Home · Solutions · Other solutions

Solution · Other solutions

Every answer carries the page it came from, and every unanswered question finds a page owner

One place to ask, an answer with its source, gaps reported

Employees ask in Microsoft Teams, the agent answers only from indexed sources and cites the page and its date, and every question it cannot answer becomes a task for the owner of that page.

DepartmentalMicrosoft TeamsHuman in the loopAI where it earns its place
1,900questions a month of the how-do-I and where-is kind reach the service desks and Teams channels of this illustrative engineering consultancy.

Executive summary

Challenge

The same questions are asked in channels and tickets because nobody can find the current answer.

What changes

The agent is the smaller half of this solution; the loop that repairs the documentation is the half that makes it hold.

Business value

Everyone receives the same sourced text, so the answer stops depending on who replies first.

Systems involved

SharePoint and Confluence Cloud pages updated by their owners; ServiceNow tickets for what the sources cannot answer; Microsoft Planner task queues

Business problem

Knowledge

Enterprise knowledge is rarely missing. It is scattered, unversioned and unranked. Policies live on SharePoint, the engineering truth in Confluence, the service desk keeps its own articles, and every team has a thread where the real answer was given once, to one person. Search returns everything and confirms nothing, so people ask a colleague, which costs two people's time, or raise a ticket, which costs a queue.

When a rule changes, the old page stays live. The new rule travels in an email, a slide and a meeting, while the page everyone finds first still says the old thing. Owners never learn which pages are wrong, because nothing connects the failed question to the page that should have answered it.

The people who pay are the ones you can least afford to interrupt. Twelve service desk agents answer questions alongside real requests and report the lot as workload, so every capacity plan is wrong by the share that was never a request. New joiners learn to trust whoever replies fastest rather than the documented rule, which is how a stale expense limit becomes a rejected claim.

How it works today

The route is the same in most organisations, and nothing it touches improves for the next person.

  1. PersonAn employee searches SharePoint, gets forty results and cannot tell which page is current
  2. PersonThe question goes into a Teams channel; two colleagues answer differently, one from a page superseded last spring
  3. WaitingThe employee waits, then raises a ticket because nobody in the thread sounded certain
  4. PersonAn agent hunts for the knowledge article, answers from memory or asks a specialist, and closes the ticket with free text
  5. Risk of errorThe resolution text holds the right answer and stays inside the ticket, where nobody will search for it
  6. WaitingThe same question arrives next week from somebody else, and the page that should have answered it is never touched
PersonWaitingRisk of error

Why the current process costs more than it appears

Nobody planned this work; it accumulated.

  • Each repeated question is billed three times: the asker's search, the colleague's interruption and, often, a ticket the service desk counts as demand.
  • Inconsistency is the expensive part. A wrong answer about a security rule or an expense limit becomes an incident or a rejected claim, never traced back to the page that caused it.
  • Specialists become the knowledge base in person, so the organisation runs on memory and holiday calendars.
  • Nobody can size the gap, because unanswered questions leave no trace, so the clean-up has no order and never starts.
  • Reporting mixes questions with requests, so the capacity plan is built on a number that includes work no request system should have received.

Cost of inaction

A year of answering what the pages should already say≈ €228,000
The span of three annual policy cycles, unchanged≈ €684,000
Twelve months once the group asks 2,300 questions a month≈ €276,000

Headcount, tools and reorganisations all add questions, and none of them adds documentation. The channels absorb the extra silently, the service desk keeps a fixed share of tickets that were never requests, and the capacity plan stays wrong by that share. Documentation decays at the speed policies change, and the decay stays invisible until an audit finds the live page two versions behind the rule people were following.

The part these rows cannot price is the answer that was simply wrong. A superseded expense limit produces a rejected claim, a superseded security rule produces an incident, and neither is traced back to the page. Employees who expect an assistant will meanwhile find an ungoverned one on their own phone, a governance problem arriving from a direction nobody budgeted for.

Illustrative scenario

A plausible organisation with realistic proportions. The figures are there to be recalculated on your data; they are not a client result.

Organisation

An engineering consultancy with 4,200 employees in three countries; policies on SharePoint (about eleven thousand pages), engineering standards in Confluence Cloud, service desk articles in ServiceNow, Microsoft 365 and Microsoft Teams in daily use.

Volume

About 1,900 questions a month of the "how do I", "where is" and "what is the rule" kind reach the IT and HR service desks and the main Teams channels; twelve agents answer them alongside real requests, and a knowledge manager keeps a list of pages she suspects are stale.

