Can the AI see data an employee shouldn't?

No. Every tool call runs with the asking employee's own roles and permissions in your ERP, so the assistant can only see what that person could see by logging in themselves. Anything it would change waits for a person's approval, and every question, tool call and answer is logged with who asked.

What stops someone tricking it?

The guard-rail stage filters what goes in and what comes out, looks for instructions hidden inside documents, messages and tool results, and masks personal data such as phone, PAN, Aadhaar and bank numbers before the model or the logs see it. Attack tests run before every change, so a new prompt or model cannot quietly reopen a gap.

Which systems can it connect to?

Your ERP, whether ERPNext, Tally, SAP Business One, Zoho or Odoo, plus your helpdesk, team chat, shared drives and email. Each connection is a tool over the Model Context Protocol, added one at a time, read-only first. Write actions such as updating a ticket or an order come later, always behind a person's approval.

How do we keep the cost under control?

Three stages work on it together. Repeated questions are answered from a cache, a router sends easy questions to small models and hard ones to strong models, and you set budgets per month and per team. Observability then shows cost per team and per job every month, so the model bill is never a surprise.

How long does a pilot take, and what does it cost?

It depends on the job, the systems it reads and where the model runs, so we scope it on a free 30-minute call rather than quote a figure here. Bring one job you want the AI to do. We map it to the stages it needs, tell you which can wait, and send a written plan with the time and the price.

Services · Private AI for your company

Your company's own AI, kept inside your rules.Private AI for companies: RAG, MCP tools, guard rails and observability

An assistant that reads your documents, checks your orders and tickets, and only ever sees what the person asking is allowed to see. We build it in 16 stages; you start with the ones your first job needs.

See the ways to start

Runs on your server or ours. Works with ERPNext, Tally, SAP Business One, Zoho and Odoo. Free 30-minute call.

“Where is order SO-1042, and can we move delivery to Friday?”

Answer: SO-1042 is packed, not yet dispatched, so it can move to Friday once you approve the change.

Plan it · stages 1–4Complete
Job one on the brief; the prompt says check with tools, never guess a date.
Connect it to your data and tools · stages 5–8Complete
Finds the delivery policy and looks up SO-1042 with the asker's own permissions.
Make it safe · stages 9–11Complete
Masks the customer's phone; the date change waits for a person's yes.
Make it fast and affordable · stages 12–14Complete
Policy from the cache, a mid-size model, one trace with its cost.
Put it to work · stages 15–16Complete
Answers in team chat; the thumbs-up goes into next month's review.

A recorded run on a sample order. Your stages are set on the first call.

Plan it (Stages 1–4)

Job one on the brief; the prompt says check with tools, never guess a date.

Connect it to your data and tools (Stages 5–8)

Finds the delivery policy and looks up SO-1042 with the asker's own permissions.

Make it safe (Stages 9–11)

Masks the customer's phone; the date change waits for a person's yes.

Make it fast and affordable (Stages 12–14)

Policy from the cache, a mid-size model, one trace with its cost.

Put it to work (Stages 15–16)

Answers in team chat; the thumbs-up goes into next month's review.

What your teams will ask it

Each team gets answers from its own records, and anything that changes data still waits for a person's yes.

Which team starts?

“Can we promise 400 units of the blue valve to our Pune dealer by the 12th?”

It checks: Stock and open orders in the ERP, the customer's price list and credit limit.
On its own: Answers, drafts the quotation.
Needs a person's yes: Sending the quotation, any discount past the rep's limit.
Stages it leans on: Search over your own documents (RAG) · Tools it can use (MCP) · Role-based access in every tool call · Human approval and audit trail

“Can we promise 400 units of the blue valve to our Pune dealer by the 12th?”

It checks: Stock and open orders in the ERP, the customer's price list and credit limit.
On its own: Answers, drafts the quotation.
Needs a person's yes: Sending the quotation, any discount past the rep's limit.
Stages it leans on: Search over your own documents (RAG) · Tools it can use (MCP) · Role-based access in every tool call · Human approval and audit trail

“This customer says the pump stopped after a month. Is it under warranty, and what did we tell them last time?”

