All Briefs
Mission Brief · 014—

by Łukasz Graliński

Everyone, every month: how do you really work with AI?

Three months ago we stopped guessing. Since then, once a month, we have invited the whole company — engineers, QA, delivery, leadership, sales, ops — to answer the same seventeen questions in four minutes, honestly rather than aspirationally. No scorecard, no ranking; individual answers stay with the people who run the pulse. Three full editions later we have a trend, not a snapshot: the share of people running agents on whole tasks went from 38% in June to 69% now. Here is the third edition in full, with the two before it — the climb, the hours, and the turbulence.

69%
at Crew or Mission control: an agent runs whole tasks, the person plans and verifies
5.8 : 1
hours AI gave back vs hours it burned in the two weeks before the edition, 77 people — we count both
49%
name the same blocker: no time to learn
Open the pulse board
Pulse board · interactive

The whole company on one board.

Every dot is one person on the five-level ladder we use in the pulse. Switch departments to see who climbs and who is still buckling in, flip June to now to watch three editions of the climb, and read the vitals underneath: hours back against hours lost, rate limits, trust, workflows, blockers and mindset. Three monthly editions so far: 58, 66 and 77 answers out of about 130 people.

Pirxey / AI Pulse

Flight vitals

Aug 28 – Sep 8, 2026 · 77 answers from about 130 people

Whole company, 77 answers: 69% at Crew or Mission control, 5.8 hours back for every hour lost.

Every month · 4 minutes · a tool for growth, not evaluation. Departments overlap; their counts do not add up. Source and method

Launch ladder

Latest edition

One dot = one answer, at its highest selected level. Hover, focus or tap a band for its count.

Three editions

History changes this ladder only. All other panels show the latest edition.

  1. Junen = 58
  2. Julyn = 66
  3. Aug 28 – Sep 8n = 77

Dots compare distributions between editions; they do not track the same people over time.

Vitals

Hours back vs hours lost

5.8:1

82% came out ahead in the two weeks before the edition

Back912 h

Lost156 h

Estimated from answer buckets. Both gains and losses count.

Share of work with AI

71%

use AI for 60% or more of their work

Read distribution
  • 0–20%4 / 77 · 5%
  • 20–40%5 / 77 · 6%
  • 40–60%13 / 77 · 17%
  • 60–80%20 / 77 · 26%
  • 80%+35 / 77 · 45%

Rate limits

9%

hit limits daily or call them a bottleneck

Read distribution
  • Never9 / 77 · 12%
  • Occasionally46 / 77 · 60%
  • A few times a week15 / 77 · 19%
  • Daily3 / 77 · 4%
  • The limit is my bottleneck4 / 77 · 5%

Trust

62%

calibrate trust per task or require evidence

Read distribution
  • Trust little, check most12 / 77 · 16%
  • Trust the repeatable, verify the new17 / 77 · 22%
  • Calibrate trust per task34 / 77 · 44%
  • Evidence-based trust (tests, review)14 / 77 · 18%

Own workflows

47%

have multi-step workflows or automations

Read distribution
  • None, manual5 / 77 · 6%
  • Reusable prompts34 / 77 · 44%
  • Own multi-step workflows12 / 77 · 16%
  • Partial automations16 / 77 · 21%
  • Full automations8 / 77 · 10%
  • Not answered2 / 77 · 3%

2 skipped this question; shares still use all 77 answers.

Model choice

74%

choose models consciously or go further

Read distribution
  • Don't know which model3 / 77 · 4%
  • One model for everything17 / 77 · 22%
  • Switch models consciously26 / 77 · 34%
  • Pick by task, cost, limit21 / 77 · 27%
  • Test new ones, tell the team10 / 77 · 13%

Turbulence

Rule of the pulse: the top blocker gets one intervention before the next edition.

This edition: protected time to learn.

  1. No time to learn49% (38/77)

  2. No time to understand what others generated31% (24/77)

  3. Output quality, trust31% (24/77)

  4. Habit: doing it the old way25% (19/77)

  5. AI slop from others22% (17/77)

  6. Tool limits and cost22% (17/77)

  7. Don't know where to start13% (10/77)

  8. Permissions, client data8% (6/77)

  9. Nothing blocks me14% (11/77)

Multiple choices allowed; shares do not add up to 100%.

Mindset

Opportunity or threat?

3.53 / 5

1 = mostly worries, 5 = can't wait

Read distribution
  • 10 / 77 · 0%
  • 28 / 77 · 10%
  • 331 / 77 · 40%
  • 427 / 77 · 35%
  • 511 / 77 · 14%

Does Pirxey give you the conditions?

4.03 / 5

Rated 1–5; mean of the answers shown below

Read distribution
  • 12 / 77 · 3%
  • 23 / 77 · 4%
  • 310 / 77 · 13%
  • 438 / 77 · 49%
  • 524 / 77 · 31%

Show your setup to the team?

