How to Build In-House Generative AI in India: Should You Buy Claude or Build Your Own Assistant?
By Parikshit Khanna | Checked 15 October 2026 | A practical guide for business owners in India
Quick answer: build an in-house assistant or start with Claude?
For many Indian companies, a useful starting point is a governed AI workspace or a private application around an existing model. Your custom app can use Claude’s API; buying and building can work together. Consider self-hosting when offline operation or infrastructure control justifies the extra operating work. Start with one workflow, test permissions and answer quality, and compare the full first-year cost before expanding.
First, decide what you want your company AI to do
Your sales team wants an approved product answer. Your project manager wants a clean handover. Your IT head wants to keep access under control. Your factory team wants the current procedure, not a confident answer from an old manual. Start with one of those problems and a person who can judge the result.
An in-house generative AI assistant is usually an application around an existing model. You can own its interface, document connections and approval process while a provider runs the model. You can also operate an eligible open-weight model on infrastructure you manage. Building a new foundation model is a much larger model-development programme.
The practical question is where you need control: the application, the information it can use, the workflow, the model-serving environment or all four. This guide gives you a way to decide, budget and run a controlled pilot. The recommendations are editorial guidance; every business example and internal project allocation is illustrative.
Build or buy? Compare four routes before approving a budget
Route | What you take responsibility for | When it deserves a look |
|---|---|---|
Buy Claude Team or Enterprise | Configure the workspace, users, approved sources and operating rules | Employees need knowledge-work assistance that the available product already supports |
Build a private application using a hosted API | Your app, login, retrieval, integrations, checks and support; the provider operates the model | You need a particular portal, ERP/CRM workflow or approval queue |
Self-host an existing open-weight model | The application plus model serving, infrastructure, updates and recovery | Offline operation or serving-environment control is important and you can operate it |
Train a foundation model from scratch | Data rights, large-scale training, post-training, evaluation and serving | You have a specialist model strategy, team and investment case |
A custom internal app can still use Claude
You do not have to choose between owning an application and using Claude. Your developers can build the internal interface, enforce permissions and call a hosted model API. Your company then owns or commissions the application under its development contract; it does not acquire the provider’s model weights.
A managed Claude workspace can also use approved connectors. Test a configured workspace before commissioning features it already provides. If you need a custom application, compare the additional workflow control against the engineering and support you will carry. Team and Enterprise access, API usage and partner-cloud arrangements have different terms and billing.
If your requirement is fully offline inference, make that explicit in the project brief. A locally hosted website calling an external API still sends authorised model inputs to that provider. A self-hosted tool sandbox is also a different arrangement from a locally operated model.
NLP, LLM, RAG and fine-tuning in simple English
NLP means natural language processing: software working with human language. For you, the useful outcomes might be finding a policy, classifying an enquiry or drafting a reply. You can ask for explanations in English, Hindi or a language your employees use, then test whether the meaning and business terms remain correct. Google Cloud: natural language processing
Term | Simple meaning | Why you care |
|---|---|---|
LLM | A model that works with language and generates responses | It prepares the answer, but still needs reliable inputs and checks |
Inference | Running a model to produce a response | This is the work billed by token usage or served by your infrastructure |
Tokens | Pieces of text processed by the model | They affect usage, context limits and cost; counts differ by language and model |
OCR | Turning text in scans or images into readable data | A wrongly read invoice number or manual limit can spoil the answer |
RAG | Find relevant approved material and supply it with the question | It helps the answer use company information available at that time |
Embeddings | Numerical representations used for similarity search | They are one retrieval method and may involve another service provider |
Fine-tuning | Further training an existing model using examples | It can change behaviour or task performance when a supported model permits it |
MCP | A protocol for connecting models and tools | It helps integration; your system still needs permissions and action controls |
Should you use RAG, fine-tuning or a simple document pack?
For a small initial task, a short approved source pack may be enough. As your collection grows, retrieval can select relevant passages rather than sending every document on every question. You can use keyword, vector or hybrid search; the choice should follow tests on your records.
