01 · route42 ms
Route
Classifies the question · picks the datasets that can answer it.
GetUp / AI
Ask in plain words. Get a cited number in under a second.
Read-only by default. Tenant-isolated. Every claim traced back to the row, file and timestamp it came from.
- Latency
- p50 · 178 ms
- Sources
- 6 datasets connected
- Accuracy
- 98.6% on 12k queries
getup.ai·live178 ms
Your question
> Why did margin drop in Łódź last week?
- routeClassifies the question · picks the datasets that can answer it.42 ms
- readPulls only the rows in scope — your venues, your period. No cross-tenant access.61 ms
- reasonComposes the answer with a structured chain of arithmetic and comparisons.63 ms
- citeAttaches every number to the exact source row, with a stable URL.9 ms
- respondStreams the answer + data card + citations to the operator who asked.3 ms
Cited answerstream · 3 ms
Margin in Łódź dropped 2.4 pts to 17.9% week-over-week.
Driver: night-shift overtime on Fri/Sat (12 Apr, 19 Apr). 38 hrs at premium rate.
Margin · Łódź · trailing 4 weeks−2.4 pts
W14
20.3%
W15
18.7%
W16
18.1%
W17
17.9%
citedpayroll_apr · row 248shifts_apr · 412 entriesp&l/2026-04.csv · row 18
›⏎
01
Anatomy of an answer.
Every question runs through the same five steps. Each step is logged, timed, and cited.
- 01
route
42 ms· 24%
- 02
read
61 ms· 34%
- 03
reason
63 ms· 35%
- 04
cite
9 ms· 5%
- 05
respond
3 ms· 2%
02 · read61 ms
Read
Pulls only the rows in scope — your venues, your period. No cross-tenant access.
03 · reason63 ms
Reason
Composes the answer with a structured chain of arithmetic and comparisons.
04 · cite9 ms
Cite
Attaches every number to the exact source row, with a stable URL.
05 · respond3 ms
Respond
Streams the answer + data card + citations to the operator who asked.
02
Specs.
What GetUp AI actually is, in numbers.
Latency
p50 · 178 ms
p95 · 412 ms · streamed first token
Sources
6 connected
shifts · employees · payroll · invoices · locations · targets
Accuracy
98.6%
measured on 12,418 operator queries · Q1 2026
Languages
EN · PL · TR · DE
ENPLTRDE
answers in the language of the question
Retention
90 / 30 days
queries kept 90d, responses 30d · zero training on your data
Isolation
Tenant-bound
row-level checks · read-only by default
Model
GPT-class
gpt-classswappable provider · same prompt + citation contract
Context window
200k tokens
large enough for a quarter of operational data in one pass
Voice
Planned
Q3
spoken Q&A · 2026 Q3
03
From the logs.
Three real questions our operators asked this week. Identifying details anonymised.
Aleksandra K.·COO · Rooftop Group · 12 venues
yesterday · 14:32Q
Where are we losing money this week?
A
Three losses, €4,820 total. Two are still preventable.
citedshifts_aprpayroll_aprinvoices_makro
Losses · week 17
list
Maria N. double-booked · Wed 14–22−€ 1,840
Piotr Z. active during sick leave−€ 2,100
Olive oil +12% at Makro−€ 880
Marek D.·Operations Lead · Stay Warsaw
Mon · 09:14Q
What will payroll be this month?
A
Projected €58,420 — €2,300 over budget. Friday is the swing.
citedpayroll_aprshifts_apr
April · forecast
metric
Budget€ 56,120
Projected close€ 58,420
Variance
+ 4.1%
Confidence
± 2.1%
Magda W.·Founder · Northwind · 8 stores
Sun · 21:08Q
Compare Wrocław to the network on labour share.
A
Wrocław runs 4.2 pts hotter than the network. The night cashier double-shifts twice a week.
citedshifts_wropayroll_wro
Labour share · trailing 4 weeks
table
Wrocław37.0%
Network avg32.8%
Gap+4.2pt
Driver Night double-shift
04
What it won't do.
We'd rather be quiet about a question than guess at it.
— 01 —
blockedIt won't write without permission.
Every action — close a shift, approve a payroll, switch a supplier — needs an explicit human confirm.
— 02 —
blockedIt won't infer past the data.
If a row isn't there, the answer says so. No filling in blanks with priors.
— 03 —
blockedIt won't forecast without a band.
Every projection comes with a confidence interval and the inputs it was built on.
— 04 —
blockedIt won't share across tenants.
Your data stays inside your workspace. Row-level checks, every query, no exceptions.
— 05 —
blockedIt won't train on your data.
Zero training. Queries kept 90 days for audit, responses 30 days, then erased.
— 06 —
blockedIt won't pretend to be deterministic.
Same question, same data, same answer — but we'll show you when a value is a range, not a point.
05
I stopped pulling reports two months ago. I just ask. And when I want to double-check, the answer points me at the row.
06
Ask your first question.
Free for 14 days. No card. Drops into your existing GetUp workspace — or starts a new one in two minutes.