How much would you let an AI agent do for you?

Two kinds of AI users, both in India

Everyone was a Millennial or Gen Z user in India, and every diary participant was a verified AI power user, so the feedback looked ahead rather than getting stuck on basic usability. We split them by how far they already let AI act for them.

Seven days of diaries, then 25 conversations

1. ScreenerBuilt on the client's criteria. The client approved every participant.
2. Daily diaryA WhatsApp group and one Google Form a day, plus a screen recording of each task as people did it.
3. Clean and chaseEach entry was checked in a shared sheet. Unclear answers meant a follow-up call before the next day's form went out.
4. Interviews90 minutes with agentic-curious users and 75 with agentic users, each with an early concept from the client to react to.
5. Sense-makingGrids for group and solo users, then journey maps, insights and the final deck.

One group says "Get this done for me". The other says "Help me decide"

Agentic users run multi-step workflows through agents because they know what to trust. Agentic-curious users only let AI help with the work: they don't yet know how far they can trust it to carry a task through.

Shopping, travel and studying fill people's days

Across 122 logged tasks, five kinds of task made up most of the diary. Agentic-curious users need agentic help most where tasks mean switching apps, comparing prices or digesting dense information, even though they don't use agents yet.

1. Shopping and dealsSolo21%26 of 122Tracking price changes, hunting coupons, comparing unit prices across Amazon, Swiggy and Zepto.
2. Group travel and commuteGroup16%20 of 122Multi-person itineraries, Airbnb reviews, logistics over WhatsApp and Goibibo.
3. Education and researchSolo14%17 of 122Digesting dense PDFs, summarising lectures, building study schedules.
4. Dining, social and eventsGroup13%16 of 122Group dinners, party logistics, shortlisting venues, finding tickets.
5. Travel and commuteSolo12%15 of 122Checking cab fares across ride-hailing apps, route costs, surge pricing.

How often, and how hard it felt

Each logged task was rated for how often people do it and how much effort it took, on a five-point scale from effortless to severe.

Today, people are the glue between apps

• The big picture

  AI chatbots handle the preparation, but the final step is scattered across apps.

  AI chatbots are the single biggest tool category people use today: the front door for research, planning and ideas. But fulfilment is split across transit, quick-commerce and e-commerce apps. Users act as the manual integration layer. A true agentic workflow has to carry them from discovery all the way to fulfilment.

• Finding 1

  Travel planning and shopping dominate daily manual workflows, with many high-stakes decisions and fragmented effort.

  High-stakes purchases and itineraries made up more than half of all tasks, with shopping first and travel second. Admin and tracking were a much smaller share.

  What this meansAgentic automation pays off most in travel and shopping, where people suffer through manual price and product comparisons again and again.

• Finding 2

  Solo tasks are about comparing; group tasks are about agreeing.

  Solo tasks centre on search, comparison and summarising, powered by AI chatbots (ChatGPT, Gemini, Perplexity) and search. Group tasks add the overhead of reaching consensus, run on WhatsApp, with money settled on Splitwise.

  What this meansSolo tasks need an agent that can dig deep across the web. Group tasks need one that can poll people and keep shared track of plans and costs.

• Finding 3

  Different blockers for different tasks.

  Group tasks are held back by trust. People fear costly mistakes and made-up answers, especially when money is involved; they'd hand the task over but step in at the payment. Solo tasks are held back by missing connections. People would happily let an agent order snacks on Blinkit, but the apps don't allow it, so they cross-check and switch between apps themselves.

In the screen recordingsWorkarounds everywhere: websites people had found just to compare prices across many apps, and group trips held together by phone calls and WhatsApp, where one person gathers everyone's preferences, does the research alone and reports back.

Agents already run parts of their lives

Agentic users use agents for financial oversight, business growth and daily execution, keeping personal and work life running without constant manual effort.

Swipe to see all seven →

• Research, comparison and analysis. Comparing products and summarising huge amounts of information instantly, instead of digging through sources one by one.
• Multi-step routines. Gathering information, tracking expenses, sorting email to pull out what matters, without hopping between apps.
• Creative and production work. Websites, presentations, social posts and teaching material, from idea to finished output with far less effort.
• Task coordination. Meetings, itineraries, calendars, reminders and daily task lists, kept in one place.

What this meansThese users deal with heavy information every day. AI should pull scattered information together, make it easy to digest and act on, narrow down the options and cut the app switching.

Hand over the work, keep the final call

Common groundControl over money was a firm no-go for both groups.

The sweet spot: AI does the prep, people make the final tap

Agentic-curious users prefer a "staged execution" model: the agent does the heavy lifting in the background and stops just before anything that commits them. That's where they save the most time at zero risk.

1. Multi-app comparisonUnit price and cart analysis

   Switching between 3 to 6 apps (Blinkit, Instamart, Zepto) to compare variable weights, hidden fees and availability causes real mental strain.

   Letting AI stage carts and build side-by-side comparison tables saves a lot of time with no money at stake.

