solution ②

Internal AI Knowledge Assistant

Answers from company documents, with citations and staff access controls.

Example knowledge assistant answering a purchasing question with citations to internal policies.
Illustrative application concept

Explore an example

Open the illustrative example

Canned replay · synthetic or public data

Illustrative data

Australian example

Scores and timings are illustrative: a scripted replay, not a measured run.

Scenario 1 of 2

A state agency officer asks about Victoria’s generative AI guideline.

Corpus: Administrative Guideline: The safe and responsible use of Generative AI in the Victorian Public Sector, with its companion Guidance, Number 2024/07, Issue 1.0, made 27 November 2024; Guidance updated 19 March 2025

  1. Question

    Can I paste a difficult stakeholder’s email chain into a public AI tool to help draft my reply?

    Retrieved passages

    1. 0.88 [2] …The information you provided to the Generative AI tool is sensitive and not publicly available…
    2. 0.83 [1] Only publicly-available information can be shared with, input or uploaded into Generative AI tools that have not been approved…
    3. 0.29 not cited (Example scenario: taking notes for a meeting)

    Answer

    No. The Guidance gives this case as an example of inappropriate use: “The information you provided to the Generative AI tool is sensitive and not publicly available. The tool may inappropriately share this information with other users.”Source 2 For a tool your organisation hasn’t approved, the guideline’s rule is: “Only publicly-available information can be shared with, input or uploaded into Generative AI tools that have not been approved for use by the in-scope organisation.”Source 1 An approved tool has a limit too: “Information shared with, input or uploaded into an agency-approved Generative AI tool must not exceed the protective marking determined by the in-scope organisation as appropriate for that tool.”Source 1

    1. Retrieve hybrid (BM25 + dense), k=5, reciprocal-rank fusion 37ms
    2. Rerank cross-encoder, keep top 2 21ms
    3. Generate grounded prompt, quote and cite 1180ms
  2. Question

    If AI helps me write a briefing and gets a fact wrong, who is responsible?

    Retrieved passages

    1. 0.86 [3] …Personnel remain responsible and accountable for the accuracy and quality of their work…
    2. 0.79 [4] Relying on Generative AI content to support decision-making without verification of its accuracy…
    3. 0.71 [5] Fact-check information against verifiable sources.
    4. 0.33 not cited (Implementation: training for personnel)

    Answer

    You are. The guideline says: “In-scope organisations’ Personnel remain responsible and accountable for the accuracy and quality of their work, including advice, decisions and content created.”Source 3 It also warns: “Relying on Generative AI content to support decision-making without verification of its accuracy could be a breach of the Codes of Conduct for Victorian Public Sector Employees and this guideline.”Source 4 What the Guidance asks you to do: “Fact-check information against verifiable sources.”Source 5

    1. Retrieve hybrid (BM25 + dense), k=5, reciprocal-rank fusion 39ms
    2. Rerank cross-encoder, keep top 3 23ms
    3. Generate grounded prompt, quote and cite 1260ms
  3. Question

    Is Microsoft 365 Copilot approved for use in my department?

    Retrieved passages

    1. 0.81 [6] Well-known Generative AI tools include: … Copilot (including Microsoft 365, Edge and GitHub) …
    2. 0.57 [7] Agency-approved Generative AI tools – Generative AI software, services or features that an in-scope organisation has approved…
    3. 0.26 not cited (Example scenario: summarising a report)

    False answer, caught by the acceptance test (not independent)

    Yes, it’s approved. The guideline lists it: “Copilot (including Microsoft 365, Edge and GitHub)”Source 6.

    What the test caught: An acceptance test marks this answer false. The cited passage lists well-known tools as examples and approves none of them: each organisation approves its own tools. The expected answer says the documents don’t state it, and points to the agency’s own approved list.

    1. Retrieve hybrid (BM25 + dense), k=5, reciprocal-rank fusion 36ms
    2. Rerank cross-encoder, keep top 2 20ms
    3. Generate grounded prompt, quote and cite 640ms
  4. Question

    I work at a local council. Does this guideline apply to me?

    Retrieved passages

    1. 0.58 [8] …It applies to all public service bodies and public entities as defined in the Public Administration Act 2004…
    2. 0.31 not cited (Further information: bodies that can assist)

    Refused: the answer isn't in these documents

    These documents don’t say whether councils are covered, so I won’t guess. The scope clause reads: “It applies to all public service bodies and public entities as defined in the Public Administration Act 2004 (in-scope organisations).”Source 8 Whether a council is one of those bodies is a question about that Act, which isn’t in these documents. Escalating to a person, with the scope clause attached.

