# Custom MES on SQL Server AI Assistant Sizing and Value Calculator

> Source: https://netray.co/tools/custom-mes-sql-server-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, hours saved and payback for a private AI assistant over Custom MES on SQL Server.

## Category

on-prem-ai

## Erp

Custom MES (SQL Server)

## Industry

Manufacturing

## Audience

- CIO
- IT manager
- Plant manager
- Solution architect

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Named users

**Key:** users

**Type:** number

**Default:** 250

**Min:** 1

**Step:** 10

**Unit:** users

#### Daily active share

**Key:** dau_share

**Type:** slider

**Default:** 35

**Min:** 5

**Max:** 100

**Step:** 5

**Unit:** %

**Help:** Share of named users who ask at least one question on a working day.

#### Questions per active user per day

**Key:** q_per_user

**Type:** number

**Default:** 5

**Min:** 1

**Step:** 1

**Unit:** questions

#### Average prompt plus answer tokens

**Key:** tokens_per_q

**Type:** number

**Default:** 1400

**Min:** 100

**Step:** 100

**Unit:** tokens

**Help:** Include retrieved context in the prompt. Retrieval-heavy questions are often 1,500 to 6,000 tokens.

#### Peak hour multiplier

**Key:** peak_mult

**Type:** number

**Default:** 3

**Min:** 1

**Max:** 10

**Step:** 0.5

**Unit:** x

**Help:** Busiest hour as a multiple of the average working hour. Shift starts and handovers raise it.

#### Throughput per GPU

**Key:** tps_per_gpu

**Type:** number

**Default:** 1500

**Min:** 100

**Step:** 100

**Unit:** tokens/s

**Help:** Aggregate generation plus prompt throughput under batching. Measure your model on your GPU; this default is a planning assumption only.