RAG retrieves material when the question is asked. Fine-tuning changes a model through additional training. A current price list or revised policy normally needs a reliable update-and-retrieval process. Tuning a supported model may help a repeated format or task after you have suitable examples and a measured reason to do it. Microsoft Learn: RAG, retrieval and evaluation considerations
Anthropic’s current glossary says the first-party Claude API does not offer fine-tuning. Ask a vendor to identify the exact model, provider and supported method if its proposal promises “fine-tuning Claude”. Uploading documents to an application should not be presented as training a new Claude model. Claude Platform: terminology and fine-tuning availability
Current features worth knowing about in October 2026
These are available capabilities or documented deployment options checked on 15 October 2026. They are not all enabled in every account, model or region. Choose the features that solve your task and confirm their status in your proposed environment.
Capability | Useful business purpose | Important boundary |
|---|---|---|
Document citations | Let employees open the passage behind an answer | A citation does not prove a source is current or that the answer is supported |
Tool calling | Request a controlled lookup or business-system operation | Your application must authorise and execute its client tools |
Remote MCP connector | Connect an API workflow to compatible tools | The documented Claude API connector is beta; permissions and retention need review |
Prompt caching and batch processing | Reduce eligible repeated-input or asynchronous-processing costs | Check cache writes, hits, feature charges and whether the workload can wait |
Self-hosted Managed Agents sandbox | Run the agent’s tool environment on your infrastructure | Claude and orchestration can remain provider-operated; this is not local model inference |
India geographic inference on Amazon Bedrock | Evaluate in-country inference routing for listed Claude models | Confirm the actual profile, model availability and the rest of the data flow |
Read the feature documentation before promising it to staff
Claude’s citations documentation describes source-linked document answers: Claude Platform: document citations. Its tool-use guide explains the split between application-executed and server-executed tools: Claude Platform: tool-use responsibilities
The remote MCP API connector is currently beta. Its documentation marks it as not eligible for zero data retention. Confirm the exact combination of features before making a retention promise. Claude Platform: remote MCP connector and beta status
For Managed Agents, a self-hosted sandbox concerns tool execution. It does not mean you download and run Claude’s model weights. Review the control-plane and tool-input/output flow. Claude Platform: self-hosted Managed Agents sandboxes
Can you keep inference in India?
An official AWS announcement dated 29 September 2026 describes India geographic inference profiles for listed Claude models, routing between Mumbai and Hyderabad. The article names Opus 5, Sonnet 5 and Haiku 4.5. Check the current model-and-region list rather than assuming every newer release is supported. AWS: India geographic inference announcement, 29 September 2026
Treat inference location, storage and the rest of your application as separate questions. Map documents, prompts, embeddings, outputs, logs, backups, tools and administrator access. A model route in India does not by itself establish that every component of your application stays in India.
First-party Claude API geography settings and partner-cloud routing are different services. Review the exact contract and deployment configuration. Retention and model-training terms are separate again. Claude Platform: data residency and Claude Platform: API data retention and ZDR
Ask separate questions about training, retention and access
Anthropic says inputs and outputs from its commercial products are not used to train models by default, with exceptions such as explicit feedback or opt-in. Review the applicable product policy and contract. This is separate from how long information is retained and where processing occurs. Anthropic Privacy: commercial model-training policy
Ask for written answers about retention by feature, administrator and support access, subprocessors, deletion and exports. Zero data retention requires specific eligibility and configuration; do not assume every stateful feature or connector is covered. Keep the data-processing agreement and actual deployment settings with your procurement record.
For a managed workspace, review staff access and connector configuration. For your own app, also review the app database, logs and model-provider request flow. For self-hosting, your team operates those controls and still needs to check any external services it uses.
If you want self-hosting, choose a model by testing your work
Current official examples include Google Gemma 4 and OpenAI gpt-oss-20b/120b. Read the exact model card, licence and usage conditions for the version you plan to deploy. An open-weight licence does not make every part of a model’s training data or development process available.
Gemma 4’s official card describes multiple model options: Google AI: Gemma 4 model card. The gpt-oss card describes the 20b and 120b models and links their applicable conditions: OpenAI: gpt-oss model card
Test your own English, Hindi, Hinglish and other required language samples, including product codes, tables and scanned documents. Measure answer quality, unsupported claims, speed and cost with your actual context lengths. A model that loads on a machine has not yet demonstrated the capacity or reliability your employees need.
Choose a workflow that is easy to check
Write one sentence: When [person] receives [input], they need [output] so they can [business action]. Then identify the source of truth and reviewer. A quotation assistant might prepare a draft from approved product records, while a sales manager approves the price and delivery commitment.
