Briefing AI Agents · Research, November 2026Nov 2026
Briefing AI Agents: Weshowpeople. Wetellagents.
AI agents are becoming the first teammate people turn to. But we still brief them by typing. We interviewed 547 people who already hand real work to AI agents about how they brief them, what gets lost, and how they check what comes back.
547 AI-moderated interviews
21 Oct to 2 Nov 2026
US and UK
Step through with your arrow keys, or scrollScroll to step through
02The show gap
This is one person we interviewed. One of 547 who already hand real work to AI agents.
Every dot is one of them. Here’s how they explain work.
78%
show colleagues the work. In person, in a meeting or in a recording.
14%
show agents the work on video. 82% type their instructions instead.
“With a colleague, I would ask them to come on a meeting, then I would do a screen share. With the agents, I would just type the message on the chat.”
Engineer, company of 50 to 249 people
1 person78% show colleagues14% show agents on video
03The handoff tax
Every task handed to an agent has two costs. Explaining it, then checking what comes back.
12hours
of a 40-hour week go on briefing and reviewing agents. About 30%, at the median.
Every finding, chart and table is below, with the methodology and the questions we asked. The PDF has the same report to keep or share, and every chart has a GIF to download.
Summary
Work goes to agents first now. We still brief them in text.
AI agents have become the first place many people take a task. Of the more than 500 people we interviewed, two in three started handing work to agents in the past year, and half now go to an agent before they go to a colleague.
The way we explain work hasn’t caught up. With colleagues, 78% show the work: in person, in a meeting or in a recording. With agents, 14% show it on video. Most type, and what’s hardest to type is context: the background, the house rules and how things should look.
That gap has a cost. We call it the handoff tax: the time spent briefing agents and checking what they hand back, about 12 hours of a 40-hour week at the median. The few who show agents the work say it helps.
The show gap
The difference between how people explain work to colleagues, by showing it, and to agents, by telling it: 78% vs 14%.
The handoff tax
The time it takes to brief an agent and review what it hands back: about 30% of the working week at the median.
AI agent
In this report, an AI tool that takes actions in other apps for you, such as updating a doc, a CRM or a ticket, or editing code. Not just a chatbot.
Key findings
Agents became teammates fast. The way we brief them didn’t change with them.
64% of people already working with AI agents started less than a year ago.
A fast shift: a third of all respondents started in the past six months.
49% go to an agent before a colleague when they need help with a task.
Among the heaviest agent users, it’s 71%.
82% type their instructions to agents, and 56% use typing alone, with no voice or video.
Voice is catching up at 28%. Only 14% use video.
76% now brief agents in more ways than when they started.
Half already show agents something visual, mostly as a screenshot.
14% show an agent the work on video.
With colleagues, 78% show the work: in person, in a meeting or in a recording.
30% of the working week goes on briefing and reviewing agents, at the median.
That’s about 12 hours of a 40-hour week. Over a third spend 40% or more.
55% say background knowledge is the hardest thing to get across.
35% say visual things: how something looks, or where it is on screen.
3 in 4 people who give agents recordings say they help more than text.
Especially with visual work: design, interfaces, step-by-step workflows and bugs.
Two paradoxes.
find agents easier to brief
yet
gave up at least once last month
Easier than a colleague, yet often abandoned.
59% say briefing an agent takes less effort than briefing a colleague. Yet 72% gave up on an agent at least once in the past month and did the task themselves. Even among those who find agents easier, 63% did.
spend less time reviewing
vs
spend more time briefing
Easier to check, harder to brief.
Since they started using agents, 47% spend less time reviewing what agents hand back. But 45% spend more time briefing them.
How this fits with other research.
Glean’s Work AI Index (June 2026) found workers spend 6.4 hours a week making AI usable. BCG’s AI at Work survey (June 2026) found 47% spend more time directing AI than doing the task itself. Both measure how much time goes on managing AI. This report looks at where that time goes when the AI is an agent: how people brief it, what gets lost, and what helps.
