Prince Pal State of AI · 2027 Hire me
The state of AI · 2027

AI can generate almost anything. What should we design next?

Teams are shipping faster than ever — yet customers aren't feeling it any sooner. A five-part story about what it now takes to design AI products people can rely on: systems, trust, guardrails and agents that act on our behalf.

Based on The State of AI for 2027, a presentation by Vitaly Friedman · Maven
Written and designed by Prince Pal ·

Scroll to begin
Not one answer anymore — a system of decisions, tools and actions.
Prologue · The efficiency trap
{{ w }}
Hiten Shah, “What AI Did To Product Teams” — quoted in the talk
In brief

The state of AI for 2027, in five sentences.

  1. 01AI is everywhere, value isn't: 99% of product teams use AI in at least one workflow, yet 80% ship faster while customers see no benefit any sooner.
  2. 02Design systems keep AI on-brand: without a closed set of tokens, specs and audits, generated interfaces drift.
  3. 03Trust comes from control: people trust AI output they can verify, shape and correct — not output that merely sounds confident.
  4. 04Reliability lives in the harness: context engineering, tool registries, guardrails, agent loops and verification matter as much as the model.
  5. 05Agents need visible autonomy: suggest, confirm and execute modes, set per action, with approvals, audit trails and recovery.
Chapter 01 / 05

The state
of AI.

AI is producing more than ever. But more output doesn't automatically mean better products — or faster outcomes for the people who use them.

AI is out of the pilot stage. In the State of Product survey shown in the talk, 99% of product teams said they had fully integrated AI into at least one workflow, and 96% said it was equal to or better than human work in at least one area.

Yet the bottleneck has moved rather than disappeared. Teams ship faster, but customers aren't benefiting any sooner. Most say AI hasn't sped up decision-making — and executive gut instinct still regularly overrides data.

Many AI experiences are no longer chatbots. Models are improving, but they hallucinate more subtly. By default, AI output is generic and verbose. It works best when it's guided — in other words, designed.

{{ s.v }}%
{{ s.label }}
{{ s.v }} {{ s.t }}

Source — State of Product report (Atlassian), as presented in the talk. Figures were recovered from slide images; read them as self-reported survey results, not independent findings.

Fig. 1.1 — Output vs. outcome
{{ l.big }}
{{ l.small }}
{{ c.label }}
The rise of AI workslop

“Work that masquerades as good work, but lacks substance to be useful.”

That's how Kate Niederhoffer describes AI workslop in Harvard Business Review: low-effort, passable output that creates more work for coworkers downstream. In the State of Product data, 82% said team members regularly produce low-effort AI work and revise it based on feedback, and roughly half spend time monitoring their agents.

Fig. 1.2 — AI productivity study, US AI doesn't reduce work — it intensifies it.
{{ b.label }} {{ b.v }}

Change in time or frequency, as listed in the talk, via Mike Rosenberg (NBC News, HBR, WSJ, ActivTrak). Outlined bar = decrease. These are reported changes alongside AI adoption — they don't show AI caused them.

What AI did to product teams.

After Hiten Shah
  1. {{ e.n }} {{ e.t }}
The state of AI, in ten lines
  1. {{ e.n }}{{ e.t }}
    2.
AI constraints
The quality of AI responses is typically constrained by three factors.
  1. 01 Training datahard to control
  2. 02 Model design and tuninghard to control
  3. 03 The user's inputeasier to influence
Probabilistic vs. deterministic
Teams say “add an agent” or “use an LLM” instead of deciding which features will be probabilistic and which deterministic — or when human judgement is critical.
Interlude · Beyond the text box

The many faces of AI.

AI can go way beyond a text box. Products are experimenting across a dozen kinds of scaffolding — and AI often works best when it's deeply integrated into the workflows people already have.

{{ s.n }} {{ s.t }} {{ s.d }}

Types of AI scaffolding, as categorised in the talk. See also Sharang Sharma, “Where should AI sit in your UI?”

The drawbacks of text prompts
“We put the burden on the user to articulate good questions, but they may not know what exactly to ask.”
Austin Z. Henley, “Natural Language Is The Lazy UI”

A good UI helps people explore the problem and solution space step by step, with guidance and nudges. “Ask me anything” interfaces do the opposite: people don't know what to ask, how to ask it, or which details matter. That's the blank-page effect — Google's Human-AI Interaction Guide calls the blank page the biggest risk to effort.

