Home · Solutions · Management & planning
Solution · Management & planningOne question, one answer with its source, and a name when the documents are silent
Policy answers in seconds instead of three emails
Employees ask in Microsoft Teams; the agent answers only from documents they are allowed to open, cites the source, and hands the question to a named owner when nothing covers it.
Executive summary
The same policy questions reach the same three busy people, and the answer is already written somewhere.
The work starts with the content, not with the agent.
An answer that used to take hours arrives in seconds, including evenings and weekends when shift and field staff actually ask.
the request forms and service-desk tool; SharePoint owner and review-date columns; the Microsoft Purview audit record
Business problem
Enterprise knowledge
Every company writes its rules down. Travel and expense policy, purchasing thresholds, supplier onboarding, IT device standards, the specification a salesperson may quote. Over a decade this becomes several hundred documents on SharePoint, a wiki nobody maintains, a shared drive that survived the last migration, and decisions taken in meetings and never written down.
The people who feel it first are the ones who know the answers. Three or four experts carry the rules in their heads: the head of procurement, the travel and expense owner in finance, an IT policy manager, a product specialist. Their calendars are full, so questions queue behind them. Feeling it next are new joiners and field staff, who cannot tell a current document from a superseded one.
At scale the failure has a shape. Search returns forty results and ranks the old ones first, because they have been opened more often. Two departments answer the same question differently and both believe they are right. An acquired site keeps its own rules for years because nobody merged them.
How it works today
One unanswered question takes this route before anything is automated.
- PersonAn employee searches SharePoint, gets forty results, opens three and gives up
- PersonThey ask in a Teams channel instead, or email whoever answered last time
- WaitingThe question waits until that person is out of meetings, typically hours, sometimes two days
- PersonThe expert finds the current document if they can, otherwise reconstructs the rule from memory
- Risk of errorThe answer stays in a chat thread, so the next person to ask gets a slightly different version
- PersonSomebody updates the document eventually, or nobody does, and the intranet keeps the old figure
- Risk of errorShift and field staff, who rarely open the intranet, act on what their supervisor remembers
Why the current process costs more than it appears
The budget shows headcount, not what it is spent on.
- Two clocks run on every question: the expert spends minutes searching and phrasing an answer, and the person who asked spends hours not doing the work that depends on it.
- Wrong answers are paid for later: a purchase split in two to stay under a threshold that changed last year, a trip refused at expense approval, a superseded specification quoted to a customer.
- Onboarding is slower than it looks. A new manager spends the first quarter learning who to ask rather than what the rules are, and the reply becomes a private copy of the policy.
- Management has no data on any of it. Nobody can say which twenty questions absorb most expert time, so nobody can decide which rule to rewrite or form to simplify.
Cost of inaction
Nobody escalates a slow tax, which is why this one has lasted. The same three people answer the same questions, the intranet drifts further from what is true, and every acquisition adds rules only its own staff understand. Document maintenance stays unfunded because nobody can show what it is worth.
Two risks compound. A wrong answer is usually discovered only after money has moved: a threshold breached, a trip refused, a specification corrected in front of a customer. The second arrives with the first broad AI assistant rollout, which reads the same estate employees do and cites the superseded file as confidently as the current one.
A plausible organisation with realistic proportions. The figures are there to be recalculated on your data; they are not a client result.
A European manufacturing and services group, 3,500 office and field employees in five countries, Microsoft 365 with Microsoft Teams in daily use, around 400 policy and how-to documents on SharePoint and an intranet three people update between other duties.
About 6,000 policy and how-to questions a month are answered by people: travel and expense rules, purchasing thresholds, IT device and access rules, product specifications, safety procedures. A quarter arrive from shift and field staff outside office hours.
Questions arrive as Teams messages, emails and corridor conversations. Experts answer from the current document when they can find it, otherwise from memory, and the answers stay in chat threads.
Twelve minutes of combined working time per question, counting the expert's search and reply plus the asker's wait and repeat. Three people carry most of the load, so they constrain everything else they own.
A curated knowledge set on SharePoint with a named owner and a review date on every document, and an agent in Microsoft Teams that answers only from that set, only from what the person asking may open, and always with a citation. What the documents do not cover goes to the owner, and a robot raises the request form.