It checks: The serial number, the warranty terms, the ticket history in the helpdesk.
On its own: Answers, drafts the reply, tags the ticket.
Needs a person's yes: Sending the reply, booking a replacement.
Stages it leans on: Search over your own documents (RAG) · Tools it can use (MCP) · Memory and context · Prompt-injection guard rails

“Which customers are past their credit terms, and what does each one owe?”

It checks: Receivables and ledgers in the books, read with the asker's own rights.
On its own: Answers with the figures and links to each invoice.
Needs a person's yes: Sending reminders, any change to an entry.
Stages it leans on: Tools it can use (MCP) · Role-based access in every tool call · Prompt-injection guard rails · Human approval and audit trail

“How many days of leave do I have left, and can I carry them into next year?”

It checks: The leave policy and the asker's own leave balance, never anyone else's.
On its own: Answers, with the policy line it used.
Needs a person's yes: Applying for leave on the person's behalf.
Stages it leans on: Search over your own documents (RAG) · Role-based access in every tool call · Prompt-injection guard rails

“What is short for next week's production plan, and who supplied it last?”

It checks: Bills of materials, stock in each warehouse, past purchase orders.
On its own: Answers, drafts the purchase request.
Needs a person's yes: Raising the purchase order.
Stages it leans on: Search over your own documents (RAG) · Tools it can use (MCP) · Cache for repeated questions · Router between models

“Why did margins drop in the south region last month?”

It checks: Sales, discounts and costs by region, within what the asker may see.
On its own: Answers with the figures and where each one came from.
Needs a person's yes: Nothing: it only reads.
Stages it leans on: Search over your own documents (RAG) · Tools it can use (MCP) · Role-based access in every tool call · Observability

What makes a company AI safe to switch on

Sixteen stages in five phases, each with what it does, what we use, and what goes wrong if it is skipped.

1 · Plan itStages 1–4

Use cases and success measures

Stage 1 of 16 · Plan it

Pick three to five jobs the AI does first, such as order status, ticket triage or policy answers, and write down what a good answer looks like for each, with today's baseline.

What we use: A one-page brief per job · Sample questions from real staff
What you get: A short list everyone agrees on, and a way to tell whether it worked.
If you skip it: A clever demo nobody uses, because it was never pointed at a real job.
The sample question: Order status is job one on the brief; changing a delivery date is job two.

2 · Connect it to your data and toolsStages 5–8

3 · Make it safeStages 9–11

4 · Make it fast and affordableStages 12–14

5 · Put it to workStages 15–16

Use cases and success measures

Stage 1 of 16 · Plan it

Pick three to five jobs the AI does first, such as order status, ticket triage or policy answers, and write down what a good answer looks like for each, with today's baseline.

What we use: A one-page brief per job · Sample questions from real staff
What you get: A short list everyone agrees on, and a way to tell whether it worked.
If you skip it: A clever demo nobody uses, because it was never pointed at a real job.
The sample question: Order status is job one on the brief; changing a delivery date is job two.

Data map and sources

Stage 2 of 16 · Plan it

List every system and folder the AI may read: ERP, helpdesk, shared drives, email, chat. Mark who owns each, what is sensitive and what stays off-limits.

What we use: ERPNext, Tally, SAP Business One, Zoho or Odoo · Drives, wikis and the helpdesk
What you get: A data map you can show your auditor.
If you skip it: The AI quietly reads payroll because a folder was shared.
The sample question: Orders live in the ERP; delivery rules live in the logistics policy.

Model choice and hosting

Stage 3 of 16 · Plan it

Choose hosted models, a private cloud deployment or open-weight models on your own server, weighing your data rules, cost and quality.

What we use: Hosted: Claude, GPT or Gemini · Your own server: open-weight models · Data kept in India or the region you pick
What you get: A written choice with its reason, which you can change later.
If you skip it: Customer data leaves the country without anyone deciding it should.
The sample question: It runs on a model hosted in the region the brief named.