Humans are also quite inefficient at articulating intent. In IEEE Spectrum, Dina Genkina reported that when a model was asked to devise its own optimal prompt, the generated prompt beat the best human trial-and-error prompt in almost every case. Many slowdowns, the talk argues, can be fixed with just enough good old UI. Further reading: Amelia Wattenberger, “Why Chatbots Are Not The Future”.

Interactive · Fig. I.1

A field guide to AI interface patterns.

Twenty patterns from the talk, grouped by where they help: shaping input, gathering context, making output usable, and running repeatable work.

{{ c.n }} {{ c.t }} {{ c.d }} {{ c.src }}

Pattern descriptions paraphrased from the talk.

Scaffolding, in ten lines
  1. {{ e.n }}{{ e.t }}
    2.

The fragments only become useful once they line up behind an outcome. That grid has a name: a design system.

Chapter 02 / 05

AI-ready design systems & drifting.

“AI output is unreliable because of tiny inconsistencies compounding over time. It's decisions made but not documented, hard-coded values never cleaned up, relying too much on AI to ‘get’ it right.”
The talk, “Design Decisions Infrastructure”
2.1 FoundationsTokens

A closed set of variables AI must choose from.

{{ w.n }}
AaAa
2.2 ComponentsSpecs

Spec files are reference docs for every component.

Start building Read the spec
Tokenised Audited 
button.primary
  radius:   radius-md      (8px)
  padding:  space-10 / 20
  fill:     primary        (one per view)
  label:    sentence case, verb-first
  states:   default, pressed, focus, disabled
2.3 ScreensAssembled

Same parts, same rules — a screen that holds together.

{{ s.tag }} {{ s.title }} Get started
2.4 DriftGenerated, three times

Each variant looks plausible. Together, the system has quietly drifted.

{{ s.tag }} Upgrade your plan {{ s.label }}
{{ s.note }}
Interactive · Fig. 2.1

Let the system drift.

Six screens, generated from the same brief. Raise the drift and watch small, individually reasonable deviations compound. Then switch on constraints: tokens become a closed set, and an audit script catches what slips through.

{{ auditText }}
Generated screen {{ s.n }} Invite your team Share the workspace and review work together. {{ s.label }} {{ s.flag }}

Illustrative simulation — not data from the talk.

Fig. 2.2 — Decisions as infrastructure

Every important decision finds a path back into the system.

The talk proposes treating design decisions as infrastructure: each one flows into a shared skill that AI then consumes. By default, AI takes the shortest path to “done” — it isn't a mind-reader, it needs skills.

A skill isn't a reference doc. It's a workflow, a way of thinking: a Markdown file injected into the AI's context when the situation calls for it. That's the beginning of guided AI, as Tey Bannerman puts it — pre-writing our thinking, principles and context rather than feeding generic input that drives generic output.

  1. {{ p.n }} {{ p.t }} {{ p.d }}

Loop as described in the talk: decisions flow into updated spec files in the same diff. Further reading: Hardik Pandya, “Expose your design system to LLMs”.

{{ q.k }}
{{ q.q }}
{{ q.a }}

The talk adds a useful frame: an agent's learning loop mirrors how a new team member onboards. It reads the codebase, finds out who knows what, and asks the right humans the right questions. Now, it's all about the quality of specs.

Look closely and the grid fractures in a few places. The most dangerous crack is the one that looks perfectly correct.

Chapter 03 / 05

Designing for trust.

“Strong AI benchmark scores don't mean good judgement. AI generates fluent answers, and hallucinations are more subtle. With it, trust is at risk.”
Hiten Shah, quoted in the talk

As Arvind Narayanan and Sayash Kapoor argue, despite AI's capabilities it isn't reliable enough to take the human out of the loop. When AI ships inside consumer products, people expect it to behave like software: predictable and accurate.

And people don't check AI output rigorously. As NN/g's research cited in the talk puts it, they decide based on how closely it fits what they already believe, know or trust.