Current process

People search, then ask in a channel, then raise a ticket. The answer comes from whoever replies first or from an agent's memory, and the resolution text stays inside the ticket, where the next asker will not find it.

Bottleneck

Eleven minutes of answering effort per question by an agent or a colleague, plus about six minutes the asker spends searching and waiting, and no signal from the failed question back to the page that should have answered it.

Solution

One place to ask in Microsoft Teams, answering only from indexed sources the asker may open and citing the page with its last-changed date; anything not covered becomes a ServiceNow ticket and a logged gap, and each week's gaps reach page owners as Microsoft Planner tasks with a draft update.

Potential outcome

Questions the documents already answer are settled in seconds at any hour, the service desk separates questions from requests for the first time, and the twenty pages that would remove the most questions are named and repaired inside a quarter. Every quantity here is a model, built on stated assumptions.

Proposed solution

The agent is the smaller half of this solution; the loop that repairs the documentation is the half that makes it hold, and we treat that loop as the product. One place to ask is a UiPath conversational agent published to Microsoft Teams as a Teams app, answering from UiPath Context Grounding indexes over your SharePoint sites and Confluence Cloud spaces. Those indexes inherit the permissions of their sources, so the agent cites only what the asker could have opened.

Every answer carries the page it came from and the date that page was last changed, which is what lets a reader judge it. Where the sources hold no answer, the agent says so, offers a ServiceNow ticket with the conversation attached, and logs the question as a gap. Those refusals are the valuable output: the only honest measurement of what your documentation fails to cover.

Weekly, UiPath Communications Mining clusters two populations by intent: questions the agent could not answer, and tickets that were really questions. Ranked by volume, the list opens with the pages that would remove the most questions rather than the pages someone remembers. A robot turns each cluster into a Microsoft Planner task for the page owner, with the questions attached and, where a closed ticket holds the answer, a draft to accept, edit or reject. Nothing publishes itself.

Where you hold Microsoft 365 Copilot licences, a Microsoft Copilot Studio agent with SharePoint knowledge can carry the same answers, the loop unchanged underneath. Power BI shows the answer rate with a source, the top unanswered intents, the failing pages and the owners who have not acted.

Native capabilities used

UiPath conversational agents in Microsoft Teams; UiPath Context Grounding indexes over Microsoft SharePoint, OneDrive and Confluence Cloud with inherited permissions; UiPath AI Trust Layer allow-list, PII masking, EU routing and prompt logging; UiPath Communications Mining intent clustering; Microsoft Planner tasks in Teams; Microsoft Purview audit and retention; Power BI; optionally a Microsoft Copilot Studio agent with SharePoint knowledge

What we build

The index scope and refresh rules, the agent instructions, citation format and refusal behaviour in each working language, the gap log, the weekly clustering and ranking, the owner-task robot with its drafts and chasing, the service desk escalation and the Power BI dashboard

Custom integration

ServiceNow tickets and knowledge articles through the UiPath Integration Service connector; Confluence Cloud spaces indexed by Context Grounding; Microsoft Planner tasks and page metadata through Microsoft Graph and the Microsoft OneDrive & SharePoint connector

How the automated process works

  1. PersonAn employee asks in Microsoft Teams, where they would have asked a colleague
  2. AutomationThe agent retrieves from the Context Grounding indexes, trimmed to what that person may open, and answers with the source page and its date
  3. AutomationIf the sources carry no answer, the agent says so, offers a ServiceNow ticket with the conversation attached, and logs the question with its intent
  4. AutomationEach week UiPath Communications Mining clusters the unanswered questions and the tickets that were really questions, and ranks the clusters by volume
  5. SystemA robot creates a Microsoft Planner task for each affected page's owner, with the questions attached and a draft where a closed ticket holds the answer
  6. PersonThe owner edits or rejects the draft and publishes the page in SharePoint or Confluence Cloud, where versioning keeps the previous text
  7. AutomationThe index refreshes, the dashboard records the closed gap, and the next person who asks gets the answer with its new date
PersonAutomationSystem

Human-in-the-loop model

Automation handles

  • Answering from indexed sources with the page and its date, or declining and offering a ticket
  • Logging every question with its outcome, intent and confidence
  • Clustering unanswered questions and question-tickets weekly and ranking them by demand
  • Creating and chasing owner tasks, drafting updates from ticket resolutions, refreshing the index after a change

