Field report 01 · ANALOG · 2026 edition
In the current ANALOG sample, AI setup is showing up as a budget, a stack, and a discipline. Treat the early read as directional; the patterns sharpen as more verified rows are accepted.
Chapter 01 — The money
Median per-person monthly AI budget, with the middle half of responses shown as a band. Normalized to USD at submission; toggle to BRL above the chart.
Figure 01 · The economics
Each row is one role across the verified sample. The dot is the median; the band is the middle 50% of responses (p25 → p75).
$263
$195–$324
$180
$180–$180
$150
$95–$162
$95
$84–$132
$74
$74–$74
Figure 02 · The economics
Stepped IQR bands across size brackets. Read the shape as directional; buckets stay hidden until they clear the privacy floor.
Figure 03 · The economics
The dominant shape of the spend — subscriptions, usage, or self-hosted compute.
Median spend
$152
Per person per month across all roles.
N = 151
Highest cohort
$263
Developers median
N = 56
Spend trajectory
60%
of respondents report spend going up.
N = 121
What this means
Chapter 02 — The stack
The current sample points to recognizable defaults, but the long tail is still doing important work. Editor, model, and agent are read as independent layers, not as a single respondent-level bundle.
Read the stack by layer first. The combinations matter, but this page does not infer co-occurrence from separate shares.
— ANALOG editors · field read
Figure 04 · The stack · editor
100 squares · one square per percent of respondents. The top category claims the largest block; the long tail pools into 'Other'.
Figure 05 · The stack · LLM
Primary LLM, ranked. The stem is share; the dot is where the percentage lands on a relative scale.
Figure 06 · The stack · agent
Of every 10 respondents, who's driving the agent layer.
Each figure ≈ 10% of respondents
Figure 07 · The stack · MCP servers
Rectangles sized by share. Multi-select: respondents can list more than one. Top tiles dominate; the small ones are where teams are improvising.
Figure 08 · The stack · open weights
Among respondents who run any OSS model, the ones they pick.
What this means
Chapter 03 — The output
Respondent estimate of the share of their daily work AI produces. Held side-by-side with budget on the next figure so the read stays careful: the two are loosely coupled.
Budget buys access. It does not, on its own, change how the day reads.
— ANALOG editors · field read
Figure 09 · The output · by role
The dot is the median; the band is the middle 50% of responses. The IQR is wider where roles disagree with themselves.
66%
60%–73%
56%
56%–56%
50%
46%–53%
37%
37%–44%
20%
20%–20%
Figure 10 · The output · budget vs. work
Each role plotted by median monthly AI spend (x) and median share of work AI produces (y). Bubble area follows paired-response sample size.
What this means
Chapter 04 — The new defaults
Rules-as-code, open-source fallbacks, and managed infrastructure (Bedrock) are three behaviors to watch as the verified sample grows. Each points to a different kind of operating constraint.
Rules files, open weights, and Bedrock mark three different reasons teams put structure around model use.
— ANALOG editors · field read
Figure 11 · The defaults · adoption
Each crowd shows the share who say yes. Filled squares are adopters; empty squares are everyone else.
Bedrock
28%
Run models through Amazon Bedrock.
N = 151
Open source
38%
Run an open-source model in their stack.
N = 151
Bedrock
1 in 4
Run models through Amazon Bedrock.
N = 151
Open source
1 in 3
Run an open-source model in their stack.
N = 151
What this means
Chapter 05 — The splits
The same survey, two audiences. We only draw a line when the named category clears the privacy floor in both groups.
The early read suggests model choice means different jobs in different cohorts. More accepted rows will make the split clearer.
— ANALOG editors · field read
Figure 13 · The splits · primary LLM
Categories ranked in each group; lines connect a category's share in each column. Steep lines are the movers.
Developers
Everyone else
What this means
Chapter 06 — Method
Every figure is built from verified, anonymous survey responses. Buckets with fewer than five responses are hidden or pooled into “Other” to protect respondents — so smaller cohorts will appear as gaps, not as data.
Budgets are submitted in the respondent’s local currency and normalized to USD at submission. The BRL toggle applies an approximate rate; the underlying figures remain the USD aggregates.
Stack share is reported by layer (editor, agent, LLM, MCP, open-weights). Layer reads are independent — they are not treated as respondent-level co-occurrence.
Figure 14 · Method · region
Self-selected and US-skewed — read everything else accordingly.