One engineer’s handoff
“When you have to give a lot of background context that comes from different folders, different videos, different environments, that’s going to be really difficult for an agent to understand.”
Engineer, company of 50 to 249 people
3 to 6 mo
since they started giving tasks to agents
Over half
of their work now involves an agent
5 to 15 min
to brief a single task
Up to half
of recent agent tasks needed significant correction
Most of their tasks now go to an agent first. Briefing one means working out the right prompt and what information to give it, and the background is the hardest part. With a colleague, they would share their screen in a meeting. With an agent, they type. Some jobs, like QA testing and deployments, they still do themselves. When they have given an agent a recording, it was “a lot more useful”, carrying the back and forth from a meeting and technical details that usually get lost in text.
A real respondent, anonymised. Quotes lightly cleaned up.
01Adoption
Agents are joining the team. Most started in the past year.
Two in three respondents began handing tasks to AI agents within the past year. Already, half go to an agent before they go to a colleague.
49% go to an agent before a colleague when they need help with a task.
An agent
Half and half
A colleague
Myself
49%18% almost always, 32% mostly
30%Split it between both
12%Mostly or always a colleague
9%Did it all themselves
Almost always an agentMostly an agentOne dot is one person in 100
Question: When you need help getting a task done at work, who do you usually hand it to? n=503.Download GIF
64% started giving tasks to agents in the past 12 months.
64% in the past 12 months
36%Over a year ago
31%6 to 12 months
23%3 to 6 months
11%Under 3 months
Now
Question: When did you start giving tasks to AI agents at work? n=547. Everyone surveyed already uses agents, so this shows how new this group is, not how fast adoption is growing overall.Download GIF
71% of the heaviest agent users go to an agent first, vs 12% of the lightest.
The more of their work involves agents, the more often agents come first.
Share of their tasks that involve an agentOut of 10 people
Under 10%
12%
10 to 25%
38%
25 to 50%
48%
50 to 75%
66%
Over 75%
71%
Share who hand tasks mostly or always to an agent, by share of their work tasks that involve an agent. n=33, 129, 177, 116 and 48 from left to right. 32% involve an agent in at least half their tasks (n=546).Download GIF
02The show gap
With people, we show. With agents, we tell.
To explain a task to a colleague, 78% show the work: in person, in a meeting or in a recording. With agents, only 14% show it on video. Voice is catching up, but showing hasn’t.
The show gap: 78% vs 14%.
Briefing a colleague
78%show the work, in person, in a meeting or in a recording
Briefing an agent
14%show the work on video
“With a colleague, I would ask them to come on a meeting, then I would do a screen share. With the agents, I would just type the message on the chat.”
Colleagues: showing them in person, sharing a screen live or in a meeting, or recording a video (n=500). Agents: giving the agent a video or screen recording in recent briefs (n=512).Download GIF
With agents, typing goes up and showing all but disappears.
How people explain a task to a colleague (n=500) vs to an agent (n=512). People could pick more than one way. Voice means a call or voice note with a colleague, and speaking or dictating instructions to an agent. For agents, showing means giving it a video; screenshots are counted separately, below. 56% give agents instructions by typing alone, with no voice or video.Download GIF
Most people start by typing. The longer they use agents, the less they only type.
76% say they now brief agents in more ways than when they started, rising from 71% of the newest users to 77% after a year.
67%
26%
Under 3 months
61%
24%
3 to 6 months
55%
29%
6 to 12 months
50%
31%
Over a year
Time using agents →
Type onlyUse voiceCircle areas are to scale
By time using agents, n=510 (55, 113, 151 and 191). Type only means typed instructions with no voice or video. Groups are compared at one point in time, not tracked over time.Download GIF
The visual gap.
What’s hard to explain is often visual. People try to show it, but mostly with a still image.