Score candidate workflows on frequency, current effort, quality of available information and ease of checking. Prefer a narrow task with clear evidence. Ask your team to show ten real examples, including difficult cases, and create anonymised test versions. If employees disagree on the correct answer, resolve that disagreement before asking a developer to automate it.
Business area | A sensible first assistant task | Keep with the responsible person |
|---|---|---|
Sales | Prepare a draft from approved product and price information | Discounts, final prices and delivery promises |
HR and administration | Find the relevant approved policy section | Policy exceptions and employment decisions |
Accounts | Organise supplied invoice records and draft follow-ups | Reconciliation, tax treatment and payment approval |
Operations | Summarise handover notes and unresolved actions | Final priorities, owners and operational decisions |
Customer service | Draft a reply using an approved service policy | Refunds, warranty exceptions and commitments |
Manufacturing documentation | Find current SOP or maintenance-document references | Plant instructions and any equipment-control changes |
Build your assistant in eight practical stages
The following process applies to a narrow pilot. You can stop after a configured workspace trial if it meets the need. Each stage should leave something your business or technical owner can review.
1. Write the outcome and the boundaries
Name the intended users, input, required output and reviewer. Record the current time and frequent mistakes. State whether the assistant may read documents, draft text, call a lookup or propose an action. Begin with one workflow and a clear manual fallback.
Create a starter test set—for example, 50–100 representative questions as a planning choice. Include normal questions, missing information, conflicting versions and questions that a user is not allowed to answer. Agree the expected outcomes before development.
2. Prepare the information your team actually trusts
List source owners, revision dates, access rights and the system that is authoritative. Remove duplicates and obsolete copies. Check OCR, tables, units, product IDs and language. Do not silently turn a draft policy into an approved rule.
Define how an update, deletion or permission change reaches the assistant. Store the source ID and version with extracted passages. Plan removal from indexes, caches and other copies according to the applicable retention rules; deleting the original file alone may not complete the job.
3. Design login and permission checks before search
Use your organisation’s identity system where appropriate, with strong authentication and least privilege. Your backend should determine the user’s role and permitted sources. Apply document or passage permission filters before content reaches the model.
Test two people who have different rights. A sales user should not receive HR-only text in an answer, citation, search result or cached response. Repeat the test after access is revoked. A shared connector credential may have broader access than an individual employee, so examine its permissions explicitly.
4. Build retrieval and evidence into the answer
For a document workflow, the application searches authorised sources, selects useful passages and sends them with the question to the model. Return the answer with source names, locations and revision information. Make “the approved material does not answer this” a normal, useful outcome.
Check retrieval and generation separately. First ask whether the right evidence was found. Then ask whether the response follows that evidence. If retrieval misses the approved policy, improve source preparation or search before assuming a larger model will solve it.
5. Connect live systems only when the task needs them
For current stock, invoice balances or order status, consider a narrowly scoped read-only tool that queries the authoritative system. Validate its inputs and enforce access on the server. Use deterministic calculations for totals and compare results with the source records.
Keep proposed actions separate from executed actions. A draft quotation can enter an approval queue. Sending it, changing a customer record or placing an order needs the relevant authorisation. Put those checks in software and the business process rather than depending on wording in a prompt.
6. Evaluate quality, access, cost and failure behaviour
Run the same reviewed test set against your workspace, hosted-API prototype or self-hosted candidate. Record accepted answers, corrections, evidence support and reviewer effort. Include denied-access cases and deliberately misleading material in the test pack.
For example, you might require every defined access-boundary test to pass and at least 90 of 100 routine answers to meet your written rubric before a limited pilot. These are sample acceptance criteria, not an industry standard. One serious access failure should be investigated even if the average answer score is high.
Measure | What you record | Why it matters |
|---|---|---|
Retrieval quality | Whether the needed approved source was found | Separates search problems from answer problems |
Answer support | Whether each important claim follows its cited material | A fluent answer can still be unsupported |
Review effort | Time spent checking and correcting the finished work | Determines practical value |
Performance | Response times, concurrency, timeouts and queueing | Shows whether staff can use the system under expected load |
Cost | Full expense per accepted output, including review | Stops a cheap first draft hiding an expensive process |
Access and recovery | Denied access, revoked access, outages and restore tests | Shows how boundaries and fallback work |
7. Release a limited pilot with named owners
Use a small group and the approved source collection. Give users clear guidance about what the assistant can answer, what needs review and where to report a problem. Let them see the source, flag an error and return to the normal process.