System prompt

Stage 4 of 16 · Plan it

The standing instructions: who the assistant is, its tone, what it answers, what it refuses and how it cites sources. Versioned like code.

What we use: Versioned prompt files · One prompt per assistant role
What you get: Answers that sound like your company and stay inside scope.
If you skip it: It answers anything, in any tone, and nobody knows which version said what.
The sample question: The prompt says: check order data with tools, never guess a date.

Search over your own documents (RAG)

Stage 5 of 16 · Connect it to your data and tools

Load your documents, split and index them, and fetch the right passages for each question, so answers quote your own sources. It re-indexes when a document changes.

What we use: pgvector or Qdrant · Keyword and meaning search, reranked · A source link on every answer
What you get: Answers with a link to the page they came from.
If you skip it: Confident answers made up from the model's general knowledge.
The sample question: It finds the delivery-change policy: changes are allowed until dispatch.

Tools it can use (MCP)

Stage 6 of 16 · Connect it to your data and tools

Give the AI tools over the Model Context Protocol: look up orders, open or update tickets, post in team chat. Read tools come first; write tools come later.

What we use: MCP servers for ERP orders, helpdesk tickets and team chat · n8n for longer workflows
What you get: The AI does the lookup instead of telling staff where to look.
If you skip it: It can describe your process but cannot check a single order.
The sample question: It calls get_order for SO-1042: To Deliver, not yet dispatched.

Role-based access in every tool call

Stage 7 of 16 · Connect it to your data and tools

Every tool call runs as the person asking, with their own ERP roles and permissions. The AI never sees more than that person could see by logging in.

What we use: The user's own ERP permissions, passed through · Row-level rules per customer, branch or team
What you get: No new back door into your data.
If you skip it: A sales rep asks about a colleague's salary, and gets it.
The sample question: The asker is on the sales team for this customer, so the lookup is allowed.

Memory and context

Stage 8 of 16 · Connect it to your data and tools

Keep the conversation's thread and a few user preferences; decide what is never remembered and when memory is cleared.

What we use: Per-user session memory · Retention rules you set
What you get: Follow-up questions work without repeating everything.
If you skip it: It forgets the order by the next message, or remembers what it should not.
The sample question: “Move it to Friday” is understood as SO-1042.

Prompt-injection guard rails

Stage 9 of 16 · Make it safe

Filter what goes in and what comes out: detect instructions hidden in documents and messages, block jailbreaks, and mask personal data before it reaches the model or the logs.

What we use: Input and output filters · Injection checks on documents and tool results · Masking for phone, PAN, Aadhaar and bank numbers
What you get: A document that says “ignore your rules” is read as text, never obeyed.
If you skip it: One poisoned PDF makes the AI send your price list to a stranger.
The sample question: The customer's phone number on the order is masked before the model sees it.

Human approval and audit trail

Stage 10 of 16 · Make it safe

Anything that changes data waits for a person to approve it. Every question, tool call and answer is logged with who asked.

What we use: An approve or reject step in the chat · An audit log you can export
What you get: The AI proposes, a person decides, and you can show it later.
If you skip it: The AI changes a delivery date on its own and nobody can say who allowed it.
The sample question: It proposes the Friday date and waits for the user's yes.

Evaluation and attack tests

Stage 11 of 16 · Make it safe

A fixed set of real questions with known good answers, run before every prompt, model or data change, plus attack tests that try to break the rules.

What we use: A question set per job · A test run on every change · An attack test set
What you get: You know a change made things better before staff see it.
If you skip it: A model update quietly breaks answers that worked last week.
The sample question: This exact question is in the order-status test set.

Cache for repeated questions

Stage 12 of 16 · Make it fast and affordable

Keep answers to questions asked again and again, such as policies and how-tos, and serve them at once. Cleared when the source changes; personal answers are never shared between users.

What we use: Exact and meaning-based matching · Cleared when the document or record changes
What you get: Faster answers and a smaller model bill.
If you skip it: You pay the model hundreds of times a day for the same leave-policy answer.
The sample question: The policy passage comes from the cache; the order lookup is live, never cached.