35%say visual things are among the hardest to get across to an agent
→
47%send agents a screenshot, a single still moment
→
14%use video, which shows the flow
Roughly three in four people who show agents something visual send only screenshots, not video.
Hardest to get across: up to three picks, n=514. Screenshots and video: what people gave agents in recent briefs, n=512. 51% gave agents something visual.Download GIF
03The handoff tax
The handoff tax. Briefing out, reviewing back.
Every task handed to an agent has two costs: explaining it, then checking what comes back. Together they take about 30% of the working week at the median, about 12 hours of a 40-hour week.
Briefing is taking longer. Reviewing is getting quicker.
45% spend more time briefing agents than when they started. 47% spend less time reviewing their work.
47%
spend less time reviewing agents’ work than when they started. The bottleneck has moved upstream, to the brief.Only 25% spend more time reviewing; 45% spend more time briefing. n=508.
Time spent compared with when they started using agents. n=508.Download GIF
1 in 5 now has another AI check the agent’s work.
Almost nobody skips review: only 1% say they didn’t really check. But reviewing is getting quicker, and AI is starting to do some of it. Twice as many people ask another AI to check as ask a colleague. People use a second AI as a first pass, then check the result themselves, especially for anything that goes to a customer.
9% asked a colleague1% didn’t really check
In their words
“It’s just easier to have another AI agent check it. LLMs have come so far, so I’m hoping if one missed it, the other will pick it up.”
Product manager, company of 1,000+ people
“First run it through another AI for a sanity check. And then I skim.”
Product manager, company of 1,000+ people
Thorough checkers catch more: 32% of people who check everything thoroughly report heavy rework, vs 14% of people who only skim. Checking closely is how problems get found.
Question: Think about the last few times an agent handed new work back to you. How did you check it? Multiple choice, n=508. Heavy rework: a quarter or more of recent agent tasks needed significant correction; thorough n=331, skim only n=133.
44% take five minutes or more to brief a single task.
People with longer briefs mostly describe three things: giving the agent enough context, writing a precise prompt, and breaking down multi-step tasks.
Under 1 minute5%
1 to 5 minutes40%
5 to 15 minutes31%
15 to 30 minutes9%
Over 30 minutes4%
It depends12%
In their words
“It’s basically just feeding it all the context. If there’s documents, Slack threads, Notion documents, Confluence documents, all that goes into feeding the content.”
Product manager, company of 50 to 249 people
“Because it’s typically a complicated multistep task, I need to make sure that the agent has all the information and the detailed protocol necessary.”
Founder, company of 50 to 249 people
Long briefs go with giving up: 79% of people whose briefs take 15 minutes or more did a task themselves in the past month, vs 67% of those whose briefs take under five.
Question: How long does it typically take you to brief an agent on a single new task, from deciding to give it to an agent to sending your first message? n=524. Giving up: n=68 and 231.Download GIF
12 of 40 hours go on briefing and reviewing agents.
That’s the median week. 37% spend 40% or more of their week on it, and the more work goes to agents, the bigger the tax.
MonTueWedThuFri
The median weekOne square is one hour.
The median week, by how much of their work involves an agent
10%Under 10% of tasks
23%10 to 25%
30%25 to 50%
33%50 to 75%
55%Over 75%
How much of their work involves an agent →
Median share of the working week spent briefing and reviewing agents, by share of tasks involving an agent. Shares are estimated from answer ranges (see methodology). n=509; the heaviest group is 49 people.Download GIF
The steps are the tip. Context sits below the surface.
Few people find the steps hard to explain. What’s hard is everything underneath: the background, the house rules, the edge cases and how things look. Each layer is sized by the share who say it’s among the hardest to get across.
Above the surface
The steps to follow17%
Below the surface
What a good result looks like32%
Visual thingsHow something looks, where it is on screen35%
Exceptions and edge cases35%
How we do things hereConventions, style, rules40%
Background knowledgeWhy it matters, what’s been tried55%
Missing context goes with rework.