Before wider use, demonstrate document updates, permission changes, backups, incident escalation and service recovery. Agree who pays the recurring bill, who handles failures and who can suspend a connection. Review the pilot evidence before adding another department.
8. Maintain it after launch
Treat the assistant as an operated business system. Re-run your evaluation pack when a model, prompt, connector, retrieval setting or important source changes. Track failures and accepted outputs, not just the number of messages sent.
Keep a record of model versions and deployment settings, maintain spending alerts and remove obsolete connections. Review source freshness and access regularly. Your business process owner and technical operator should both know how to stop the assistant and restore the usual workflow.
Security controls to put in the project brief
The controls below are engineering recommendations. OWASP’s GenAI guidance describes prompt injection and permission-related retrieval risks; it does not certify a system simply because you use RAG. OWASP GenAI: prompt injection and OWASP GenAI: vector and embedding weaknesses
Area | What you should require | A useful acceptance test |
|---|---|---|
Identity and access | Strong login, least privilege and server-side authorisation | A restricted user cannot retrieve another role’s documents |
Permission-aware RAG | Carry source permissions into retrieval and keep them current | Revoking access removes it from search, answers and citations |
Tenant separation | Separate customer indexes, caches and request context where needed | One customer cannot receive another customer’s material |
Prompt injection | Treat documents and tool results as untrusted content | A document instruction cannot override the application’s tool boundaries |
Tools and actions | Narrow tool scopes, validated inputs and enforced approvals | A read-only pilot cannot send, pay, order or change records |
Secrets and network | Keep credentials outside prompts; constrain access and outbound traffic | Sensitive tools are unavailable to unauthorised requests |
Storage and transfer | Encrypt data and control keys, backups and permitted routes | Verify actual deployment settings and restoration behaviour |
Logs and retention | Useful audit records with controlled access and deliberate retention | An authorised reviewer can trace an incident without exposing unnecessary content |
Updates and dependencies | Patch systems and assess connectors before enabling them | A model or connector change passes the regression tests |
Incident response | Named owner, contact route, containment steps and fallback | Practise disabling the affected integration and continuing normal work |
What Indian companies should check about data protection
As of 15 October 2026, India’s DPDP framework has phased commencement. The official Act notification places core provisions, including much of sections 3–17, in an 18-month phase after Gazette publication. The Rules also use staged dates; Rule 6’s security safeguards belong to the later phase. Do not describe every operational duty as already effective. Have your legal and security teams map applicable dates, sector rules and contracts. MeitY Gazette: DPDP Act commencement notification and MeitY Gazette: DPDP Rules and phased commencement
CERT-In’s directions are a separate current consideration for covered entities. They include reporting specified cyber incidents within six hours of noticing or notification and securely maintaining rolling 180-day ICT logs within Indian jurisdiction. Establish the relevant reporting process with your security team. This should not be turned into a claim that every type of company data must stay in India. CERT-In: cybersecurity directions
For deployment, document the personal data involved, purpose, permitted users, processors, locations, retention and deletion process. Check the rules that apply to your sector and customer contracts. The hosting label “private” does not establish compliance on its own.
If you are an OT or manufacturing head
Operational technology monitors or controls physical processes. Safety, availability and reliability can require different decisions from an office chatbot. For your first assistant, use approved document copies or reviewed exports in a segmented, read-only environment.
Keep PLC, SCADA, safety-system and production-parameter write access outside this pilot. Give the assistant no plant-control credentials. It can help locate an authorised procedure or summarise a handover; responsible engineers still decide what to do.
This is a recommended application of NIST’s OT-security principles, not a claim that NIST prescribes this exact AI architecture. Any later connection to operational changes needs an OT-specific assessment and the plant’s change-control process. NIST: OT Security Guide, SP 800-82 Rev. 3
Total cost: compare the whole service you will operate
Ask for a first-year budget and a recurring operating budget. Include people, data work, integration, evaluation, security, inference, hosting, storage, monitoring, support, training and taxes. An API token rate or GPU hourly price covers only part of the work.