Router between models

Stage 13 of 16 · Make it fast and affordable

Send easy questions to a small, cheap model and hard ones to a strong model, switch to a backup when a provider is down, and stop at the budget you set.

What we use: A model router with fallbacks · Budgets per month and per team
What you get: Quality where it matters, low cost everywhere else, and no single point of failure.
If you skip it: Every “thanks” goes to the most expensive model there is.
The sample question: It goes to a mid-size model: one lookup and one policy, no deep reasoning.

Observability

Stage 14 of 16 · Make it fast and affordable

Trace every step of every answer: what was retrieved, which tools ran, which model, how long it took, what it cost and whether the user liked it. Alerts when something drifts.

What we use: Step-by-step traces · Cost per team and per job · Thumbs up or down from users
What you get: When someone says “the AI was wrong”, you can see why in a minute.
If you skip it: Complaints with no way to find the cause, and a bill with no breakdown.
The sample question: The whole answer is one trace: two tool calls, one policy passage, its time and its cost.

Channels

Stage 15 of 16 · Put it to work

Put the assistant where people already work: inside the ERP, on the website, on WhatsApp, in team chat and in email.

What we use: Inside the ERPNext desk · A website widget · WhatsApp · Team chat
What you get: People use it because it is already in front of them.
If you skip it: Another tab nobody opens.
The sample question: The sales rep asked in team chat; the answer comes back there.

Roll-out, training and running it

Stage 16 of 16 · Put it to work

Start with one team, collect feedback, train people, then widen. Each month: review the traces, re-index, update models, and report cost and use.

What we use: A pilot team first · A monthly review and report
What you get: An assistant that gets better after launch instead of going stale.
If you skip it: A launch, a burst of interest, then silence.
The sample question: This question and its thumbs-up go into next month's review.

1. Use cases and success measures

Pick three to five jobs the AI does first, such as order status, ticket triage or policy answers, and write down what a good answer looks like for each, with today's baseline.

2. Data map and sources

List every system and folder the AI may read: ERP, helpdesk, shared drives, email, chat. Mark who owns each, what is sensitive and what stays off-limits.

3. Model choice and hosting

Choose hosted models, a private cloud deployment or open-weight models on your own server, weighing your data rules, cost and quality.

4. System prompt

The standing instructions: who the assistant is, its tone, what it answers, what it refuses and how it cites sources. Versioned like code.

5. Search over your own documents (RAG)

Load your documents, split and index them, and fetch the right passages for each question, so answers quote your own sources. It re-indexes when a document changes.

6. Tools it can use (MCP)

Give the AI tools over the Model Context Protocol: look up orders, open or update tickets, post in team chat. Read tools come first; write tools come later.

7. Role-based access in every tool call

Every tool call runs as the person asking, with their own ERP roles and permissions. The AI never sees more than that person could see by logging in.

8. Memory and context

Keep the conversation's thread and a few user preferences; decide what is never remembered and when memory is cleared.

9. Prompt-injection guard rails

Filter what goes in and what comes out: detect instructions hidden in documents and messages, block jailbreaks, and mask personal data before it reaches the model or the logs.

10. Human approval and audit trail

Anything that changes data waits for a person to approve it. Every question, tool call and answer is logged with who asked.

11. Evaluation and attack tests

A fixed set of real questions with known good answers, run before every prompt, model or data change, plus attack tests that try to break the rules.

12. Cache for repeated questions

Keep answers to questions asked again and again, such as policies and how-tos, and serve them at once. Cleared when the source changes; personal answers are never shared between users.

13. Router between models

Send easy questions to a small, cheap model and hard ones to a strong model, switch to a backup when a provider is down, and stop at the budget you set.

14. Observability

Trace every step of every answer: what was retrieved, which tools ran, which model, how long it took, what it cost and whether the user liked it. Alerts when something drifts.

15. Channels

Put the assistant where people already work: inside the ERP, on the website, on WhatsApp, in team chat and in email.