People decide

  • What a page should say: owners accept, edit or reject every draft, and nothing publishes without them
  • What the documentation cannot answer, resolved by the service desk as a normal ticket
  • Which sources belong in scope and which questions the agent must never attempt, set by the knowledge manager
  • Whether an HR or security page is fit to be indexed, decided by the team that owns the policy

Before and after

BeforeAfter
Time to an answerhours, sometimes two daysseconds, in Teams, at any hour
What the answer rests onwhoever replied first, or memoryan indexed page, cited with its last-changed date
Questions the documentation failedinvisiblelogged, clustered weekly and ranked by volume
Who repairs a wrong pagenobody, until an audit finds itthe named owner, with the questions and a draft in a task

Systems and integrations

Where a rule suffices we do not use a model. Where judgement is needed, a person decides.

Inputs

  • questions asked in Microsoft Teams
  • SharePoint policy pages
  • Confluence Cloud engineering spaces
  • ServiceNow articles and closed question-tickets
  • page metadata with owners and dates

Automation layer

  • UiPath Orchestrator
  • UiPath conversational agents
  • UiPath Context Grounding
  • UiPath Communications Mining
  • UiPath Integration Service

Target systems

  • SharePoint and Confluence Cloud pages updated by their owners
  • ServiceNow tickets for what the sources cannot answer
  • Microsoft Planner task queues
  • Power BI

Human touchpoints: the agent conversation in Microsoft Teams; owner tasks in Microsoft Planner; service desk escalations; the knowledge manager's weekly review of the gap list

questions asked in Microsoft TeamsUiPath OrchestratorUiPath conversational agentsSharePointthe agent conversation in Microsoft Teams

Technologies used

UiPath conversational agents in Microsoft Teams

one place to ask, published to Teams as a Teams app

A
UiPath Context Grounding

permission-inherited indexes over SharePoint, OneDrive and Confluence Cloud

A
UiPath AI Trust Layer

allow-listed model, PII masking on prompts, EU region routing, prompt and answer log

A
UiPath Communications Mining

weekly intent clustering of unanswered questions and question-tickets

A
UiPath Robots + Orchestrator + Integration Service

owner tasks, drafts, chasing, index refresh; ServiceNow and Microsoft connectors

A
Microsoft Planner and Microsoft Graph

the owner's task queue inside Microsoft Teams, with due dates

A
Power BI

answer rate with a source, top unanswered intents, failing pages, owner response times

A
Averified product capability (vendor documentation)

Illustrative economic model

Start by questioning the assumptions.

Illustrative model
1,900 questions a month × 12 minutes of combined working time= 380 h / month
380 h × €50 blended fully loaded hourly cost= €19,000 / month
× 12 months= €228,000 / year
Annual pool of working time spent answering questions the documentation should answer (illustrative)≈ €228,000

The 55% is the number worth arguing about, so it comes first: once the ranked gaps are closed, the model assumes that share of questions can be answered from documentation, measured by a pilot on one month of your real questions. Two pools sit under it. Answering effort: 1,900 questions a month at 11 minutes by an agent or a colleague, €47 an hour fully loaded, of which 55% is addressable. Asker effort: the same questions at about 6 minutes of searching and waiting, at €54 an hour. Blended, that is roughly 12 minutes per question at €50, and the rounding moves the annual figure by about one and a half percent. Wrong answers, onboarding and Copilot Credit consumption stay outside the model, and nothing here was measured at a client.

Run the numbers on your data

hours released per month
of annual capacity released

An illustrative estimate from your own inputs. It models released capacity; it is not a promise of savings.

Business benefits

  • Everyone receives the same sourced text, so the answer stops depending on who replies first
  • Time to an answer falls from a thread that runs for days to seconds in Teams, with a page and date the asker can check
  • Service desk capacity returns as questions stop arriving as tickets, and reporting separates questions from requests
  • Documentation improves in the order demand dictates, because the gap list is ranked by how often each answer was needed
  • Stale pages are found by the questions they fail, not by an annual review nobody finishes

The management view

  • Question volume becomes a measure of where the organisation is confused, worth reading before a policy roll-out or a system change lands
  • The top unanswered intents are visible weekly, with the pages and owners behind them, so knowledge work is prioritised like any other backlog
  • Every indexed page has a named owner, a task queue and a response time, and service desk demand splits into questions and requests

Board-level KPIs

share of questions answered with a sourceescalation rate to the service deskowner task completion and time to close a gapquestion-tickets as a share of service desk volumepages updated per month against pages that failed

Security and governance

Where the data sits and who can see it.