2. Group consensusNeutral task coordination

   Outings stall on group indecision and clashing schedules; aligning everyone takes heavy effort.

   AI as a friendly, neutral "group co-host" that runs polls, tallies preferences and works out who owes what.

3. Synthesis and researchStructuring information

   Manual research is buried in ads, dense PDFs, scattered schedules and messy search results.

   Handing over the heavy lifting (filtering ads, summarising PDFs, organising multi-day itineraries) gives immediate, risk-free value.

Four domains cover about 80% of all daily workflows, making them the primary golden zones:

When they want control back

Right before the final decision, payment or anything irreversible, control has to pass from the AI back to the person for review and final approval.

What users kept asking forA hybrid: AI does all the background work, research and cart-building, but stops just short of the final decision so they can review and approve with one tap.

Share what the task needs, guard everything sensitive

People were comfortable giving full access to what a task actually needs, like files for an assignment or order history and a phone number for a food order. They hesitated over anything sensitive: bank details, PINs, and constant live tracking.

What to build: four trust blockers and how to answer them

1. Blocker 1Fear of costly errors and made-up answers
   →
   OpportunityAn "undo" window. Instant cancellation for any automated booking or out-of-routine action.
2. Blocker 2Apps aren't deeply connected
   →
   Opportunity1-tap links. Don't make the final payment; fill the cart and send the user straight to Blinkit or Uber to check out.
3. Blocker 3"Black box" anxiety
   →
   OpportunityTransparent logs. A step-by-step trail of exactly which apps AI checked, with side-by-side proof.
4. Blocker 4Privacy and personal data worries
   →
   OpportunityGranular permissions. Strict "ask before acting" prompts before AI opens photos or private files, or messages anyone.

Three early concepts, tested in the interviews

The client brought three early concepts and wanted to know how people perceived them, whether they'd accept them, and what feedback could shape the next version. Screens below are from the client's prototypes, with our annotations. Tap any screen to see it larger.

Concept 1 · Shown to agentic users and group-task users

An agentic coordinator for group plans

Ask the AI assistant to plan a weekend getaway with friends, and it creates a group task that gathers everyone's preferences, shortlists places and runs a vote.

People understood it as an AI agent helping with group planning, but expected it to move from explaining to acting early on: asking the right questions, gathering preferences, prompting for the next input.

Task card

People didn't notice the task card until prompted. They read the text answer first, and expected the agent to start asking questions early.

They wanted it to collect dates, budget, trip preferences and transport before planning, not show screens full of explanation.

What this meansMake the first interaction actionable by starting to gather information straight away.

"Working" status

"Working" confused people: was the agent collecting, waiting for, or analysing the group's inputs?

They couldn't tell whether it was gathering data, making a link, contacting friends or just doing something in the background.

What this meansSay clearly what the agent is doing at each point, so progress and next steps are obvious.

"Needs Input" status

People read "Needs Input" as a call for human direction, but weren't sure what to do next.

Give their own preference? Review the AI's suggestions? Wait for the others? Many expected the agent to spot clashes in people's availability and sort them out itself.

What this meansTell people exactly what's needed from them, so they're ready for the next step.

Preference cards

People liked the structured cards for choosing preferences, but wanted room to edit and add specifics like food.

Picking from cards cut the effort and the mistakes of typing everything out. They preferred typing extra wishes in the last card.

What this meansGather preferences visually and proactively, so people can set their intent without remembering or typing every detail.

Voting

People wanted to know how and why the shortlist and the winner were chosen, not just see the result.

They wanted the reasons behind each option, everyone's preferences, who voted for what, and a clear confirmation step before anything was booked.

What this meansBe open about the reasoning at every step of research and coordination.

1. How might we make the coordinator actionable from the first interaction to the end, without adding formal steps, screens or conversations?
   Open with the right questions (dates, budget, destination, transport, group size), cut screens with nothing to do, and use WhatsApp for updates and reminders.
2. How might we make the agent's progress and next steps transparent?
   Replace vague labels like "Working" and "Needs Input" with what the agent is doing and what it needs, and show which stage it's at.
3. How might we help groups move from individual wishes to a shared decision?
   Show places, dates, budgets, travel and stays side by side, so the group can vote and settle clashes.
4. How might we give AI enough autonomy to coordinate while people keep control of important, irreversible decisions?
   Let the agent prepare the plan, but ask the group to confirm before booking; add edit and override controls, shared preferences and access settings.

What people expect from a group coordinator, and the personality they want

1. Conversational preference gathering: ask relevant questions before acting, instead of long explanations up front.
2. Voting and polling: a clear, WhatsApp-poll-like way to compare itineraries (Coorg vs Pondicherry), see budgets and vote on places, dates and transport.
3. Rich decision support: pictures, credible reviews, deeper research and real alternatives, not one recommendation.
4. WhatsApp integration: updates and reminders where people already are, not in a separate app.
5. Group and access controls: who can join and interact, sharing or hiding preferences, adding questions, deciding which answers count.