The following amounts are worked planning examples, not market-standard prices, salary benchmarks, quotations or a proposal from Parikshit Khanna. The staffing allocations and tax/FX reserves are invented assumptions. Replace them with your scope, internal costs and written vendor quotes. Different routes below do not establish equal capacity or performance.
A provisional tax or foreign-exchange reserve is a budgeting placeholder, not a tax rate. Have your finance owner verify the actual treatment and invoice terms.
Example A: ten people use a managed Claude workspace
Using the published US Team Standard annual rate of $20 per member per month, ten seats cost $2,400 for a year before applicable taxes. At a deliberately assumed budgeting conversion of ₹90 per US dollar, the licence illustration is ₹2,16,000. ₹90 is not a live exchange rate. Regional checkout terms can differ. Claude: plans and published pricing and Claude Help: Team plan, regional pricing and seats
That licence figure is not the full adoption cost. One fictional first-year budget could allocate ₹1,20,000 to setup and enablement, ₹1,80,000 to source preparation and business review, ₹90,000 to security/access review and ₹1,20,000 to ongoing administration and support. Including the licence gives ₹7,26,000; a chosen 10% contingency gives ₹7,98,600, before any additional tax or other unbudgeted cost.
These allocations show how to build your own budget. They are not mandatory expenses or evidence that your company needs that level of spending. A narrower pilot may use existing staff and require less work; complex governance or integrations may require more.
Example B: calculate model tokens for a custom API app
Assume 10,000 requests a month, each with 4,000 input tokens and 1,000 output tokens, and one model call per request. At the current standard uncached Sonnet 5.5 rates of $2 per million input tokens and $10 per million output tokens, the calculation is: 40 million input tokens × $2 = $80; 10 million output tokens × $10 = $100; total = $180 a month.
At the same assumed ₹90/$ budgeting conversion, that is ₹16,200 a month, ₹48,600 over three months or ₹1,94,400 a year. It excludes retries, multi-step agent calls, tools, embeddings, OCR, application services, staff and taxes. Use observed token counts once you have a prototype. Claude Platform: API and feature pricing
If only 9,000 requests yield accepted outputs, token expense alone is ₹1.80 per accepted output. Calculate the full cost per accepted output after adding review and operating expenses. An employee’s chat subscription should not be treated as an unlimited backend API licence.
Example C: understand a self-hosted GPU bill
E2E Networks’ H100 page currently advertises an on-demand starting price of ₹379.05 per GPU-hour. At that displayed rate, 480 experimental GPU-hours cost ₹1,81,944. One GPU running for 8,760 hours costs ₹33,20,478 before other costs. Confirm the current configuration, availability, taxes and commercial terms. E2E Networks: advertised H100 pricing
This is a price illustration, not a recommendation that every assistant needs an H100. A smaller model or different serving pattern may need different hardware. Conversely, your workload may require additional capacity. A single GPU also supplies no separate high-availability replica.
Memory use includes more than model weights, and context length and simultaneous requests affect serving capacity. Ask for load tests with the intended model, precision, input lengths and traffic. Recheck answer quality after quantisation. NVIDIA: model-serving memory considerations
Illustrative 90-day custom-build pilot budgets
Every project allocation below is invented for planning; only the model/GPU lines follow the stated calculations. The self-hosted column assumes 480 experimental GPU-hours, not continuous service. Figures are rounded to the nearest rupee.
Cost bucket | Hosted-API app pilot | Open-weight serving pilot |
|---|---|---|
Engineering and integration | ₹4,80,000 | ₹6,00,000 |
Data preparation and connectors | ₹1,20,000 | ₹1,20,000 |
Evaluation and business review | ₹90,000 | ₹1,50,000 |
Security and access testing | ₹1,00,000 | ₹1,50,000 |
Model tokens / experimental GPU allocation | ₹48,600 | ₹1,81,944 |
Application, CPU and network services | ₹30,000 | ₹60,000 |
Storage, logs and backups | ₹30,000 | ₹45,000 |
Support, training and handover | ₹60,000 | ₹1,20,000 |
Provisional tax/FX/transaction reserve | ₹50,000 | ₹75,000 |
Subtotal | ₹10,08,600 | ₹15,01,944 |
Chosen 10% contingency | ₹1,00,860 | ₹1,50,194 |
Illustrative pilot envelope | ₹11,09,460 | ₹16,52,138 |
Illustrative first-year custom-build envelopes
This is a separate worked first-year budget including initial implementation; do not add the pilot totals above. It uses the same API traffic assumption and, for self-hosting, one H100 running continuously. It does not demonstrate that the two deployments handle equivalent work or meet your availability target.