Share reporting heavy rework, out of 10 people.
Background is hardest to brief
33%
Everyone else
20%
Question: What’s hardest to get across to an agent in a brief? Up to three picks, n=514. Heavy rework: a quarter or more of recent agent tasks needed significant correction (27% of all respondents, n=503).Download GIF
72%
did a task themselves in the past month because explaining it to an agent took too long.13% do this often. n=517.
Showing helps. Especially when the work is visual.
3 in 4 people who give agents videos as context say they help, especially for explaining visual things.
Most say video is clearly more useful than text.
3 in 4
people who give agents recordings say they beat text alone.
Clearly more useful than text ~74%Useful for some tasks ~21%No better than text ~5%
One tile is about 5% of people who use recordings with agents.
Compared with text alone. Base: about 150 people who use recordings with agents and gave a clear answer. Answers coded by hand, so figures are approximate.Download GIF
Where it helps most.
Design, interfaces and UI
“Screen recordings eliminate tedious written explanations for UI interactions.”
“I had to fix an issue. Then I can’t describe the location properly.”
Data and dashboards
“Information, data, spreadsheets, that kind of stuff.”
Meetings as context
“I needed to review something for my meeting where we’ve recorded it. It can watch the video and give me an answer.”
Themes from respondents’ answers.
Why they say it works.
“Some things you can’t explain with just words.”
“Text alone can be a bit ambiguous, given a screen recording showing exactly what I’m talking about.”
“I can quickly do a screen grab rather than typing everything. It saves a huge amount of time.”
05Barriers
Habit and friction. Why people don’t show agents more.
Most people know agents can take recordings. Few use them, and the reasons are habit and effort, not doubt about what agents can do.
42% use recordings with agents. 32% know they can but don’t. 27% didn’t know.
1 in 3 people using agents for under six months didn’t know it was possible (33%, n=153).
Even those who use recordings rarely reach for them by default: only 14% gave an agent a video in their last few briefs.
Question: Did you know you can give AI agents screen recordings as context? n=474. Newest users: n=153.Download GIF
Of those who know but don’t, most say they’ve never needed it, or it’s too much effort.
Only about one in twenty doubt agents can read video.
~37%Never needed it, not a habit
~19%Too much effort
~10%Privacy or policy
~10%Text or screenshots are enough
~5%Don’t know how
~5%File size or cost
~5%Doubt agents can read video
Main reason, one each, from the 151 people who know agents can take recordings but don’t use them. Coded by hand, so figures are approximate. Square areas are to scale. Too much effort means recording, editing, saving and uploading. Privacy includes confidentiality and company policy.Download GIF
“I use AI to speed things up. Doing a screen recording of something I’m trying to explain to it seems counterintuitive.”
“When you record a screen, you capture everything. Not just the Excel file you want help with.”
Looking ahead
What to expect in 2027. Three signals in the data.
This is Supercut’s view, based on how people who’ve used agents longest already brief them, compared with newer users.
Signal 1
Context moves out of the prompt and into the setup.
31%→44%
As people get used to agents, they stop retyping context and set it up once. 31% of new users give agents saved instructions or skills, rising to 44% of those using agents for over a year. Connected tools rise from 29% to 35%.
Signal 2
Voice keeps closing the gap.
26%→31%
26% of new users speak or dictate their briefs, rising to 31% after a year. With agents, voice is already close to how people talk to colleagues: 28% vs 32%.
Signal 3
Visual briefs move from stills to video.
47%vs14%
Nearly half already give agents visual context, but mostly as screenshots, and video use hasn’t grown much with experience. Most who use video say it beats text, and 27% didn’t know agents can take a recording. As that changes, we expect video to follow.
New users: under 3 months (n=55). Over a year: n=191. Groups are compared at one point in time, not tracked over time.
Methodology
How we ran the research. Who we spoke to, and what we asked.
