Tool fatigue isn’t a tooling problem — it’s a selection problem, and the agent-lens solves it: a tool earns a slot only if it improves what an agent can perceive, compute, learn from, need, or change in your design work. The question this post asks is what tools an agent needs, and how a team can tell which ones earn a place.
The average designer now uses 7 off-the-shelf AI tools, up from 3 a year earlier, while 91% use AI for design work at least weekly, up from 54%. Nearly half still haven’t settled on go-to tools, according to the State of AI in Design tools chapter. The path here: read the numbers, examine why tool selection breaks, then apply PROVE and the agent-lens to decide what stays.
The numbers: adoption is up and selection is unfinished
The AI toolstack is growing faster than designers are settling on a stable set of tools. The average rose from 3 tools in 2025 to 7 in 2026, while weekly AI use climbed from 54% to 91%, according to the State of AI in Design tools chapter. The report’s own headline on confidence is “Confidence in tools is up, but so is the noise.” Asked how their confidence in choosing tools compares with a year ago: 37% are much more confident and have clear go-to tools, 49% are somewhat more confident but still searching, 6% are about the same, 6% are less confident because it feels more overwhelming than a year ago, and 2% were not using AI tools a year ago. The overwhelmed 6% is the closest the survey comes to measuring fatigue directly; the rest measure adoption and confidence.
The survey included 906 designers in more than 60 countries and was fielded in March 2026 by AI in Design, a project of Designer Fund and Foundation Capital; the survey’s about page gives its scope. The authors say the sample may skew toward respondents who have embraced AI more than the average designer, and present findings as directional rather than absolute benchmarks.
Anthropic, the maker of Claude, is a partner of the AI in Design 2026 survey; the survey was conducted independently of its partners. The tools chapter reports that Claude is used by 78% of respondents, ChatGPT by 65%, and Gemini by 48%; Claude overtook ChatGPT as the primary general AI tool. Anthropic launched Claude Design just after the survey closed, and the report says it will track that tool’s adoption. These figures describe reported use, not proof that any tool belongs in a particular team’s stack.
Figma’s own vendor research, its State of the Designer 2026, reports that 91% say new AI tools improve their designs, 89% say they work faster, and 80% say they collaborate better. Those are Figma’s findings from a NewtonX survey of 906 designers, not a neutral industry benchmark. See Figma’s blog summary and its full report page.
Why a wider stack is harder to choose from
The sources above show that a wider stack is harder to manage, but they do not measure the cost of managing it. The reasoning in this section is the author’s, not a survey finding. Tools can overlap: two may accept similar inputs or produce similar outputs. Each addition then brings an evaluation question: does it fit the workflow, meet the quality bar, and save enough effort to justify its place? Moving between tools also asks people to carry context across different interfaces.
This is a selection framework, not a claim that every team experiences the same cost. For an individual stack, make the cost visible: list what each tool receives, what it returns, where its output goes, and what a person must do before the result is usable.
The old way we picked tools (and why it broke)
A convincing demo is weak evidence of reliable design output. Nielsen Norman Group (NN/G) compared AI prototyping tools with designs made by its own designers, using heuristic evaluation, and found that prompt detail and design artifacts materially affected accuracy. Even the best outputs had visible shortcomings; the authors summarized the result as “good from afar, but far from good” in their AI prototyping evaluation. The selection signal should be fit for real work, not first impression.
NN/G’s testing methodology describes the study design: three phases covering prompt writing, generation, and heuristic assessment. Broad prompts produced widely varying outputs. Detailed prompts brought results closer to the human design, while prompts that included artifacts such as a Figma frame link produced the most accurate translation. A separate NN/G status update from May 2025 found that narrow-scope AI features, not full-design generators, were the ones designers most often kept in everyday work, and that most design-specific AI could not yet match human designers’ output quality (NN/G, May 2025).
Even the strongest outputs showed weak visual hierarchy, overused colors, poor contrast, and inconsistent margin spacing. That matters when a tool’s demo is judged on a polished image but the real task involves translating requirements and artifacts into usable design work. Surface quality is not the same as dependable output. Evaluate against the work the team actually needs done.
PROVE: a framework that already exists
PROVE is an existing way to decide whether an AI tool deserves a place in a team’s stack, or whether “no” is the smarter call, as described on the Future Product Days PROVE session page. The framework was developed at Nielsen Norman Group by Caleb Sponheim, NN/G’s AI Lead.
The page names three evaluation questions: workflow alignment, quality and reliability against human standards, and true efficiency, including hidden time, oversight, and cognitive cost. PROVE is not a reason to add AI by default; it is a way to assess whether a tool’s contribution merits its cost.
The agent-lens in this post extends that decision with five questions about what a tool makes possible for an agent. Use the two together: PROVE asks whether the tool fits and pays off; the agent-lens asks what it can see, do, learn from, require, or durably change.
The agent-lens: five questions that replace “do I like this tool?”
Reported adoption does not tell a team what an agent can usefully do. Claude’s 78% reported use, in the tools chapter, says little about whether it belongs in a given work loop. Ask what the tool contributes to an agent’s work loop, then test the answer against the team’s actual requirements.
-
Perceive — what can an agent see or intake through this tool? Can it receive relevant design context, or only a narrow prompt? Identify the input it can access and what must be carried in manually.
-
Compute — what can it do with that input? Name the concrete design operation or output. “It uses AI” is not a capability; a useful answer describes what the agent can produce or transform.