  • The index inherits the permissions of SharePoint and Confluence Cloud, so a question about a document the asker may not see returns nothing rather than a summary
  • Prompts and answers pass through the UiPath AI Trust Layer: an allow-listed model, PII masking, processing kept in the EU region, and a log of every exchange retained under Microsoft Purview with the Teams conversation
  • The agent answers only from indexed sources and declines otherwise; HR and security pages enter the index once the team that owns the policy confirms them
  • Drafts never publish themselves. A page changes when its owner approves it, where versioning keeps the previous text and the audit trail shows who changed what, and when

Why now

01

Employees compare an internal tool with the assistant on their own phone. A channel that answers with a source is the expectation, and the alternative is not "no assistant" but an ungoverned one nobody approved

02

Retrieval over your own documents is now a platform feature rather than a project: UiPath Context Grounding indexes SharePoint and Confluence Cloud with permissions inherited, so the weak point is the documentation and the loop that maintains it

03

The modelled €19,000 a month is paid a few minutes at a time by people who never report it, and it grows with every tool, policy and reorganisation the year adds

Relevant executive roles

CIO

A governed answer channel with sources and an audit log, instead of the ungoverned assistant employees will otherwise use

COO

Fewer interruptions of the people who know, and consistent answers in place before a process change lands

CHRO

Policy questions answered from the current policy, with a record of what was told to whom and when

Head of Shared Services

Question volume separated from request volume, and a service desk that stops being a search engine

Common questions and objections

Microsoft 365 Copilot already answers from SharePoint.

For a licensed user it does, and where you hold those licences we can put the same answers there. What it does not do is log the questions it could not answer, cluster them by intent, route them to the page owner and chase the fix. That loop is what we add, on top of Copilot rather than instead of it.

A wrong answer is worse than no answer.

Which is why the agent cites and dates every answer, draws only on sources you approved, and declines when they are silent. A refusal costs one ticket; a confident invention costs trust.

Our documentation is a mess.

That is the argument for starting, not for waiting. The gap list ranks the clean-up by demand, so the first month names the twenty pages that would remove the most questions.

When this is not the right solution

  • Documentation so thin that less than about a fifth of questions could be answered from it; write the pages first, then index them
  • No page owners willing to accept tasks: without the loop this is search with citations and the pages stay wrong
  • A single-domain need such as HR policy alone, better served by a narrower agent with its own validation rules

A question for the next management meeting

When our last policy change went live, how long did the superseded page keep answering people, and who in this company could have told us that it was?

Implementation approach

We start with one slice of the process and extend only once it is proven.

We deliver

  • A classification of one month of your real questions and question-tickets, which sets the answerable share before anything is built
  • Index scope, refresh rules and the permission model, agreed with the owner of every source, and the agent itself: instructions, citation format, refusal behaviour, working languages, guardrails and the question bank it is tested against
  • The gap loop: logging, weekly clustering in UiPath Communications Mining, ranking, owner tasks in Microsoft Planner and the draft robot
  • The escalation into ServiceNow with the conversation attached, so nothing the agent declines is dropped
  • The Power BI dashboard, owner training and the weekly review routine with your knowledge manager

We need from you

  • Page owners who will accept tasks, and a decision on what happens to pages that have no owner
  • One month of questions and tickets, and access to the sources you want indexed with permissions intact
  • A rule for retiring superseded pages, because an index of contradictions answers confidently and wrongly

Stages

Sizing

Classify a month of questions and measure what the documentation actually answers

Design

Index scope, agent behaviour, refusal and escalation rules, the owner and task model

Pilot

Two domains, typically IT and HR policy, for one business unit, with the service desk watching the log

Scale and run

The remaining sources and domains, then the weekly gap cycle with the knowledge manager

Departmental. The effort sits in the content rather than the agent: how many sources, how much they contradict each other, how many pages have no owner, and how many languages the answers need.

Three answers in ten minutes, and the page that should have settled it is two years old.

Export one month of "how do I" questions from your service desk and busiest Teams channels. You get back the share your documentation could answer today, the twenty pages that would close the biggest gaps, and a written assessment of what the loop needs from your owners.

Send us one month of questions

The neighbouring process usually has the same problem

Industries we deliver this in most oftenManufacturing & industryTransport & logisticsServices & ITShared services

Browse all 173 solutions