Supercut ran AI-moderated interviews with Typeform Research Flow between 21 October and 2 November 2026. Respondents in the US and UK were recruited through research panels (Prolific and User Interviews). An AI interviewer asked each structured question, then followed up with open questions about the answer.
547
interviews analysed after screening
499
completed in full; 48 partial, counted for the questions they answered
99%
held by voice (539 voice, 8 text)
US and UK
respondents, recruited through research panels
Where we interviewed.
547 people across two markets. Maps not to the same scale.
United StatesUnited Kingdom
Screening
We only interviewed people already handing real work to AI agents. The first question asked how they had used AI tools at work in the past month. Only one answer qualified; 2 people who gave a different answer were screened out.
Which of the following best describes how you’ve used AI tools at work in the past month?
I’ve had an AI tool make changes for me directly in other apps, such as updating a doc, spreadsheet, CRM or ticket, publishing or sending something, or editing code.Qualified
I’ve mostly chatted with an AI assistant to get answers, ideas, or drafts that I then used myself.Screened out
I haven’t used AI tools for work in the past month.Screened out
Role
n=499
Engineering
24%
Product
14%
Operations
13%
Data
9%
Founder
8%
Marketing
7%
Sales
7%
Design
6%
Support
5%
Finance, HR, legal, security
7%
One dot is 1% of respondents.
Company size
Employees, n=499
4%Just me
15%2 to 9
17%10 to 49
23%50 to 249
14%250 to 999
26%1,000+
Building size shows company size; the number is the share of respondents.
Time using agents
n=547
Under 3 months
11%
3 to 6 months
23%
6 to 12 months
31%
Over a year
36%
One dot is 1% of respondents.
The questions we asked.
In order, after the screening question. Each was followed by open questions from the AI interviewer.
When did you start giving tasks to AI agents at work?
Roughly what share of your work tasks do you involve an AI agent in today?
How long does it typically take you to brief an agent on a single new task, from deciding to give it to an agent to sending your first message?
In the past month, have you done a task yourself because explaining it to an agent took too long?
What’s hardest to get across to an agent in a brief? Pick up to three.
Think about the last few times you briefed an agent. How did you give it the task, and what did you give it to work with? Select all that apply.
Compared with when you started working with agents, do you now use more ways to brief them, for the prompt or the context?
In a typical work week, roughly what share of your time goes on briefing agents (explaining tasks, giving context, follow-ups)?
In a typical work week, roughly what share of your time goes on reviewing and correcting agents' work?
Compared with when you started using agents, how has your time spent briefing agents changed?
And your time spent reviewing and correcting their work?
Think about the last few times an agent handed new work back to you. How did you check it?
Out of the last tasks you gave an AI agent, how many needed significant correction or re-doing?
When you need help getting a task done at work, who do you usually hand it to?
When you explain a task to a colleague (today, or before you used agents), how do you usually do it?
Compared with briefing a colleague on the same task, briefing an agent takes more, less or about the same effort?
Think about briefing a colleague versus briefing an agent on the same task. What’s the biggest difference in how you do it?
Did you know you can give AI agents screen recordings as context?
Is there anything you can easily give a colleague that’s hard to give an agent?
How many people work at your company?
Which best describes your role?
Margin of error
For results based on all respondents, the margin of error is about ±4 percentage points at 95% confidence. It is wider for smaller groups, and figures for groups under 50 people are directional.
Who funded it
Supercut funded and designed this research. Supercut was not named in any question.
How we calculated the figures
Percentages are of the people who answered each question; n is shown with every figure. Select-all questions can add up to more than 100%. Shares of the working week are estimated from the midpoint of each answer range. “Product and engineering” means engineering and product management roles (n=189). Figures may not add up to 100% because of rounding.
Coded answers
How useful recordings are, and why people don’t use them, come from open answers coded by hand to one main theme each. Those figures are approximate. Quotes are lightly cleaned up for readability and respondents are anonymous.