Trust, in Mary Borysova's definition, means believing — despite uncertainty — that a system is competent, predictable, aligned with your goals and transparent about its limits.

Fig. 3.1 — “In general, how much do you trust AI?” “Do not trust at all” rose 9 points, to 36%.
{{ r.when }}
{{ g.v }}
■ Trust a lot■ Trust some■ Trust a little■ Do not trust at all

US adults, Aug 2025 vs May 2026, as shown in the talk. The pollster isn't named in the converted slide text; values recovered by OCR. Only 11% said AI should be adopted widely.

Fig. 3.1b — Net trust to use AI responsibly, US People only trust themselves and their circle.
{{ r.label }} {{ r.v }}

Share who say each can be trusted to use AI responsibly, minus the share who say they cannot. As shown in the talk; values recovered by OCR. Dark = net positive, red = net negative.

{{ d.v }} {{ d.t }}

Trust varies sharply by audience: in the talk's data it was lowest among Gen Z (4.1) and highest among upper-income earners (23.7). The headwind is concentrated in the English-speaking developed world — the US, Canada and Australia were the only markets where ChatGPT's net trust declined.

Morning Consult brand and category tracking, as shown in the talk. Figures recovered by OCR.

AI fatigue
{{ q.q }}
{{ q.a }}
Interactive · Fig. 3.2

Polished is not the same as right.

The same answer, twice. First as most AI products present it — fluent, confident, finished. Then with its evidence exposed, and with controls to correct it. A disclaimer alone doesn't create trust; seeing where an answer comes from, and being able to fix it, does.

Why did churn rise in Q3, and what should we do?
Assistant {{ confText }}

{{ c.text }}

{{ c.tag }} {{ c.note }}

Illustrative interface — the company, figures and sources are fictional.

Fig. 3.3 — The cost of fixing errors

“Just fix it” is six steps of real thinking.

Fixing errors might feel easier than writing from scratch, but it has a cost. And as systems researcher David D. Woods notes, a finished product is harder to evaluate than one you worked on in parts.

  1. {{ s.n }} {{ s.t }}

Steps as listed in the talk. Step 06 often loops — “a few times”.

Fig. 3.4 — How trust emerges

Trust doesn't emerge by default.

It comes from understanding a product, using it successfully, achieving goals — and then expanding usage. It's about consistent reliability. And it isn't just good intentions: people need confidence that no harm will come their way, whether hidden fees, privacy, security, sensitive data or inaccurate outcomes.

{{ f.t }}{{ f.d }}

Foundations of trust, as listed in the talk.

{{ f.n }} {{ f.t }} {{ f.d }} {{ f.a }}
Pattern {{ p.t }} {{ p.d }}
Trust
Value
Effort
Fig. 3.5 — The trust flywheel

“Trust is earned through interaction. Value must be perceived before it's imagined. Effort must be absorbed, rather than exported to users.”

Buzz Usborne, “Designing AI experiences people actually use”

Trust
willingness to invest effort
Value
motivation to engage at all
Effort
mental energy to achieve a goal

Trust isn't linear: it builds slowly and drops fast when a product proves fragile — and the best way to rebuild it is to recover quickly. Most AI products don't want blind reliance either. The goal is calibrated trust: people who know when to lean on AI and when to double-check it.

One answer is only the visible tip. Behind it sits context, tools, execution and verification — the harness.

Chapter 04 / 05

Harnesses & guardrails.

An AI harness is the infrastructure that gives a model tools, context and memory handling, guardrails, verification steps, approvals and recovery loops. We engineer the environment to reduce mistakes.

Interactive · Fig. 4.1 — Five ingredients of a harness
Deterministic rule Probabilistic judgement
Model
probabilistic core
{{ selNode.n }} · {{ selNode.s }}

{{ selNode.t }}

{{ selNode.d }}

Run log · illustrative task
  1. {{ l.n }}{{ l.m }}
    2.
Select a module to read about it, or run a task to watch information move through the harness.

Ingredients as listed in the talk. The run is an illustrative scenario. A guardrail lowers risk; it doesn't guarantee safety.