    1. Retrieve hybrid (BM25 + dense), k=5 35ms
    2. Abstain no passage says whether councils are covered, so quote the scope clause and stop 2ms

Sources

  1. [1] Guideline · Statement · safeguard 2 (p. 4)

    Only publicly-available information can be shared with, input or uploaded into Generative AI tools that have not been approved for use by the in-scope organisation. Information shared with, input or uploaded into an agency-approved Generative AI tool must not exceed the protective marking determined by the in-scope organisation as appropriate for that tool.

  2. [2] Guidance · Examples of inappropriate use · Drafting an email

    No. The information you provided to the Generative AI tool is sensitive and not publicly available. The tool may inappropriately share this information with other users.

  3. [3] Guideline · Statement · safeguard 5 (p. 4)

    In-scope organisations’ Personnel remain responsible and accountable for the accuracy and quality of their work, including advice, decisions and content created.

  4. [4] Guideline · Consequences of non-compliance · para 2 (p. 5)

    For example, as part of demonstrating Accountability under the Code of Conduct for Victorian Public Sector Employees, public sector personnel are required to make transparent accountable decisions. Therefore, the information considered in making those decisions should be accurate, available, explainable and stand up to scrutiny. Relying on Generative AI content to support decision-making without verification of its accuracy could be a breach of the Codes of Conduct for Victorian Public Sector Employees and this guideline.

  5. [5] Guidance · Maintain accountability, transparency and explainability · Do

    Fact-check information against verifiable sources.

  6. [6] Guideline · Introduction · Generative AI (p. 3)

    Well-known Generative AI tools include: ChatGPT (OpenAI); Google Gemini; Midjourney; Copilot (including Microsoft 365, Edge and GitHub); DALL.E; Synthesia; Perplexity; QuillBot

  7. [7] Guideline · Definitions · Agency-approved Generative AI tools (p. 7)

    Agency-approved Generative AI tools – Generative AI software, services or features that an in-scope organisation has approved for official work purposes. These tools are to have undergone privacy, cyber security and legal assessments in accordance with that organisation’s existing information technology policies and procedures.

  8. [8] Guideline · Scope (p. 3)

    This guideline sets out minimum requirements for the use of Generative AI in the VPS. It applies to all public service bodies and public entities as defined in the Public Administration Act 2004 (in-scope organisations). The guideline also applies to all employees, contractors, consultants and volunteers engaged directly or indirectly by in-scope organisations with access to public sector information (Personnel).

Excerpts from Administrative Guideline: The safe and responsible use of Generative AI in the Victorian Public Sector (Number 2024/07, Issue 1.0), © State of Victoria (Department of Premier and Cabinet) November 2024, and from Guidance for the safe and responsible use of generative artificial intelligence in the Victorian public sector (updated 19 March 2025), © Copyright State Government of Victoria. Both are licensed under CC BY 4.0. Sources: Guideline, Guidance. Changes: split into passages and labelled with section references for retrieval; wording unchanged. No endorsement by the State of Victoria is implied.

Scenario 2 of 2

A risk analyst asks about APRA CPS 230 notification timeframes.

Corpus: Prudential Standard CPS 230 Operational Risk Management, the version commencing 1 July 2026

  1. Question

    What are our incident-notification timeframes under CPS 230?

    Retrieved passages

    1. 0.91 [1] …must notify APRA as soon as possible, and not later than 72 hours, after becoming aware of an operational risk incident…
    2. 0.86 [2] …must notify APRA as soon as possible, and not later than 24 hours after, if it has suffered a disruption to a critical operation outside tolerance…
    3. 0.74 [3] Critical operations are processes…which, if disrupted beyond tolerance levels, would have a material adverse impact…
    4. 0.58 [4] …not more than 20 business days after entering into or materially changing an agreement for the provision of a service…
    5. 0.31 not cited (roles and responsibilities of the Board)

    Answer

    CPS 230 sets two notification clocks, plus a separate rule for service providers. An operational risk incident must be notified to APRA as soon as possible, and no later than 72 hours after you become aware of it, where you determine it is likely to have a material financial impact or a material impact on your ability to maintain critical operationsSource 1. A disruption to a critical operation outside tolerance has a shorter clock: notify APRA as soon as possible, and within 24 hours, covering the nature of the disruption, the action taken, the likely impact and the timeframe for returning to normal operationsSource 2. “Critical operations” means processes that, if disrupted beyond tolerance levels, would have a material adverse impact on your depositors, policyholders, beneficiaries or other customers, or on your role in the financial systemSource 3. Separately, service-provider arrangements have their own notice: as soon as possible and within 20 business days of entering into or materially changing an agreement for a service a critical operation relies onSource 4.

