# SYSPRO AI Assistant Sizing and Value Calculator

> Source: https://netray.co/tools/syspro-ai-assistant-sizing-and-value-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Estimate GPUs, index size and monthly value for a private AI assistant on SYSPRO. Enter your named users, usage pattern and cost assumptions to see peak concurrency, daily token volume and payback.

## Category

other-erp

## Erp

SYSPRO

## Audience

- CIO
- IT manager
- ERP architect
- Finance lead

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Named SYSPRO users

**Key:** users

**Type:** number

**Default:** 250

**Min:** 1

**Step:** 10

**Unit:** users

#### Daily active share of users

**Key:** active

**Type:** slider

**Default:** 30

**Min:** 5

**Max:** 90

**Step:** 1

**Unit:** %

**Help:** Share of named users who ask the assistant something on a given day.

#### Questions per active user per day

**Key:** qpd

**Type:** number

**Default:** 10

**Min:** 1

**Max:** 100

**Step:** 1

**Unit:** questions

#### Average prompt plus answer tokens

**Key:** tokens_q

**Type:** number

**Default:** 3500

**Min:** 200

**Step:** 100

**Unit:** tokens

**Help:** Includes retrieved context. Short SQL answers run lower; document questions run higher.

#### Documents to index

**Key:** docs

**Type:** number

**Default:** 40000

**Min:** 0

**Step:** 1000

**Unit:** documents

#### Average pages per document

**Key:** pages

**Type:** number

**Default:** 6

**Min:** 1

**Step:** 1

**Unit:** pages

#### Minutes saved per question

**Key:** min_saved

**Type:** number

**Default:** 4

**Min:** 0

**Step:** 0.5

**Unit:** minutes

**Help:** Time the user would otherwise spend finding the answer. Measure it in a pilot.

#### Loaded hourly cost per user

**Key:** rate

**Type:** number

**Default:** 45

**Min:** 0

**Step:** 1

**Unit:** $/hr

#### GPU monthly cost

**Key:** gpu_cost

**Type:** number

**Default:** 1800

**Min:** 0

**Step:** 100

**Unit:** $/GPU/month

**Help:** Amortised hardware plus power and hosting per GPU, or the monthly rate for private cloud GPUs.

#### Tokens per second per GPU (sustained)

**Key:** tps

**Type:** number

**Default:** 1500

**Min:** 50

**Step:** 50

**Unit:** tokens/s

**Help:** Aggregate throughput with batching for your chosen model and quantization. Measure with vLLM on your hardware.

**Outputs:**

#### Questions per day

**Key:** q_day

**Expr:** users * active / 100 * qpd

**Format:** integer

#### Tokens per day

**Key:** tok_day

**Expr:** q_day * tokens_q

**Format:** number

#### Peak concurrent requests

**Key:** peak_conc

**Expr:** max(1, ceil(q_day * 0.2 / 3600 * 20))

**Format:** integer

**Description:** Assumes 20% of daily questions land in the busiest hour and each request holds a slot for about 20 seconds.

#### GPUs needed (estimate)

**Key:** gpus

**Expr:** max(1, ceil(tok_day * 0.2 / 3600 / tps))

**Format:** integer

**Primary:** true

**Description:** Peak-hour tokens per second divided by tokens per second per GPU, minimum one. Add one for resilience in production.

#### Index size estimate

**Key:** index_gb

**Expr:** docs * pages * 3 * 9 / 1000000

**Format:** decimal

**Description:** About 3 chunks per page at roughly 9 KB each for text and vector. Result in GB.

#### Hours saved per month

**Key:** hours_month

**Expr:** q_day * 22 * min_saved / 60

**Format:** hours

#### Monthly value of time saved

**Key:** value_month

**Expr:** hours_month * rate

**Format:** currency

**Primary:** true

#### Days of value to cover one month of GPU cost

**Key:** payback_days

**Expr:** gpus * gpu_cost / (value_month / 30)

**Format:** days

**Description:** GPU monthly cost divided by value per day. Under 30 means value exceeds GPU cost.

**Chart:**

#### GPUs needed vs questions per active user

**Type:** bar

**X:** qpd

**From:** 2

**To:** 60

**Steps:** 20

**Series:**

- gpus
- peak_conc

**Disclaimer:** Planning model. Formulas: questions/day = users x active share x questions each; tokens/day = questions x tokens; GPUs = ceil(peak-hour tokens per second / tokens per second per GPU); value = questions/day x 22 working days x minutes saved / 60 x hourly cost. The 20% peak-hour share and 20 second request time are fixed assumptions. Validate throughput with your own model and prompts.

## Content

**Intro:** Estimate GPUs, index size and monthly value for a private AI assistant on SYSPRO. Enter your named users, usage pattern and cost assumptions to see peak concurrency, daily token volume and payback. It uses plain formulas you can check and change.

**Sections:**

#### How to use it

**Body:** Enter the number of named SYSPRO users and the share who use the assistant daily. Add questions per active user, average tokens per question, documents and pages for the index, minutes saved, loaded hourly cost and GPU monthly cost. The tokens per second per GPU input is the most important assumption, so replace the default with a number from a test on your hardware.

**Bullets:**

- Start with a pilot group size and read hours saved.
- Raise tokens per question if you use long documents.
- Use the chart to see how demand scales with users or usage.

#### The method behind it

**Body:** Daily tokens are questions times tokens per question. Peak load assumes 20% of daily questions land in the busiest hour, converted to tokens per second and divided by per-GPU throughput. Concurrency uses the same arrival rate times a 20 second request time. Index size assumes about three chunks per page, each around 9 KB for text and vectors. Value is minutes saved converted to hours, times 22 working days and hourly cost.

#### How to read the result

**Body:** Typical SYSPRO questions are stock by warehouse, order and job status, shortages and purchase receipts, so many answers are short SQL results plus a few lines of explanation. If the GPU count is one, plan a second for resilience once the assistant is relied on. If payback days exceed 30, either the usage is low, the time saved is small, or the hardware is oversized for the load.

#### Common mistakes

**Body:** Counting named users as active users, ignoring retrieved context in the token count, trusting vendor throughput numbers for a different model, and claiming time saved without a pilot measurement. Another is forgetting that model quality, not just speed, decides whether people keep using the tool. Treat value as a hypothesis until a pilot confirms it.

**Cta:** Want these numbers tested against your real SYSPRO data and prompts? Netray can run a measured pilot on your own GPUs and report throughput and accuracy.

**Faq:**

#### Why is tokens per second per GPU an input?

**Answer:** It depends on the model, quantization, context length and batching. Public numbers rarely match your workload, so measure with vLLM on your hardware and enter the sustained figure.

#### Does the GPU estimate include redundancy?

**Answer:** No. It is the minimum for peak-hour load. For a production assistant, add at least one spare GPU or a failover plan.

#### How reliable is the value figure?

**Answer:** It is only as good as minutes saved per question. Measure that in a pilot by timing a sample of tasks with and without the assistant.

#### What does the index size include?

**Answer:** An estimate of text chunks and vectors for the documents you enter. It excludes the ERP database copy, which is sized separately by your DBA.

## Search terms

SYSPRO AI sizing calculator, SYSPRO LLM GPU sizing, on-prem AI assistant ROI calculator, SYSPRO private AI cost, GPU count for ERP assistant, tokens per day calculator ERP, AI assistant payback calculator

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json