Poor-quality context
  • Poor data returns generic, poor results
  • Contradictions return confusing responses
  • Competing data adds noise and fluff
  • Many irrelevant details lead to degradation
  • Data overload overwhelms AI's “reasoning”
Relevant, orchestrated context
  • Task description and background data
  • Documents, attachments, spreadsheets
  • Guidelines, rules and templates to use
  • Examples — good and bad — patterns, specs
  • Conversation history, tools, limits and guardrails
“It's the careful, dynamic construction and maintenance of a model's information environment, so it can solve the problem reliably. Think of it as IA of AI prompts.”
Context engineering, as framed in the talk, citing Chroma's “Context Rot” report
Context isn't static

Context is everything AI considers when answering: sources, level of expertise, code, history. And it changes as people interact. Good interfaces help answer six questions:

  1. {{ q.n }}{{ q.t }}
    2.
{{ c.t }} {{ c.d }} {{ c.a }}
Fig. 4.2 — Setting up AI evals

Benchmarks pick a model. Product evals tell you if your product works.

Evals answer three questions every team has: did a change make things better or worse? Is it good enough to ship? Can we catch regressions before users do? The talk notes engineering teams often spend 60–80% of their time on error analysis and evaluation.

{{ e.n }} {{ e.t }} {{ e.stat }} {{ e.statLabel }} {{ e.d }}
Frequent failure categories
{{ f.n }}{{ f.t }}
AI judges are biased, too
{{ b.t }} — {{ b.d }}

Editor's noteThe talk recommends working bottom-up: sample 50–100 real inputs and outputs, categorise the failures, and let metrics emerge — because an eval score is evidence, not proof of real-world reliability.

Once a system can execute and verify work, a new question arrives: what happens when we let it act for us?

Chapter 05 / 05

Agentic experiences.

With agents, we move from replies to outcomes. An agent pursues a goal — it plans, chooses tools, takes multiple steps, observes results and adapts. So we design for trust, control, autonomy and recovery around actions.

{{ g.n }} {{ g.t }} {{ g.d }}

The six ingredients of an AI agent

“The most significant difference between both is autonomy. No need to give agents precise instructions — they figure it out as an ant hill.”
Greg Nudelman, UX for AI — on AI vs. AI agents
Interactive · Fig. 5.1

Autonomy is set per action, not globally.

Agentic AI needs at least three modes: suggest (propose only), confirm (ask before each step) and execute (auto-run with prior permission). Some actions are harmless, like creating a draft; others have consequences, like submitting a report. Pick a mode and run the agent.

Goal: file the Q3 churn report
  1. {{ s.title }} {{ s.stateLabel }}
    {{ s.detail }} · {{ s.policy }}
    {{ s.ask }}
Audit trail{{ statusText }}
  1. {{ l.t }}{{ l.m }}
    2.
Every action, approval and recovery will be recorded here.

Modes and per-action autonomy as described in the talk. The task and failure are illustrative.

Fig. 5.2 — Autonomy levels
{{ l.n }} {{ l.t }} {{ l.d }}

Many products span several levels depending on the task — a delegation loop where you hand over control, then take it back. Start at the lowest level by default and raise it only when it's earned.

“Agents should have partial autonomy — we should keep human in the loop, and whenever output is vague, flag it and ask for human's input.”
Andrej Karpathy, “Software Is Changing (Again)”

Agents also need access — to data, browsers, calendars, payment details — which raises serious privacy and security questions, as Signal's Meredith Whittaker warns. And the talk is blunt: most companies don't need AI agents at all. Often, simple automations or AI workflows are the more reliable choice.

Agentic UX challenges
Agents run in loops, with tools, memory and sub-agents. That creates new design problems.
  1. {{ c.n }}{{ c.t }}
    2.
“Book a ticket” — what an agent needs
{{ a }}

Meredith Whittaker, President of Signal, warns that agents need deep access across data, browser, apps, operating system and cloud — and that this raises severe privacy and security issues.

“You can either fit AI in existing jobs or you can fit jobs into and around AI. What doesn't really work is replicating structures, mistakes, limitations of the past with an entirely new paradigm.”
Tom Goodwin — on AI agent org charts
{{ n.n }} {{ n.t }}
Epilogue · One connected system

Designing AI is always about designing for trust.