Cost bucket | Hosted-API app | Open-weight service |
|---|---|---|
Engineering, integration and maintenance | ₹18,00,000 | ₹36,00,000 |
Data and connectors | ₹4,00,000 | ₹6,00,000 |
Evaluation and review | ₹3,00,000 | ₹6,00,000 |
Security and access operations | ₹3,00,000 | ₹5,00,000 |
Model tokens / GPU serving | ₹1,94,400 | ₹33,20,478 |
Application, CPU and network | ₹1,20,000 | ₹3,00,000 |
Storage, logs and backups | ₹1,20,000 | ₹2,40,000 |
Support and enablement | ₹3,00,000 | ₹6,00,000 |
Provisional tax/FX/transaction reserve | ₹1,00,000 | ₹3,00,000 |
Subtotal | ₹36,34,400 | ₹1,00,60,478 |
Chosen 10% contingency | ₹3,63,440 | ₹10,06,048 |
Illustrative first-year envelope | ₹39,97,840 | ₹1,10,66,526 |
How you can control cost without hiding work
Route simple tasks to a model that passes the required quality tests; reserve more expensive reasoning for tasks that justify it.
Retrieve relevant authorised passages instead of sending the whole library every time.
Evaluate eligible caching and asynchronous batch options, including write/read charges and latency requirements.
Limit output length, tool loops and retries, and set spending alerts at the application and provider levels.
Use measured concurrency and operating hours to size self-hosting; include idle capacity, recovery and any replicas.
Review the full cost per accepted business output. Retain the human checks needed for the task.
Whom should you hire?
Assign responsibilities before advertising job titles. Your first business and document owners may already be employees. For a custom application, you need application engineering, identity, data and operating ownership in addition to prompt design. Add specialist ML-serving skills before committing to self-hosting.
Role or responsibility | What you need delivered | When to involve them |
|---|---|---|
Business process owner | Scope, correct outcomes, reviewer and go/no-go decision | Before procurement or development |
Technical lead / solution architect | Deployment choice, data flow, integration and operating plan | During discovery |
Application / integration engineer | Working app, authentication, connectors, approvals and failure handling | For a custom application |
Data engineer / document owner | Reliable ingestion, versions, permissions, updates and deletion | When the source collection needs it |
Domain reviewer / evaluation owner | Representative test set and repeatable acceptance checks | Before and throughout the pilot |
Security / privacy / legal support | Threat review, access tests, applicable obligations and contract checks | Early enough to change the design |
Platform / ML-serving engineer | Model deployment, load testing, monitoring, patching and recovery | For self-hosted inference |
Support and change owner | User guidance, issue handling and ongoing adoption | Before employees depend on the service |
Questions to ask a developer or implementation agency
Which part will we own: repository, cloud accounts, pipelines, configuration and evaluation set? Put ownership and access in the contract.
Which exact model and licence will you use, and can we change provider or export our data?
Can you demonstrate our real test cases, including denied access, changed permissions and missing evidence?
Where does each type of data go: inference, embeddings, tools, logs, backups and support access?
Which operations can the system execute, and where is approval enforced?
What is included in the quote, and which recurring charges are excluded?
What load and availability assumptions did you use? Show measured results rather than a hardware-fit claim.
Who handles an incident, a failed connector, a model retirement or an employee leaving?
Can our team restore the service and operate it after handover? Show the runbook and exit plan.
Use a staged six-to-ten-week plan
Treat this as a suggested sequence, not a guaranteed delivery time. Give each stage a visible decision. You can stop, simplify or repeat a stage when the evidence is weak.
