank is an answer

Automate, enhance, or advise

Sorted by what they actually do

Five methods to take with you

Map the workflow yourself first

Tag each task Automate, Enhance or Advise

Prime a workspace instead of starting blank

Run one question through a model council

Ground the answer in your own sources

The homework is the last two rows

Where this class connects

Tools and references

Tools that night

Named in the room

Every frame from this class