    1. Retrieve hybrid (BM25 + dense), k=5, reciprocal-rank fusion 41ms
    2. Rerank cross-encoder, keep top 4 24ms
    3. Generate grounded prompt, cite-as-you-write 1420ms
  2. Question

    And what’s the financial penalty if we miss the 72-hour deadline?

    Retrieved passages

    1. 0.34 [1] …notify APRA as soon as possible, and not later than 72 hours…

    Refused: the answer isn't in these documents

    That isn’t stated in the provided documents. CPS 230 sets the 72-hour notification obligation, but the excerpts here don’t specify any financial penalty for late notification, so I won’t guess. Escalating to a person, with the relevant clause attached.

    1. Retrieve hybrid (BM25 + dense), k=4 35ms
    2. Abstain top score below the grounding threshold, so refuse 2ms

Sources

  1. [1] CPS 230 · Operational risk incidents · ¶32

    An APRA-regulated entity must notify APRA as soon as possible, and not later than 72 hours, after becoming aware of an operational risk incident that it determines to be likely to have a material financial impact or a material impact on the ability of the entity to maintain its critical operations.

  2. [2] CPS 230 · Business continuity plan · ¶41

    An APRA-regulated entity must notify APRA as soon as possible, and not later than 24 hours after, if it has suffered a disruption to a critical operation outside tolerance. The notification must cover the nature of the disruption, the action taken, the likely impact on the entity’s business operations and the timeframe for returning to normal operations.

  3. [3] CPS 230 · Critical operations and tolerance levels · ¶34

    Critical operations are processes undertaken by an APRA-regulated entity or its service provider which, if disrupted beyond tolerance levels, would have a material adverse impact on its depositors, policyholders, beneficiaries or other customers, or its role in the financial system.

  4. [4] CPS 230 · Monitoring, notifications and review · ¶60(a)

    An APRA-regulated entity must notify APRA: (a) as soon as possible and not more than 20 business days after entering into or materially changing an agreement for the provision of a service on which the entity relies to undertake a critical operation; …

Excerpts from Prudential Standard CPS 230 Operational Risk Management (the version commencing 1 July 2026, made by Banking, Insurance, Life Insurance, Health Insurance and Superannuation (prudential standard) determination No. 1 of 2026). © Australian Prudential Regulation Authority 2024. Licensed under CC BY 4.0. Source: apra.gov.au/standards/cps-230. Answers paraphrase the cited paragraphs, which the source panel quotes verbatim; headings added for retrieval. No endorsement by APRA is implied.

Staff questions, answered with citations or refused.

The job

An internal AI knowledge assistant answers staff questions from an agreed collection of company documents. It shows the source and version, respects access permissions and declines unsupported questions.

What you receive

An assistant in the agreed workspace, plus a test report on your questions. Correct answers, wrong answers and unnecessary refusals are reported separately, with instructions for updating the documents.

Who it's for

Packages

One manual, one team

For
A team whose staff ask the same questions of one collection of manuals, policies or procedures, and need answers that show their source.
Scope
One collection for one user group, delivered inside Microsoft 365 (Copilot, SharePoint or Teams) where you license it, otherwise on the web.
Included
  • Ingestion and clean-up of the collection, with the current version of each document marked
  • Permission mapping, so staff get answers only from documents they can already open
  • Refusal behaviour tuned to the collection
  • An acceptance test of 50–100 questions written with your staff
  • An admin guide on adding a document, and a staff briefing
Every package includes
  • A test on your own examples, with thresholds agreed before testing and every failure shown, or, for AI Switch-On, the hours saved, measured against a baseline taken before anything changes.
  • Where the work has a test set, a re-run of it when the AI model changes or may have changed: by us on a system we build or run, and by your team, with the kit we hand over, after an evaluation.
  • A one-page data note: where your data sits, where the model runs, retention and deletion.
  • A runbook and a handover.
  • The delivery choice that fits the package, named in its region note.
  • Your time commitment, stated in hours and weeks.
Your time
A document owner for about 2 hours a week, 5–10 staff writing test questions for about an hour each, and access to the collection.
Timeline
4–6 weeks, with a pilot group live around week 4 and go-live after the acceptance test.
Out of scope
  • Questions about matter files, ERP data or live records
  • Legal research, clinical advice and student-facing chat
  • Procedures already held in a platform whose own AI answers from them
Go/no-go gate
Before go-live, you see the pilot's results on 50–100 staff-written questions against the thresholds agreed in week 1, and you decide whether to roll it out.
Region note
Delivered on the web, the search index and the model run in the cloud region you choose, in your own Microsoft or AWS account, or in an Australian region if we manage the assistant for you. If the model you want isn't offered in your chosen region, you hear that before you decide. Delivered inside Microsoft 365, Microsoft sets where model calls are processed, and its published statement for your geography may allow processing outside it; the data note quotes it with its date. Our team reaches the assistant remotely from Sydney, under the data-transfer terms in your contract.