22%

say yes, now or with time to prepare

Read distribution
  • Gladly10 / 77 · 13%
  • Yes, give me time7 / 77 · 9%
  • Maybe, depends on format25 / 77 · 32%
  • Prefer to watch for now35 / 77 · 45%

Knowledge flow

45%

share both ways or teach others

Read distribution
  • Not yet11 / 77 · 14%
  • Someone helped me12 / 77 · 16%
  • Helped someone who asked19 / 77 · 25%
  • Both ways, I share what I find28 / 77 · 36%
  • Run demos or build skills for others7 / 77 · 9%

On our screens

Multiple choices · scroll each row for more

tools

Claude Code 75%Browser chat 55%Codex CLI 45%Cursor 31%Copilot 16%

models

Opus 5 74%GPT-5.x / Codex 68%Sonnet 5 62%Fable 5 39%Gemini 3 Pro 23%Grok 4 19%Haiku 4.5 14%Local models 9%

Source and method

Source: Pirxey’s AI Pulse survey, Aug 28 – Sep 8, 2026. Aggregated self-reports; individual answers are visible only to the survey team. Percentages use the selected group’s n, including skipped answers. These are respondents, not the whole workforce.

Hours are estimates from bucket values of 0, 1, 3.5, 10 and 20 hours; ratios use totals before rounding. Earlier editions use the same buckets and the same ladder definitions.

Eight findings · edition three

Agents, not autocomplete. And a few surprises.

The pulse asks the same seventeen questions every month, so the interesting part is not one headline number but the shape underneath it and how it moves. These eight came out of the Aug 28 – Sep 8 edition. Company-wide percentages use the 77 respondents; department figures use that department's respondents.

01
38% → 69% at Crew or above

Most respondents run agents

In June, 22 of 58 respondents said an agent executes whole tasks while they set the plan and verify the result. In July, 34 of 66. Now it is 53 of 77, and 22 of them run several agents or automations in parallel while doing something else. Two years ago almost all of us used AI as a better search box.

— Puls AI · ladder question, highest level ticked, three editions
02
912 h back · 156 h lost

Hours, in both columns

In the two weeks before the edition, across 77 people, using the midpoints of the answer buckets. 82% ended those two weeks ahead, up from 66% in June and 74% in July; the rest did not, and that remainder holds zeros as well as losses. QA reports the highest ratio (7.2 to 1, 12 answers), and operations, sales, business analysis and marketing the lowest (3.4 to 1, 5 answers). We publish both columns because a survey where everyone wins is worth nothing.

— Puls AI · hours given back vs hours wasted
03
74% Opus 5 · 68% GPT-5.x

Two model families in most stacks

Sonnet 5 sits at 62%, Fable 5 at 39%, Gemini 3 Pro at 23%, and 9% run local models — a few run open models on their own machines. 55% use both an Anthropic and an OpenAI model; 70% use two or more families, 31% three or more. 34% switch models consciously, 27% pick by task, cost and limit, 13% test the new ones and tell the team. 4% have no idea which model is on.

— Puls AI · models used before the edition, model choice
04
44% calibrate per task

Trust is calibrated, not blind

16% trust little and check most of what comes back. 22% trust the repeatable and verify anything new. 44% calibrate trust to the task and know where checking pays. 18% trust on evidence — tests and review — and consciously let go where the risk is low. For comparison, in Stack Overflow's 2025 survey 33% of developers trust AI output and 46% distrust it.

— Puls AI · trust question; Stack Overflow 2025
05
29% hit limits weekly or worse

More than one in four hit limits weekly

60% hit rate limits occasionally, 19% a few times a week, 4% daily, and 5% say the limit is their bottleneck. Four in ten are on a $90–100 tier, four in ten on a $15–20 tier, a handful on the $180 plans. Where the ceiling sits depends on the plan; how much it costs in pace is a question for the next edition.

— Puls AI · rate limits, subscription tier
06
75% worked on AI techniques

Learning happens, mostly on AI itself

51% worked on understanding the client's product and business, 49% on writing clear requirements, 47% on deep reading and analysis, 35% on people and English. Only 12% on teaching others; 5% say the time flew by. Our kick-off note argues the durable skills are the human ones — the pulse shows they get less deliberate time than the tools.

— Puls AI · what did you consciously work on
07
43% set up Pirxey World

The shared system is half adopted

Pirxey World is our shared AI workspace: skills, integrations with Slack, Google Workspace, ClickUp, Figma and our own tools, scheduled automations. 22% use the integrations, 17% the skills, 9% have contributed something. 21% want help with the setup; 36% see no need yet. Personal tools spread on their own. A shared system needs pairing, and we know who asked.

— Puls AI · Pirxey World question
08
81% rate the conditions 4–5 · 49% their own future

Optimism has a ceiling

Asked whether Pirxey gives them the conditions to catch the opportunities, 81% answer 4 or 5 out of 5 (mean 4.03). Asked how they feel about their own professional future with AI, 49% pick 4 or 5 (mean 3.53), 40% sit in the middle and 10% lean towards worries. People trust the company more than the moment. That gap is the number to watch across editions.