Limits
Answers are self-reported. Changes over time are as respondents remember them, not tracked. Everyone interviewed already uses AI agents, so results describe that group, not all workers. Figures for groups under 50 people are directional.
Data
Full tables are available to journalists on request.
Authors
Research and analysis by Cristina G Avila, Supercut.
Appendix
Results by role and company size. The same measures, cut by group.
Red marks the highest value in each column.
By role.
Group
n
Go to an agent first
Did a task themselves
Briefing takes longer
Heavy rework
Use video
Type only
Week on agents
Engineering
121
62%
71%
46%
28%
9%
68%
33%
Product
68
60%
76%
47%
37%
19%
51%
30%
Operations
63
41%
75%
44%
24%
14%
59%
23%
Data or analytics
46
41%
83%
30%
30%
20%
48%
30%
Founder
41
54%
73%
46%
27%
12%
46%
33%
Marketing
34
32%
62%
65%
21%
12%
65%
30%
Sales
34
38%
62%
38%
9%
12%
50%
18%
Design or UX
31
48%
71%
55%
45%
19%
48%
40%
Other roles
61
36%
64%
36%
21%
11%
49%
23%
All respondents
547
49%
72%
45%
27%
14%
56%
30%
By company size.
Group
n
Go to an agent first
Did a task themselves
Briefing takes longer
Heavy rework
Use video
Type only
Week on agents
1 to 49 people
181
50%
73%
43%
30%
13%
57%
33%
50 to 999 people
187
50%
73%
49%
25%
14%
51%
30%
1,000+ people
131
47%
67%
41%
26%
15%
61%
23%
All respondents
547
49%
72%
45%
27%
14%
56%
30%
n is the number of people in each group; some questions were answered by slightly fewer. Groups under 50 people are directional. Go to an agent first: mostly or always. Did a task themselves: at least once in the past month because explaining it took too long. Briefing takes longer: than when they started. Heavy rework: a quarter or more of recent tasks needed significant correction. Use video: gave an agent a video in recent briefs. Type only: typed instructions with no voice or video. Week on agents: median share of the week spent briefing and reviewing agents, estimated from answer ranges.
FAQ
Questions about briefing AI agents. Answered with the data.
How do most people brief AI agents?
Mostly by typing. 82% type their instructions to agents, and 56% use typing alone, with no voice or video. 28% speak their instructions, and 14% give agents video.
Can you give an AI agent a screen recording?
Yes. Record your screen with Supercut and share the recording with your AI agent. Agents like Claude, ChatGPT and Cursor connect to Supercut over MCP and read the recording as context: its transcript, comments and frames.
How much time do people spend briefing and reviewing AI agents?
About 30% of the working week at the median, roughly 12 hours of a 40-hour week. More than a third spend 40% of their week or more.
What’s hardest to explain to an AI agent?
Background knowledge: why the task matters and what’s already been tried (55%). Then how things are done here (40%), exceptions and edge cases (35%), and visual things like how something looks or where it is on screen (35%).
Who was surveyed for Supercut’s Briefing AI Agents report?
For Briefing AI Agents: We show people, we tell agents, Supercut interviewed 547 people in the US and UK who had an AI agent make changes directly in a work app in the past month. The interviews were AI-moderated and run with Typeform Research Flow from October 21 to November 2, 2026.
For journalists
Using this research. Charts, data and how to cite.
Charts
Every chart above has an animated GIF to download, and the full report (PDF) has them all.
Data
Full tables are available to journalists on request.
Supercut is fast, polished screen recording for business communication. Video shows the work, and agents move it forward: AI agents like Claude, ChatGPT and Cursor connect to Supercut over MCP and read recordings, transcripts, comments and frames.
How to cite this report
Supercut (2026). Briefing AI Agents: We show people, we tell agents. supercut.ai/blog/briefing-ai-agents
You’re free to quote and republish these findings with credit and a link (CC BY 4.0).