When | Work and deliverable | Go / no-go outcome |
|---|---|---|
Weeks 1–2 | Name the business owner, interview users, select one workflow and measure the current process. Produce a scope and anonymised test set. | Go when the problem and expected answers are clear. Pause if the task or reviewer remains undefined. |
Weeks 3–4 | Test the simplest approved approach. Record drafts, source evidence, corrections and full task time. | Go when quality is acceptable and the team sees a repeatable benefit. Fix source problems before adding software. |
Weeks 5–6 | If justified, build a narrow prototype with the agreed document collection, user roles and approval steps. | Go when the prototype passes the agreed tests, including missing evidence and access boundaries. Otherwise revise or stop. |
Weeks 7–8 | Run a limited user pilot. Practise document updates, support requests and returning to the normal process. | Go when users can operate it and named owners can maintain it. Delay expansion if support or ownership is unclear. |
Weeks 9–10, if needed | Resolve remaining issues, document operating instructions and decide the next workflow. | Expand only with accepted results, a support owner and an agreed operating budget. Keep the scope narrow otherwise. |
Three fictional decisions a business owner might make
These teaching scenarios are not client case studies or promised implementation outcomes.
Fictitious example: a 20-person distributor
You have three salespeople who repeatedly ask the owner for product details and quotation wording. Your approved catalogue and price sheet are current, but enquiry notes are inconsistent.
Begin with a small, reviewed drafting pilot in an approved tool. Standardise enquiry fields and use the same approved documents. You do not yet need to commission an application merely to avoid rewriting a reply.
Consider a custom assistant later if the tested workflow needs controlled access to many records, a shared interface or a specific approval queue. Ask whether improving the catalogue and enquiry form would solve the remaining problem more simply.
Your success check: Compare total preparation and review time on similar quotations. Continue only if prices, specifications and promises remain correct and another employee can repeat the method.
Fictitious example: a 200-person service company
You have several delivery teams using different versions of project templates. Employees need answers from approved internal documents, and managers struggle to trace where a summary came from.
Choose one department and one document collection. Agree which version is authoritative and which users should see it. Run a question-answer test with expected answers, source references and examples that should receive an unknown response.
A controlled internal assistant may be worth piloting if permissions, document refresh, answer evidence and a consistent user interface are essential. Keep the initial scope small instead of connecting every department at once.
Your success check: Test correct answers, source traceability and access boundaries with real user roles. Proceed only when the document owner can maintain the source collection and the support owner can handle problems.
Fictitious example: a manufacturing operations or OT head
You want shift teams to find approved instructions and summarise handover notes. Some information belongs to office systems; other information concerns operational technology and controlled plant processes.
Start with an isolated, read-only pilot using approved, non-sensitive copies of documents. Keep plant control and equipment commands outside the pilot. Ask authorised engineering, quality and operations owners to define acceptable answers and escalation points.
Consider a custom interface only after the document task works and the responsible technical team has reviewed the proposed system boundaries. A successful office-document demonstration does not establish readiness for a plant-connected workflow.
Your success check: Check revision accuracy, source references, handling of conflicting instructions and refusal to guess missing limits. The pilot should direct an employee to the authorised procedure or person when evidence is insufficient.
Copy this project brief before you ask for a quote
Complete the fields below with your business and IT owners. A useful supplier response should explain the assumptions it has made and the evidence it will deliver.
Project: [one workflow] assistant pilot.
Business problem: [specific delay or repeated task].
Users and pilot size: [roles and number].
Current process: [steps, approximate time and common errors].
Required output: [format, necessary facts and supporting source references].
Approved sources: [documents/systems, owners and version dates].
Data boundaries: [permitted fields, excluded information and user access rules].
Allowed actions: [draft/read-only outputs].
Actions requiring human approval: [sending, publishing or changing records].
Unknown-answer rule: State what is missing and route to [responsible person].
Test set: [representative anonymised cases, expected answers and difficult examples].
Acceptance measures: [agreed quality standard, full task time and reviewer effort].
Operating plan: [document updates, issue reporting, named support owner and fallback process].
Commercial review: [development effort, account costs, maintenance responsibilities and approval for further work].
Go/no-go meeting: [date and decision-maker].
Deliverables: Working prototype, test results, operating instructions and handover.
Out of scope: [excluded departments, connections and decisions].
Before wider rollout, ask for these deliverables
A working demonstration on your agreed cases, with evaluation results and known limitations.
A data-flow and permission design that your technical and security owners have reviewed.
Document-update, deletion and access-revocation tests.
An operating budget with traffic, staffing, hosting and support assumptions.
