A private knowledge assistant looks simple in a demo: upload the documents, ask a question, get an answer with a citation. In production, the hard part is not the model. It is deciding which documents are current, who is allowed to see them, how long they are retained and what happens when the answer is wrong.
This is why so many internal assistants fail after the pilot. The model is adequate; the knowledge base is not. For a Bangladesh business, the Personal Data Protection Act, 2026 adds another layer: personal data must be handled with a lawful basis, purpose limitation, retention rules and appropriate security. A private assistant that retrieves everything for everyone is a compliance incident waiting for a question.
Where things stand in 2026
Retrieval-augmented generation is now the default architecture: retrieve relevant chunks, add them to the prompt and ask the model to answer with citations. It works well when the corpus is clean and permissions are simple. It fails when documents are duplicated, outdated or confidential. Most teams underestimate the cleanup by a factor of three.
The market is also noisy. Every productivity suite now offers an assistant, and every vendor claims "enterprise-grade security". The useful questions are narrower: where is the data stored, who can access it, is the content used for training, how are permissions inherited, how are deletions handled and can the organisation export its index at exit?
2027: make permissions and ownership real
2027 should be the year of the boring foundation. Build a document inventory with an owner for every source. Classify content into public, internal, confidential and personal data. Remove duplicates and superseded versions. Define a freshness rule: policies reviewed annually, price lists monthly, product manuals per release. An assistant that cites a 2023 policy is worse than no assistant.
Permission-aware retrieval is the core requirement. The assistant must inherit the access rules of the source system, not a separate allow-list that drifts. If a user cannot open a document in the source system, the assistant must not quote it. Test this with real users and real restricted documents, not with a demo account.
- Data processing terms: confirm the vendor's role, subprocessors, retention and breach notification.
- Residency: know where the index and the model run. The PDPA's data residency provisions may affect your design.
- Evaluation: build 50 to 100 questions with expected answers and required citations.
- Feedback: give users a way to mark a bad answer and route it to the document owner.
2028: connect structured and unstructured knowledge
By 2028, the assistant should answer questions that span documents and systems: "Which customers have overdue invoices and an open support ticket?" That requires controlled access to structured data, not just PDFs. The architecture becomes a retrieval layer plus a tool layer, with permissions and audit logs across both.
This is also the year to add actions. An assistant that can draft a reply is useful. An assistant that can update a CRM record, issue a refund or send a quotation needs approval rules, idempotency and a complete audit trail. Start with draft-and-approve. Add direct execution only for low-consequence actions with a rollback path.
Multi-source retrieval raises the quality bar. The assistant must show which source it used, when that source was last updated and whether the user has permission to see it. A confident answer without provenance is a liability.
2029: consolidate and make it portable
2029 is a consolidation year. Most SMEs will have several assistants: one in the intranet, one in the helpdesk, one in finance and one embedded in a productivity suite. Users will not know which one to trust. Consolidate around a common knowledge layer, a common permission model and a common evaluation set.
Portability becomes a procurement requirement. Can you export the index? Can you change the model without rebuilding the corpus? Can you run the same evaluation across vendors? Can you delete a document everywhere it appears? A knowledge assistant without an exit path is a long-term dependency, not a tool.
Retention and deletion need to be tested, not just documented. Ask the vendor to show how a deleted document disappears from the index, the cache and the logs. If the answer is "within 30 days, maybe", that is a compliance risk.
2030: knowledge as infrastructure
By 2030, the assistant should be part of the company's knowledge infrastructure, not a separate product. New documents enter a governed pipeline. Owners are accountable for freshness. Access reviews happen on a schedule. Quality is measured with the same discipline as a service desk: answer accuracy, citation rate, stale-answer rate and time saved.
The most valuable use cases will be narrow and repetitive: onboarding, policy questions, product specifications, supplier terms and troubleshooting. The assistant will not replace judgment. It will reduce the time spent finding the document that contains the judgment.
What to do in the next 90 days
- List the ten sources people ask about most often. Assign an owner and a review date to each.
- Classify the content and remove duplicates, drafts and superseded versions.
- Choose one permission boundary and test it with a restricted document.
- Build a 50-question evaluation set with expected answers and citations.
- Run a pilot with 10 to 20 users and a weekly review of bad answers.
- Write the retention, deletion and exit terms before signing.
How to measure progress
- Answer accuracy: the share of evaluation questions answered correctly.
- Citation rate: the share of answers that cite an accessible, current source.
- Stale-answer rate: answers based on documents past their review date.
- Access violations: attempts to retrieve content the user is not allowed to see.
- Time saved: measured against a baseline task, not self-reported enthusiasm.
The assistant will only be as good as the knowledge it is allowed to use. Treat the corpus, permissions and ownership as the product. The chat window is just the interface.
Where this fits in our work
If you want help applying this to your business, start with Business Automation. We publish scope and starting prices on the pricing page. For the next step in this series, read AI governance policy: 2027, 2028, 2029, 2030.
Sources
- DataGuidance, "Bangladesh: Personal Data Protection Act 2026", accessed 16 November 2026.
- NIST, AI Risk Management Framework, accessed 16 November 2026.
- ISO/IEC 42001, AI management systems, accessed 16 November 2026.
- NIST Privacy Framework, accessed 16 November 2026.
- McKinsey, "The State of AI: Global Survey 2026", 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.