Generation is the easy part now. The products that last will pair it with solid foundations, visible evidence, real guardrails and actions people can follow, approve and undo. The value of any technology lies in the real impact it brings to real people.

Key questions to ask before you add AI
  1. {{ q.n }}{{ q.t }}
    2.

Ten takeaways for designing AI.

Design patterns for AI — summary from the talk
  1. {{ e.n }}{{ e.t }}
    2.
Questions, answered

Frequently asked questions.

Short answers drawn from the story above and the talk it's based on.

What is the state of AI in 2027?
AI is out of the pilot stage: in the State of Product survey shown in Vitaly Friedman’s talk, 99% of product teams had fully integrated AI into at least one workflow. But while 80% of teams say AI helps them ship faster, customers aren’t benefiting any sooner — the bottleneck has moved to decision-making.
What is AI workslop?
AI workslop is low-effort, AI-generated work that masquerades as good work but lacks the substance to be useful, creating more work for colleagues downstream. The term is described by Kate Niederhoffer in Harvard Business Review.
What makes a design system AI-ready?
An AI-ready design system treats design decisions as infrastructure. Decisions flow into spec files for every component, a token layer gives AI a closed set of variables to choose from, an audit script catches what AI gets wrong, and a sync routine flags specs to update after shipping.
How do you design AI products people trust?
Trust doesn’t emerge by default. People trust systems they can query, shape, verify and override. Show evidence and missing context, flag uncertain claims, let people correct output, and aim for calibrated trust — knowing when to lean on AI and when to double-check it.
What is an AI harness?
An AI harness is the infrastructure around a model that gives it tools, context and memory handling, guardrails, verification steps, approvals and recovery loops. The talk lists five ingredients: context engineering, a tool registry, deterministic guardrails, agent loops, and verification with feedback.
What is context engineering?
Context engineering is the careful, dynamic construction and maintenance of a model’s information environment so it can solve a problem reliably — task, data, documents, rules, examples, history, tools and limits. Too little context produces vague answers; too much makes it hard for AI to focus.
What are AI evals?
AI evals are test suites that track AI behaviour over time. Each combines a dataset of 50–500 cases per capability, a runner that sends every case through the pipeline several times, and a grader — a human or an LLM-as-a-judge — whose agreement with humans is itself checked.
How much autonomy should an AI agent have?
Agentic AI needs at least three modes: suggest (propose only), confirm (ask before each step) and execute (auto-run with prior permission). Autonomy should be set per action, not globally — creating a draft is harmless, submitting a report has consequences.
Do most companies need AI agents?
According to the talk, most companies don’t: agents are less reliable and may produce unpredictable outcomes. Simple automations or AI workflows are often the better choice.
Taking on new projects

Building with AI? Let's design the parts that make it trustworthy.

I'm Prince Pal, a UI/UX designer. I wrote and designed this story because these are the problems I like working on: turning capable AI into products people understand, control and come back to.

{{ s.n }} {{ s.t }} {{ s.d }}

Tell me about your product.

A few lines is plenty. I reply within one business day.

Form not loading? Open it in a new tab

Sources & notes

This page is an editorial interpretation of The State of AI for 2027, a presentation by Vitaly Friedman (Maven, 15 October 2026). Quotations are attributed as they appear on the slides; statistics are reported results cited in the talk, some recovered from slide images by OCR. Interactive scenarios are illustrative and labelled as such. Narrative transitions and editor's notes are my own.

Vitaly's AI design patterns course Checklists
  • {{ s.t }} (opens in new tab) {{ s.a }}
  • Also quoted from the talk: Hiten Shah, Kate Niederhoffer (HBR), Arvind Narayanan & Sayash Kapoor, Dina Genkina (IEEE Spectrum), Austin Z. Henley, Nilay Patel, Pavel Samsonov, Morning Consult, Chroma (“Context Rot”), Luke Wroblewski, Mary Borysova, Buzz Usborne, Salesforce “State of the Connected Customer”, International Journal of Human-Computer Interaction, Google “Human-AI Interaction Guide”, David D. Woods, NN/g, Tey Bannerman, Ryan Rumsey, Addy Osmani, Greg Nudelman, Tom Goodwin, Meredith Whittaker, Andrej Karpathy.