# Epicor Kinetic AI Assistant Sizing and Value Calculator

> Source: https://netray.co/tools/epicor-kinetic-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, vector index size, hours saved and payback for a private AI assistant on Epicor Kinetic on-premises.

## Category

epicor

## Erp

Epicor Kinetic

## Industry

Discrete Manufacturing

## Audience

- CIO
- IT manager
- Finance controller
- ERP administrator

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Named users

**Key:** users

**Type:** number

**Default:** 220

**Min:** 1

**Step:** 10

**Unit:** users

**Help:** Users licensed to use the assistant.

#### Daily active share

**Key:** active_pct

**Type:** slider

**Default:** 35

**Min:** 5

**Max:** 100

**Step:** 5

**Unit:** %

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

#### Questions per active user per day

**Key:** q_per_user

**Type:** number

**Default:** 8

**Min:** 1

**Max:** 200

**Step:** 1

**Unit:** questions

**Help:** Pilot logs give the real figure; 5 to 15 is common once habit forms.

#### Prompt plus answer tokens per question

**Key:** tokens_per_q

**Type:** number

**Default:** 4000

**Min:** 500

**Max:** 32000

**Step:** 250

**Unit:** tokens

**Help:** Question, retrieved BAQ rows and document chunks, plus the answer. 3,000 to 6,000 is typical for grounded answers.

#### Documents to index

**Key:** docs

**Type:** number

**Default:** 150000

**Min:** 0

**Step:** 1000

**Unit:** documents

**Help:** Travelers, drawings, inspection reports, certs and SOPs you want searchable.

#### Average pages per document

**Key:** pages_per_doc

**Type:** number

**Default:** 3

**Min:** 1

**Max:** 500

**Step:** 1

**Unit:** pages

**Help:** Average across the document set.

#### Minutes saved per question

**Key:** minutes_saved

**Type:** number

**Default:** 6

**Min:** 0

**Max:** 120

**Step:** 0.5

**Unit:** minutes

**Help:** Time a person would otherwise spend finding the answer across screens or asking a colleague.

#### Loaded hourly cost

**Key:** hourly_cost

**Type:** number

**Default:** 55

**Min:** 0

**Step:** 1

**Unit:** per hour

**Help:** Salary plus benefits and overhead for the people asking.

#### GPU monthly cost

**Key:** gpu_monthly

**Type:** number

**Default:** 1500

**Min:** 0

**Step:** 100

**Unit:** per GPU per month

**Help:** Hardware amortised over 36 months plus power, cooling, support and rack space.

#### Throughput per GPU

**Key:** tps_per_gpu

**Type:** number

**Default:** 1200

**Min:** 100

**Max:** 10000

**Step:** 100

**Unit:** tokens/s

**Help:** Aggregate tokens per second under batching for your model and quantisation. Measure it; this is the most sensitive assumption.

#### One-time implementation cost

**Key:** impl_cost

**Type:** number

**Default:** 60000

**Min:** 0

**Step:** 5000

**Unit:** one-time

**Help:** Integration, document pipeline, security review and pilot.

**Outputs:**

#### Questions per day

**Key:** q_day

**Expr:** users * active_pct / 100 * q_per_user

**Format:** integer

**Description:** Named users x active share x questions per active user.

#### Tokens per day

**Key:** tokens_day

**Expr:** q_day * tokens_per_q

**Format:** number

**Description:** Questions per day x tokens per question.

#### Peak concurrent requests

**Key:** peak_conc

**Expr:** max(1, ceil(q_day * 0.25 / 3600 * 15))

**Format:** integer

**Description:** Assumes 25 percent of daily questions land in the busiest hour and each request holds a slot for about 15 seconds.

#### GPUs needed (before redundancy)

**Key:** gpus

**Expr:** max(1, ceil(q_day * 0.25 / 3600 * tokens_per_q / tps_per_gpu))

**Format:** integer

**Primary:** true

**Description:** Peak-hour tokens per second divided by throughput per GPU, rounded up. Add a spare GPU if you need failover.

#### Vector index size

**Key:** index_gb

**Expr:** docs * pages_per_doc * 2 * 1024 * 4 * 1.5 / 1000000000

**Format:** decimal

**Description:** Gigabytes, assuming 2 chunks per page, 1024-dimension float32 embeddings and 50 percent index overhead. Original files are extra.