A restore procedure, incident process, named support owner and manual fallback.
Company access to the agreed code, configuration, accounts and handover material.
Discuss the right first step for your company
If you want help defining business use cases, comparing tools or preparing your employees for a pilot, speak with Parikshit Khanna. Start with the work you want to improve, the systems you use and the decisions your team must retain.
Parikshit Khanna is a corporate AI trainer, TEDx speaker and founder of Digital Training Jet, with 3,50,000+ professionals trained. Discuss an owner/leadership workshop, practical Claude training or a use-case and readiness session for your team in Delhi NCR, Mumbai, Kolkata or online. Onsite dates and logistics are agreed after your enquiry.
Request a workshop or readiness-discussion proposal: Request a workshop proposal. View current private sessions: View coaching and sessions on Topmate. Share your business type, team size, location, current tools and two workflows you want to assess.
For enquiries, use WhatsApp Parikshit Khanna or email parikshitkhanna@digitaltrainingjet.com. The guidance is independent; any software-build engagement needs its own agreed scope, technical responsibilities and commercial terms.
Frequently asked questions
Can your company build an assistant that works like Claude?
You can build a conversational assistant with company sources, permissions and workflows using an existing hosted or open-weight model. That is different from reproducing Claude’s underlying model. Define the business task and compare a managed workspace, a custom API app and self-hosting before choosing.
Does in-house AI mean your information never leaves your company?
Only if the complete design supports that boundary. A private interface may still call an external model, embedding provider or tool. Map the entire data flow and check contracts, logs, backups and support access. A locally hosted application alone does not establish an offline system.
Do you have to train a new model on all your company documents?
Usually you can start with approved context or permission-aware retrieval. RAG supplies relevant material when a question is asked. Fine-tuning is additional model training for supported models and tasks. Claude’s first-party API currently does not offer fine-tuning.
Is a vector database compulsory?
No. A small approved document pack or a suitable keyword search may meet the first task. Larger collections can benefit from vector or hybrid retrieval. Test whether the right authorised evidence is found before choosing the search stack.
Can you run Claude itself on your own server?
The deployment options discussed here do not provide Claude model weights to run offline. A custom app can call Claude, and certain agent sandboxes can run tools on your infrastructure. If local model inference is required, evaluate an eligible open-weight model and its licence.
Can you use Claude inference in India?
AWS documents India geographic inference profiles routing listed Claude models between Mumbai and Hyderabad. Check current model availability and the chosen profile. Separately verify storage, logs, embeddings, tools and other processing; an inference route is not an all-component residency guarantee.
How much will an in-house GenAI assistant cost in India?
There is no responsible fixed total without your scope, traffic, sources, controls and support needs. This guide shows transparent licence, API and GPU calculations plus illustrative full budgets. Replace the invented project allocations with your internal costs and written quotations.
Do you need a large AI research team?
A managed-workspace trial may use existing business, document and IT owners. A custom app needs engineering and operating support. Self-hosting adds model-serving expertise. Training a foundation model is a separate specialist programme; it should not be assumed necessary for a first internal assistant.
Can your first assistant connect to factory controls?
Start with a segmented, read-only document or reviewed-export pilot. Keep plant-control credentials and PLC/SCADA write access outside it. A later operational-control use case requires an OT-specific assessment and the organisation’s engineering change process.
When should you choose a paid Claude workspace rather than build?
Choose a managed-workspace pilot when its actual capabilities and controls meet the task. Consider a custom app when a tested need for your own workflow, interface or integration justifies the operating burden. Use the same reviewed examples to compare routes.
Can one prompt make the assistant secure?
A prompt helps describe behaviour, but security also needs identity, backend authorisation, source permissions, tool boundaries, monitoring and tested recovery. Treat retrieved content as untrusted and test what the system does when documents contain misleading instructions.
How can your team get help with the decision?
Use the contact or Topmate links in this guide to discuss a business-use-case or readiness session with Parikshit Khanna. Share your users, systems, data boundaries and two repeated tasks. Confirm the workshop scope and any separate implementation responsibilities before booking.
Official references and update note
Product references were checked on 15 October 2026. Features, access and terms can change. Company scenarios and budget allocations are illustrative; dated vendor rates are identified separately. Select a workflow through a controlled pilot and confirm current provider terms before buying or implementing a system.