— Puls AI · mindset, two 1–5 scales
Turbulence · the challenges

Adoption climbs. The hard part is not the models.

The pulse has a rule: the top blocker of every edition gets one concrete intervention before the next one. In June it was not knowing where to start (31%), so we paired people on setup and ran the first demos. In July it was tool limits and cost (35%), so plan tiers went up. Now it is time. These are the blockers people named, with what they wrote about them — paraphrased, because individual answers stay private.

T1
49% · the top blocker

No time to learn

The most common note: a couple of hours a day, at most, for learning and rest, while the knowledge changes week to week. Several people asked for the same thing in different words — explicit, blessed time to learn, with something taken off the priority list to make room. This edition's intervention: protected time. The blocker moved twice in three months; that is the pulse doing its job.

T2
31% · no time to understand others' output

Code review is the new bottleneck

Engineers wrote the same thing independently: code can be produced far faster than it can be reviewed, and pull requests wait for days. One conclusion recurred: if we want to move at AI speed, the whole delivery cycle has to adapt, review included — with a model reading the code and a person reading the requirements.

T3
22% · low-quality input from others

AI slop, the input problem

Content is generated in seconds; the reader pays. A research summary lands and nobody knows what is true and what was hallucinated, so the team burns time debunking it. The arithmetic several people described: a check skipped upstream costs a multiple of that time downstream, paid by everyone who has to read the output.

T4
82% of engineers · 33% of QA at Crew or above

The ladder is uneven across roles

49 of 60 engineers are at Crew or Mission control; 4 of 12 in QA are. Leadership and delivery sit in between, 8 of 15. QA asked for role-specific training and step-by-step setup guides; several noted that tools built on a Mac take real effort to run on Windows. Same company, different altitudes.

— Puls AI · ladder by department
T5
31% · output quality

Trust with a caveat

A year of daily use has not turned into blind trust, and people say why: the agent that was told not to touch something touches it and reports that it only looked; the model that repeats the same change and insists it is fixed. The habit people described is a boundary — when to stop the loop and change the approach.

T6
45% prefer to watch

Watching, not showing — yet

Asked to show their setup or one task from their workflow, 13% said gladly, 9% asked for time to prepare, 32% said maybe, 45% prefer to watch for now. 36% share knowledge both ways and 9% run demos or build skills for others. And 26% had not heard of our AI transformation initiative. Communication is a blocker too; it just does not appear on the list.

— Puls AI · demo willingness, knowledge sharing
Evidence · sourced

Our numbers, next to the public ones.

How the pulse is run, how it compares, and where it came from.

77
answers in edition three; 58 and 66 in the two before.
Aug 28 – Sep 8, 2026, about 59% of the company: 60 in engineering, 12 in QA, 15 in leadership and delivery, 5 in operations, sales, analysis and marketing (people could tick more than one area). Participation grew from 45% in June to 51% in July to 59% now. Individual answers are seen only by the team that runs the pulse; everything here is an aggregate.
Puls AI · form responses
4 min
per edition; the same seventeen questions, once a month.
Two parts — practice (tools, models, ladder, hours, limits) and growth (what you worked on, where you got stuck, what blocks you, how you feel about the future). Results go back to the team within three days. The top blocker gets one intervention before the next edition. When the response rate drops, we shorten the pulse rather than add reminders; so far it has grown every month.
Puls AI · kick-off note and mechanics
5.6 : 1
an earlier snapshot of the same edition, at 62 answers.
Brief 013 quoted 734 hours given back against 130 wasted while responses were still coming in. At 77 answers the ratio is 5.8 to 1: 912 against 156. The picture held as the late answers arrived, which is the point of asking everyone rather than the early adopters.
Mission Brief 013 — the receipts
90%
of DORA respondents use AI at work; 30% have little or no trust in its code.
DORA's 2025 report: more than 80% feel more productive, throughput goes up, stability goes down, and AI amplifies whatever a team already has. Our pulse reads the same way — the climb is real, and so is the review debt.
DORA 2025 (Google Cloud)
33%
of developers trust AI output; 46% distrust it; 66% cite nearly correct answers as a frustration.
Stack Overflow's 2025 developer survey, 84% using or planning to use AI tools. Against that backdrop our 44% who calibrate trust per task and 18% who trust on evidence look like the habit we want to spread, not an outlier.
Stack Overflow Developer Survey 2025
95%
of measured commits on a client project carried AI-authored code.
The pulse is self-reported; the git log is not. Brief 013 shows the commit prefix that records the AI share, model and tool on every commit, and the weekly report a client receives. Two instruments, one story: adoption is high, supervision is the work.
Mission Brief 013 — the receipts
Mission control standing by

Choosing an engineering partner? Ask them the same questions.

We write custom software. Some pieces are ready-made. We join your team and work alongside it. We will share the questionnaire, the five-level ladder and the mechanics that keep a pulse alive past the third edition — or run the first edition with you. What you get is your own numbers, not ours. Four minutes a month, aggregates only, no scorecard.

Ask how we run the pulseBook a call