#### Documents to index

**Key:** docs

**Type:** number

**Default:** 30000

**Min:** 0

**Step:** 1000

**Unit:** documents

**Help:** Records, work orders, procedures or files, counted as documents.

#### Average pages per document

**Key:** pages_per_doc

**Type:** number

**Default:** 4

**Min:** 1

**Step:** 1

**Unit:** pages

#### Minutes saved per question

**Key:** min_saved

**Type:** number

**Default:** 6

**Min:** 0

**Step:** 1

**Unit:** minutes

#### Loaded hourly cost

**Key:** hourly_cost

**Type:** number

**Default:** 55

**Min:** 0

**Step:** 5

**Unit:** $/hour

#### GPU monthly cost

**Key:** gpu_cost

**Type:** number

**Default:** 900

**Min:** 0

**Step:** 100

**Unit:** $/month

**Help:** Amortised hardware, power and support per GPU, or private cloud rental.

#### One-time implementation cost

**Key:** impl_cost

**Type:** number

**Default:** 60000

**Min:** 0

**Step:** 5000

**Unit:** $

**Outputs:**

#### Daily active users

**Key:** active_users

**Expr:** users * dau_share / 100

**Format:** integer

#### Questions per day

**Key:** q_day

**Expr:** active_users * q_per_user

**Format:** integer

#### Tokens per day

**Key:** tokens_day

**Expr:** q_day * tokens_per_q

**Format:** integer

#### Peak questions per hour

**Key:** peak_q_hour

**Expr:** q_day / 8 * peak_mult

**Format:** integer

**Description:** Assumes an 8-hour working day.

#### Peak concurrent requests

**Key:** peak_conc

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

**Format:** integer

**Description:** Little's law with a 20-second average request time.

#### Peak tokens per second

**Key:** peak_tps

**Expr:** peak_q_hour * tokens_per_q / 3600

**Format:** integer

#### GPUs needed (no redundancy)

**Key:** gpus

**Expr:** max(1, ceil(peak_tps / tps_per_gpu))

**Format:** integer

**Primary:** true

**Description:** Add one GPU if you need redundancy.

#### Index chunks

**Key:** chunks

**Expr:** ceil(docs * pages_per_doc * 600 * 1.15 / 400)

**Format:** integer

**Description:** 600 tokens per page, 400-token chunks, 15% overlap.

#### Index size

**Key:** index_gb

**Expr:** chunks * 7000 / 1000000000

**Format:** decimal

**Description:** About 7 KB per chunk for vector, text and metadata.

#### Hours saved per month

**Key:** hours_saved

**Expr:** q_day * 21 * min_saved / 60

**Format:** hours

**Description:** Assumes 21 working days.

#### Monthly value

**Key:** monthly_value

**Expr:** hours_saved * hourly_cost

**Format:** currency

#### Monthly GPU cost

**Key:** gpu_spend

**Expr:** gpus * gpu_cost

**Format:** currency

#### Net monthly value

**Key:** net_month

**Expr:** monthly_value - gpu_spend

**Format:** currency

#### Simple payback

**Key:** payback

**Expr:** if(gt(net_month, 0), impl_cost / net_month, 0)

**Format:** months

**Primary:** true

**Description:** Zero means no payback at these inputs.

**Chart:**

#### Monthly value and GPU cost vs named users

**Type:** line

**X:** users

**From:** 20

**To:** 1500

**Steps:** 30

**Series:**

- monthly_value
- gpu_spend

**Disclaimer:** Planning estimate only. Throughput per GPU depends on model, quantisation, context length and serving stack, so benchmark your own setup. Time saved is a user-reported assumption, not a measured result.

## Content

**Intro:** This calculator sizes a private AI assistant over Custom MES on SQL Server and estimates the value it returns. Change the inputs to see how many GPUs you need, how large the retrieval index is and when the project pays back.

**Sections:**

#### How to use it

**Body:** Start with named users and the share who use the assistant on a typical day. Set questions per active user from a pilot or a survey; three to six a day is a common planning range for a plant tool. Prompt plus answer tokens should include retrieved context, which dominates cost. Then set your measured throughput per GPU. The chart shows how value and GPU cost scale with the number of named users.

**Bullets:**

- Use pilot data for questions per day
- Include retrieved context in token counts
- Benchmark tokens per second on your own GPU
- Treat minutes saved as an assumption to verify

#### Formulas and assumptions

**Body:** Questions per day = users x daily active share x questions per user. Peak questions per hour = questions per day divided by 8, times the peak multiplier. Peak concurrency uses Little's law with a 20-second request time. GPUs = ceiling of peak tokens per second divided by tokens per second per GPU, minimum one. Index chunks = documents x pages x 600 tokens x 1.15 overlap divided by 400. Hours saved = questions per day x 21 days x minutes saved divided by 60. Payback = one-time cost divided by net monthly value.

#### System notes for Custom MES on SQL Server

**Body:** The ai schema is the contract: only reviewed views are visible, columns that hold sensitive data are left out or masked, and every query runs with the end user's identity in the session context so row-level security applies. Write actions go only through the MES service layer with human approval, never through direct updates.

#### Common mistakes

**Body:** Sizing errors are usually in tokens and in value, not in arithmetic.

**Bullets:**

- Counting only the question and answer, and forgetting retrieved context
- Using vendor peak throughput instead of measured throughput under real batching
- Assuming every named user is active every day
- Counting all minutes saved as cash savings

**Cta:** If you would like these numbers checked against a pilot on your own Custom MES on SQL Server data, Netray can run the sizing and a read-only proof of value on your hardware.

**Faq:**

#### What default throughput per GPU should I use?

**Answer:** The default of 1,500 tokens per second is a planning assumption for a mid-sized open-weight model with batching. Real figures vary a lot with model size, quantisation and context length, so measure your own setup with vLLM or Ollama before buying hardware.

#### Why is peak concurrency calculated from a 20-second request?

**Answer:** Little's law says concurrent requests equal arrival rate times time in system. Retrieval plus generation often takes 10 to 30 seconds for a grounded answer, so 20 seconds is a reasonable default. Change the expression if your measured time differs.

#### How reliable is the hours-saved figure?

**Answer:** It is only as good as the minutes-saved input. Measure it in a pilot by timing how long the same question takes without the assistant. Treat early numbers as a range, and do not count the saved minutes as cash unless the work is actually reduced.

#### Does the index size include the model?

**Answer:** No. Index size covers vectors, chunk text and metadata for the documents you list. Model weights are separate and range from a few GB to tens of GB depending on model size and quantisation.

## Search terms

custom mes on sql server ai sizing, custom mes on sql server ai roi, on-prem ai assistant sizing calculator, llm gpu count estimator, private rag index size calculator, ai assistant payback calculator, tokens per day calculator

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