-
Learn from — what feedback does the tool return that improves future work? Look for a signal that helps an agent or its operator judge whether an output worked. If all the tool returns is an artifact with no useful feedback, future work may not improve through this tool.
-
Need — what does the agent require from the human? Account for the brief, decisions, corrections, oversight, and cleanup a person must provide. Human input is not automatically a flaw; hidden or repeated effort is what needs to be made visible.
-
Change — what in the design file, system, or workflow does the tool durably change? Distinguish an output that can inform the next step from a durable change to design work. State exactly where the result lands and what must happen for it to persist.
Rule of thumb: a tool scoring on fewer than three of the five is a novelty, not infrastructure. This is a decision rule for the framework, not a finding from the survey.
The decision table: what earns a slot
A tool earns a slot when its contribution survives questions about context, output, effort, and lasting change. Figma’s vendor research reports that 91% say new AI tools improve their designs, as described in Figma’s State of the Designer 2026; that is a useful signal, not a substitute for the criteria below.
| Criterion | Question to ask | Red flag | Source/anchor |
|---|---|---|---|
| Perceives context | What relevant input can the agent receive? | The context has to be recreated or re-entered each time. | NN/G AI prototyping evaluation |
| Computes real design output | What concrete design task can it produce or transform? | The output looks persuasive but misses the actual task. | NN/G AI prototyping evaluation |
| Returns learning signal | What feedback helps judge or improve later work? | It returns an artifact but no useful feedback. | Framework criterion (this post) |
| Human overhead needed | What must a person supply, check, correct, or clean up? | Oversight and cognitive effort are invisible in the efficiency claim. | PROVE framework |
| Durable change to files or systems | What does the tool change, and where does that change persist? | The result is isolated from the design file or workflow. | Framework criterion (this post) |
| Survives PROVE evaluation | Does it align with workflow, meet human standards, and deliver true efficiency? | It fails any of those evaluation questions. | PROVE framework |
| Team-shareable | Can the team understand and assess the tool’s role consistently? | Its value depends on one person’s undocumented process. | Framework criterion (this post) |
| Cost of keeping vs. dropping | What ongoing effort does the tool justify compared with removing it? | No one can name the benefit that would be lost if it disappeared. | Framework criterion (this post) |
To run a tool audit in one afternoon, first list the tools people actually use and the workflow each is meant to support. For each one, fill in the table with a specific input, output, human requirement, and durable change, or mark the answer unknown. Apply PROVE’s three questions, then score the five agent-lens questions. Compare overlaps, discuss any uncertain claims, and record a keep, investigate, or remove decision with its reason. This is a proposed review process, not a report of an audit we conducted.
What agents will need next (and what to prune now)
A wider stack is not automatically a better one. The average toolstack rose from 3 to 7 tools between 2025 and 2026, and designers’ go-to sets are still forming. A grounded next step is to favor fewer, deeper integrations and review output-only tools with no learning signal first.
Some tools are rebuilding the canvas so agents can read it directly: Paper renders its canvas in HTML and CSS, which the tools chapter describes as connecting the canvas to agentic tools. Prefer a tool that can receive relevant context, return something useful, and take part in a durable workflow over a collection of disconnected outputs. Keep tools that have a clear role and pass the criteria; investigate ones whose value is uncertain.
Beyond this framework, a related question is how agents work. The organizing idea may shift from tool-per-task to tool-per-loop. That does not mean every loop needs another product. It makes the agent-lens more useful: judge whether each tool contributes to the loop, and whether it helps an agent or person carry work forward.
FAQ
Isn’t 7 tools just the new normal?
It is the reported average toolstack in the 2026 State of AI in Design survey, not a target every designer should match. The same survey says nearly half of designers are still searching for their go-to tools. Treat seven as a reason to review fit, not as a quota for your team.
How is this different from just using PROVE?
PROVE asks whether a tool aligns with the workflow, meets human standards for quality and reliability, and delivers true efficiency after hidden time and oversight are counted. The agent-lens adds five questions about what an agent can perceive, compute, learn from, need, and change. Use both; neither requires adding a tool.
Should I wait for tools to consolidate before choosing?
You do not need to wait to make a decision about your current stack. Record what each tool contributes and where its value is unclear, then keep, investigate, or remove it using the criteria here. This approach does not depend on predicting which products will consolidate or when.
What’s the fastest way to cut my toolstack this week?
List the tools you use, then ask what each one uniquely receives, produces, teaches, requires, and changes. Flag overlaps and tools with no clear output or durable role. Apply PROVE before removing anything the team depends on, and record the reason for each decision so the change is reviewable.
Did you test these tools yourselves?
No. This review is based on published sources: the State of AI in Design survey, NN/G articles, Figma’s vendor research, and the PROVE session page. We did not run any of these tools hands-on. We did not test, query, or evaluate products. The criteria synthesize those sources; they are a decision framework, not comparative test results.
The Bottom Line
The growing stack and the continuing search for go-to tools make addition an easy default, not a reasoned decision. This framework’s recommendation is to require a tool to pass the criteria before it keeps its slot, and to remove or investigate the ones that fail. Use the five agent-lens questions alongside PROVE as the subtraction criteria: keep tools with a clear, dependable contribution, and question the rest. Audit your toolstack against the table this week.
How This Guide Was Built
This review is based on published sources: the State of AI in Design survey of 906 respondents, with Anthropic disclosed as a survey partner; NN/G articles, including its testing methodology; the PROVE session page; and Figma’s State of the Designer 2026, explicitly treated as vendor research. We did not run any of these tools hands-on. We conducted no testing, ran no queries, and evaluated no products.