16. Roll-out, training and running it

Start with one team, collect feedback, train people, then widen. Each month: review the traces, re-index, update models, and report cost and use.

Your server, your cloud or ours

The same sixteen stages in each. What changes is where your data sits, which models answer, and who keeps it running.

Show only differencesOne row is the same for all 3, hidden

Your own server, your cloud account and our managed hosting, compared
Compared on	SelectedYour own server	Your cloud account	Our managed hosting
Data sits	In your building, on your hardware.	In your cloud account, in the region you pick.	On servers we run for you, in the region you pick.
Models	Open-weight models on your own GPU server.	Claude, GPT or Gemini through your cloud's own AI service, or open-weight models.	Claude, GPT or Gemini on business terms, or open-weight models.
Used to train others' AI	Never.	Never.	Never.
You pay for	A GPU server once, then power and upkeep.	Your cloud bill, by use.	A monthly hosting fee, plus model use.
Who runs it	Your IT team, with our support.	Your IT team or us.	Us.
Not for you if	Nobody can look after a GPU server.	You have no cloud account and do not want one.	Your rules say data may not leave your building.

Selected

Your own server

If nothing may leave your building

Data sits: In your building, on your hardware.
Models: Open-weight models on your own GPU server.
Used to train others' AI: Never.
You pay for: A GPU server once, then power and upkeep.
Who runs it: Your IT team, with our support.
Not for you if: Nobody can look after a GPU server.

Your cloud account

If your IT team already runs AWS, Azure or Google Cloud

Data sits: In your cloud account, in the region you pick.
Models: Claude, GPT or Gemini through your cloud's own AI service, or open-weight models.
Used to train others' AI: Never.
You pay for: Your cloud bill, by use.
Who runs it: Your IT team or us.
Not for you if: You have no cloud account and do not want one.

Our managed hosting

If nobody in-house should have to run it

Data sits: On servers we run for you, in the region you pick.
Models: Claude, GPT or Gemini on business terms, or open-weight models.
Used to train others' AI: Never.
You pay for: A monthly hosting fee, plus model use.
Who runs it: Us.
Not for you if: Your rules say data may not leave your building.

You don't build all 16 on day one

Start with one job, prove it, then add stages.

Pilot: one job
One job, such as order status, for one team. Read-only tools, with access, guard rails, approval and tests in place from day one.
What you get
A working assistant one team uses every day, and its test results.
What it needs
10 of 16 stages
Stages 1–7, 9–11
Full company AI
All sixteen stages, several jobs, write actions behind approval, on every channel your teams use.
What you get
One assistant across teams, with cost and quality reported monthly.
What it needs
16 of 16 stages
Stages 1–16
Fix the AI you have
Already running a chatbot or copilot? We test it against the access, safety, cost and tracing stages and fix what is missing.
What you get
A written gap report, then the fixes.
What it needs
8 of 16 stages
Stages 7–14

Questions IT and finance ask first

Where the data goes, who can see what, which systems it connects to, what stops tricks, and what it costs to run and to start.

Does our data leave our servers or the country?

Only if you choose that. The model and hosting stage decides where the model runs: a hosted model in a region you pick, a private cloud deployment, or an open-weight model on your own server. Your documents, the search index and the logs stay where you keep them, and the choice is written down with its reason.

Safety

Cost and time

Where it runs: managed hosting, in your cloud account or ours. How it connects: custom integrations.

Your company's own AI, kept inside your rules.Private AI for companies: RAG, MCP tools, guard rails and observability

What your teams will ask it

What makes a company AI safe to switch on

Use cases and success measures

Use cases and success measures

Your server, your cloud or ours

Your own server

Your cloud account

Our managed hosting

You don't build all 16 on day one

Pilot: one job

Full company AI

Fix the AI you have

Questions IT and finance ask first

Does our data leave our servers or the country?

Can the AI see data an employee shouldn't?

What stops someone tricking it?

Which systems can it connect to?

How do we keep the cost under control?

How long does a pilot take, and what does it cost?

Start with one conversation

Cookies and analytics