The personality: a fair decision helper, neutral and able to mediate with empathy. Friendly and human, like an "AI cousin", but clear and concise, with no fluff, ads or chat spam.

• Teammate and helper: coordinates without taking overMost
• Neutral mediator: fair pros and cons in disagreementsMany
• Empathetic and calm: understands budgets, babies, elderly needsMany
• Non-intrusive: works quietly, respects privacySome
• Witty and reasonable: smart, practical, groundedFew

Concept 2 · Shown to solo-task users

A multi-cart grocery journey

Ask the AI assistant to build a grocery cart, and it fills carts across several quick-commerce apps, compares them side by side and offers one checkout.

People valued several carts in one view over a single cart for comparing prices across apps, but expected to edit the carts right there, with clear information about missing items in each.

Choosing apps

People held conflicting ideas about where the journey would happen: inside the assistant, or in the shopping app.

Most thought "Continue" would open the app; some feared it would charge them. "Connect" was understood straight away. Staying in one view pleasantly surprised most people.

What this meansMake clear what every button will do.

Editing the cart

People liked having the cart in the same place, but wanted to edit it the way shopping apps allow, not by typing.

Most wanted "+" and "−" buttons and swipe-to-delete, instead of describing every change in the chat box.

What this meansKeep the cart editable on the same screen.

Checkout

Different cart totals confused people, because they couldn't tell which items were missing where.

Some didn't realise their list had been split across three carts, and checked each one by hand. A few wanted one best shop with everything, or clear flags for what's missing.

What this meansShow exactly what's in, and missing from, each app's cart.

What people expect from a multi-cart grocery journey

1. Single-tap payment in one place, automatically split between shops, rather than paying in three apps.
2. All deliveries tracked together, ideally arriving in the same window.
3. Cart controls: "+" and "−", dietary toggles, missing-item alerts.
4. Discounts found and applied automatically across apps, so no deal is missed.

Concept 3 · Shown to solo-task users

Restaurant table booking with deals built in

The AI assistant finds a restaurant, gathers the offers from several booking platforms, lets you pick one and reserves the table.

People wanted every offer in one place with its details, clear prices upfront (cost for two, cover charge, total), and the menu without leaving.

Deal stacking

People liked deals gathered in one place, but wanted to see the details and choose themselves.

"We found the best deal" wasn't enough: they wanted to know where each offer came from and its conditions, like "only with a certain credit card". The booking label was read two ways: helping, or booking on their behalf?

What this meansShow offer details openly so people can pick one they actually qualify for.

Booking

People wanted the whole booking handled in one place, with costs shown first and confirmation on familiar apps like WhatsApp.

Every agentic-curious user expected to book without leaving. Most were happy for AI to reserve, as long as charges like cover fees came first.

What this meansThe AI should be the start and end of the booking journey.

Across both solo concepts

1. How might we make it easy to change or refine what AI has prepared, without asking the AI?
   Familiar controls like shopping apps have: "+" and "−", remove, edit preferences, change selections, right inside the AI experience.
2. How might we make the agent's progress and next steps transparent?
   Say what each button will do: open an app, create a cart, apply an offer, or book or order.
3. How might we help people understand options well enough to choose with confidence?
   Show the total and per-person cost, offer conditions, reviews, menu prices, availability, missing items and alternatives upfront.
4. How might we let AI do the work while people keep the final decision?
   AI prepares the order or booking; people pick the offer, review the amount, and approve the booking or payment themselves.

What the client cared about most: an agent that talks to local businesses

We asked where people would want an agent to reach out for them, for example to a local salon. They saw it as a go-between for them and local businesses, handling the back-and-forth they'd rather not.

• Do the legworkChecking timings, availability and prices, without making the same call to vendor after vendor.
• Find the contactWhen they don't have a number, the agent can find it and reach the business directly.
• Be polite, like a person wouldStart with a text (SMS or DM), and call only if the business doesn't reply in reasonable time.
• Bring the decision back to meAI can't judge whether a business is trustworthy or protect their accounts, so people want the final information and transaction, checked step by step.
• Sound human, not roboticPeople are tired of robotic customer-service voices that don't take feedback well. Many felt businesses would engage with a human-sounding voice, and might hang up on a bot.
• Speak like a localSome felt a local accent would get a better price, and a foreign one a higher quote.

What the research gave the client

The project wrapped up recently, so what the client builds from it isn't public yet. What it gave them was a clear, evidence-based picture to make product decisions with.

1. Sharper conceptsScreen-by-screen feedback on three early concepts, including the buttons and statuses they'd planned.
2. Where AI can take the main seatThe tasks people are ready to hand over, and how agents fit into their day.
3. The limitsWhere people draw hard lines, how much data they'll share, and their firm "no"s. Control over money is a no-go for both groups.
4. How a voice agent should behaveText first, call as a fallback, and sound human, the area the client was most interested in.

From the clientThe client lead said the work gave them the depth they were looking for.

What I took away