#### Hours saved per month

**Key:** hours_saved

**Expr:** q_day * minutes_saved / 60 * 21

**Format:** hours

**Description:** Questions per day x minutes saved, over 21 working days.

#### Monthly value

**Key:** value_month

**Expr:** hours_saved * hourly_cost

**Format:** currency

**Primary:** true

**Description:** Hours saved x loaded hourly cost. Treat as an upper bound until a pilot confirms the time saved.

#### Simple payback

**Key:** payback

**Expr:** if(gt(value_month - gpus * gpu_monthly, 0), impl_cost / (value_month - gpus * gpu_monthly), 0)

**Format:** months

**Description:** One-time cost divided by monthly value minus GPU cost. Shows 0 when value does not exceed GPU cost.

**Chart:**

#### Monthly value vs named users

**Type:** line

**X:** users

**From:** 20

**To:** 1000

**Steps:** 25

**Series:**

- value_month

**Disclaimer:** Planning estimate only. Tokens per second, time saved per question and peak share vary widely; replace the defaults with measurements from your own pilot before committing to hardware.

## Content

**Intro:** This calculator sizes a private AI assistant on Epicor Kinetic: how many GPUs peak load needs, how large the vector index gets, and what the time saved is worth against GPU cost. Every assumption is an input you can replace.

**Sections:**

#### How to use it

**Body:** Start with named users and the share who ask a question on a typical day. Set questions per active user from a pilot if you have one. Tokens per question include retrieved BAQ rows and document chunks, so grounded answers cost more than chat. Enter your measured tokens per second per GPU if you have it; otherwise keep the default and treat the GPU count as an estimate.

**Bullets:**

- Use a pilot log for questions per user
- Count retrieved context in the token figure
- Use amortised monthly GPU cost, not purchase price
- Run the sweep to see where a second GPU becomes necessary

#### The formulas

**Body:** Questions per day = users x active share x questions per user. Peak-hour tokens per second = daily questions x 0.25 / 3600 x tokens per question. GPUs = peak tokens per second / throughput per GPU, rounded up. Index size = documents x pages x 2 chunks x 1024 dimensions x 4 bytes x 1.5 overhead. Hours saved = daily questions x minutes saved / 60 x 21 days. Payback = one-time cost / (monthly value minus GPU cost).

#### Reading the result

**Body:** GPU count is driven by the busiest hour, not by headcount, so two sites with the same users can need different hardware. Monthly value is an upper bound until a pilot confirms the time saved. If value barely exceeds GPU cost, payback shows zero and the case rests on better decisions rather than time savings. Add a spare GPU if the assistant must stay up during a failure.

#### Common mistakes

**Body:** The common errors are using the vendor's peak throughput figure instead of a measured one, forgetting retrieved context in the token count, and counting every named user as active. Also remember that Epicor's own BAQ dashboards may answer fixed questions at no extra cost; the assistant is worth the most on cross-module questions.

**Cta:** If you want these numbers validated against your Kinetic environment and a real pilot, Netray can run that with you. The calculator works without any contact.

**Faq:**

#### Why does GPU count depend on peak and not users?

**Answer:** Because requests arrive unevenly. A GPU serves a limited number of tokens per second, so the busiest hour sets the requirement. Two hundred named users who ask a few questions each may need only one GPU.

#### What throughput per GPU should I enter?

**Answer:** Measure it with your model, quantisation and context length under batching. A 7B to 14B model on a 24 GB to 48 GB GPU often delivers roughly 1,000 to 2,500 aggregate tokens per second, but results vary widely, so the default is only a starting point.

#### Does the index size include the original files?

**Answer:** No. It estimates the vector index only. Original documents, extracted text and the SQL replica need separate storage, so plan additional capacity for them.

#### How reliable is the monthly value?

**Answer:** It is as reliable as your minutes-saved input. Pilot with a small group, log answers that saved real time, and replace the default before using the figure in a business case.

## Search terms

Epicor Kinetic AI sizing, Epicor AI ROI calculator, private LLM GPU sizing ERP, Epicor Kinetic AI cost, on-prem AI assistant payback, ERP chatbot sizing calculator

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