We show you where your data is processed.

This page describes a packaged offer, not a delivered engagement.

SOP & Work-Instruction Assistant

SOP & Work-Instruction Assistant

For
Manufacturers, and resources and utilities field crews, whose SOPs, work instructions and OEM manuals sit outside the safety platform.
Scope
One collection of SOPs, work instructions and OEM manuals for one site, line or crew, reached on the web. Answers show the step and page used, can come in the language the worker asks in with the English step alongside, and say so when the procedure is silent.
Included
  • Ingestion and clean-up of the collection, with the current version of each document marked
  • Permission mapping, so staff get answers only from documents they can already open
  • Refusal behaviour tuned to the collection, so it says when the SOP is silent
  • Answers in each language you choose, each tested separately
  • An admin guide on adding a document, and a staff briefing
Every package includes
  • A test on your own examples, with thresholds agreed before testing and every failure shown, or, for AI Switch-On, the hours saved, measured against a baseline taken before anything changes.
  • Where the work has a test set, a re-run of it when the AI model changes or may have changed: by us on a system we build or run, and by your team, with the kit we hand over, after an evaluation.
  • A one-page data note: where your data sits, where the model runs, retention and deletion.
  • A runbook and a handover.
  • The delivery choice that fits the package, named in its region note.
  • Your time commitment, stated in hours and weeks.
Your time
A document owner for about 2 hours a week, 5–10 staff writing test questions for about an hour each, and access to the collection.
Timeline
4–6 weeks, with a pilot group live around week 4 and go-live after the acceptance test.
Out of scope
  • SOP questions about procedures that already live in SafetyCulture (now Mitti), where you try its own AI Assistant first
  • Questions about your ERP data
  • Offline answers, because the index and the model run in the cloud
Go/no-go gate
Before go-live, you see the pilot's results on 50–100 questions your crews wrote against the thresholds agreed in week 1, and you decide whether to roll it out.
Region note
The search index and the model run in the cloud region you choose, in your own Microsoft or AWS account, or in an Australian region if we manage the assistant for you, and staff reach the assistant on the web from a phone, tablet or desktop. If the model you want isn't offered in your chosen region, you hear that before you decide. Our team reaches the assistant remotely from Sydney, under the data-transfer terms in your contract.

Your data stays in your region.

This page describes a packaged offer, not a delivered engagement.

Policy & Procedure Assistant

Policy & Procedure Assistant

For
Local government, independent schools, non-clinical health services and financial-services procedure teams whose staff ask questions of a policy and procedure library.
Scope
One policy and procedure library for one staff group, delivered inside Microsoft 365 (Copilot, SharePoint or Teams) where you license it, otherwise on the web. Answers cite the policy, section and version, and say "not in our documents" when the library is silent.
Included
  • Ingestion and clean-up of the library, with the current version of each policy marked
  • Permission mapping, so staff get answers only from policies they can already open
  • Refusal behaviour tuned to the library
  • An admin guide on adding and retiring a policy, and a staff briefing
Every package includes
  • A test on your own examples, with thresholds agreed before testing and every failure shown, or, for AI Switch-On, the hours saved, measured against a baseline taken before anything changes.
  • Where the work has a test set, a re-run of it when the AI model changes or may have changed: by us on a system we build or run, and by your team, with the kit we hand over, after an evaluation.
  • A one-page data note: where your data sits, where the model runs, retention and deletion.
  • A runbook and a handover.
  • The delivery choice that fits the package, named in its region note.
  • Your time commitment, stated in hours and weeks.
Your time
A policy owner for about 2 hours a week, 5–10 staff writing test questions for about an hour each, and access to the library.
Timeline
4–6 weeks, with a pilot group live around week 4 and go-live after the acceptance test.
Out of scope
  • Clinical advice: a health service's library is non-clinical only
  • Student records and student-facing chat
  • Decisions about people, which stay with your staff
Go/no-go gate
Before go-live, you see the pilot's results on 50–100 staff-written questions against the thresholds agreed in week 1, and you decide whether to roll it out.
Region note
Inside Microsoft 365, Microsoft sets where model calls are processed, and its published statement for your geography may allow processing outside it; the data note quotes it with its date. The parts we build, such as a search index, run in your account in the region you choose, or in ours in an Australian region. Our team reaches them remotely from Sydney, under the data-transfer terms in your contract.

We show you where your data is processed.

This page describes a packaged offer, not a delivered engagement.

  • On request Ask the Firm Manual: Answers staff questions at larger accounting and law firms from the firm manual, precedents and procedures, not matter files, and shows the page it used.
  • On request PM Procedures & Onboarding Assistant: Answers new property managers' questions from your procedures manual, with the page shown.

The program

Measured knowledge layer

One tested, permission-aware source of cited answers that the assistants your staff already use, such as Microsoft Copilot, can call through a connector.

12 weeks, in three 4-week phases

  • Prove: one collection, with thresholds agreed before testing
  • Harden: permissions, logging, and a re-test on every model change
  • Hand over: the runbook, the test kit and a session for your team
  • A retention and disposal position for logs

How we test it

Acceptance test (not independent)

Every assistant is tested on 50–100 questions your staff write, with pass thresholds agreed in week 1, before anyone relies on it.

  • Correct answers, and whether the cited page says what the answer claims
  • False answers, counted on their own: confident answers your documents don't support
  • False refusals, counted on their own: "not in our documents" when the answer was there
  • The sample size in plain words: at 50 questions, one miss moves the rate by 2 points
  • A re-test on the same questions when the AI model changes

Read the evaluation method

Where it runs

  1. (a) In your own Microsoft or AWS account

    Runs in your own Microsoft or AWS account, in the cloud region you choose. Our team reaches it remotely from Sydney, under the data-transfer terms in your contract.

  2. (b) Managed for you in an Australian region

    We run it for you in an Australian region, with an exit pack of code, data, configuration and runbook. If you're outside Australia, your data is transferred to Australia under the terms in your contract.

  3. (c) Inside a platform you already license

    Runs inside software you already pay for, such as Microsoft Copilot, Xero or Dext, where the vendor sets the processing location. Where we build part of it, such as a retrieval index, that part runs in your own account in the region you choose, or in ours in an Australian region. Our team reaches it remotely from Sydney.

On the web, the assistant runs in your own Microsoft or AWS account, in the cloud region you choose, or in an Australian region we manage. Inside Microsoft 365, it works through the Copilot you already license. The search index we build runs in your account in the region you choose, or in ours in an Australian region.

The region note

  • Where inference actually runs, which can differ from where the platform is hosted.
  • The provider's retention, abuse-monitoring and human-review settings.
  • What happens when the newest model isn't available in your chosen region.
  • For a platform you already license: the vendor's published processing location, or "not published", with its source and an as-at date, re-checked quarterly.

What we don't do

Use your platform's AI first

If your staff have Microsoft Copilot licences, SharePoint's built-in AI can already answer from a document library within each person's permissions, and SafetyCulture (now Mitti) has an AI Assistant whose own examples include SOP questions. Try them first. If Copilot already does this well enough, we'll say so.

  • We don't build a chat platform. Inside Microsoft 365, the assistant uses the Copilot you already license.
  • We don't answer from matter files or ERP data, or rebuild SOP answers already inside SafetyCulture (now Mitti).
  • We don't give clinical advice or build student-facing chat.

Questions

We already have Copilot. Why would we need this?
You may not. If Copilot or SharePoint already answers your staff's questions well enough, we'll say so. What we add is a test on questions your staff write, with false answers and false refusals counted separately, and an assistant tuned to say "not in our documents" instead of guessing.
Will it show staff documents they can't already open?
No. The assistant keeps the permissions staff already have, so each person gets answers only from documents they can open today. Permission mapping is part of every build.
Have you built this for another organisation?
No. Techsider has no clients yet, and this page describes a packaged offer, not a delivered engagement. You can start with the Two-Week Trial on Your Own Files, on questions your own staff write, or, for government and enterprise, the Feasibility Sprint on public or synthetic data. Either way, you see the error rate before you rely on anything.
Where is our data processed?
It depends on where the assistant lives. On the web, the index and the model run in your own account in the cloud region you choose, or in ours in an Australian region. Inside Microsoft 365, Microsoft sets where model calls are processed, and we show you what it publishes for your geography, including whether they may run outside it. Our team reaches the assistant remotely from Sydney, under the data-transfer terms in your contract.