In the modelled case, questions the documents already answer are settled in seconds at any hour, expert time moves to real exceptions and rule changes, and a monthly gap report shows which documents to fix first. These figures are a model, not a client result.
Proposed solution
The work starts with the content, not with the agent. We inventory the documents that are supposed to answer questions, give each a named owner and a review date as SharePoint columns, retire duplicates and superseded versions, and agree what stays out: drafts, board material, anything under a sensitivity label that encrypts the file. An agent grounded on an unmanaged library repeats that library's contradictions, faster and with more confidence.
Option one runs entirely on Microsoft 365. We build a Microsoft Copilot Studio agent whose knowledge is the curated SharePoint sites and publish it to Microsoft Teams and Microsoft 365 Copilot. Retrieval is trimmed to the permissions of the person asking, and sensitivity labels are respected the same way. Ungrounded responses are switched off, so the agent answers only when it can attach an in-text citation; otherwise it says so and offers the owner. Users with a Microsoft 365 Copilot licence are zero-rated in Teams and Copilot Chat, everyone else is metered in Copilot Credits, so the licence mix is a design decision.
Option two suits a company whose automation platform is already UiPath. We build a UiPath conversational agent with Context Grounding indexes over the same SharePoint libraries, deploy it to Microsoft Teams as a Teams app, and route what the documents do not answer to UiPath Action Center as an escalation with an owner and a response time. Indexes inherit the permissions of their source, guardrails for prompt injection and PII come from the AI Trust Layer, and the agent can call an existing automation as a tool.
Both do the same job for the employee: an answer with a source, a name when there is no answer, a next step when the question was really a request. Where a question ends in a form or a ticket, the agent offers it and a robot creates the record. Licensing and where your follow-up actions already live usually settle the choice.
Microsoft Copilot Studio agents with SharePoint knowledge, permission-trimmed retrieval, in-text citations and publication to Microsoft Teams and Microsoft 365 Copilot; SharePoint columns and versioning; Microsoft Purview labels and audit; SharePoint Advanced Management reports; or UiPath conversational agents in Teams with Context Grounding over SharePoint, Action Center escalations and AI Trust Layer guardrails
The curated knowledge set with its owner and review-date model, the agent instructions and answer style in each working language, the grounding, citation and refusal settings, escalation routes, the follow-up automations, the evaluation question bank and the monthly gap report
None on the knowledge path; follow-up actions write to your existing request or ticket system through UiPath Integration Service connectors or an agent flow
How the automated process works
- PersonAn employee asks in Microsoft Teams, in their own words, in Polish, English or German, at any hour
- AutomationThe agent retrieves from the curated knowledge set only, and only from documents that person may open
- AutomationThe answer carries an in-text citation to the source document, its owner and review date; what cannot be grounded is not sent
- SystemWhere the question is really a request, the agent offers the next step and a robot creates the form entry or ticket
- PersonWhere the documents do not answer, the question goes to the named owner with the original wording
- AutomationThe owner's answer is logged as a gap and the document that should have contained it is flagged to its owner
- AutomationA monthly report ranks the most asked questions, the documents actually cited, the documents nobody cites and recurring escalations
Human-in-the-loop model
Automation handles
- Answering the questions the curated documents already answer, with the source attached
- Enforcing who may see what, by trimming retrieval to the asker's permissions and to sensitivity labels
- Creating the simple follow-up record: a request form entry, a ticket, a booking on the right template
- Logging every question, citation and refusal for the monthly review
People decide
- Every question the documents do not cover, answered by the owner and written into the document
- What belongs in the knowledge set, who owns each document and when it is reviewed
- Rule changes and interpretations, which stay with the function that owns the rule
- Whether a repeated question means the document is wrong or the process is wrong
Before and after
Systems and integrations
Every entry can be checked in vendor documentation. The evidence class is stated next to each one.
Inputs
- curated SharePoint libraries
- intranet policy pages
- the product specification library
- a decision log for rules not yet written up
Automation layer
- Microsoft Copilot Studio or UiPath Agents with Context Grounding
- Microsoft Entra ID
- UiPath Orchestrator and Robots
Target systems
- the request forms and service-desk tool
- SharePoint owner and review-date columns
- the Microsoft Purview audit record
Human touchpoints: the agent in Microsoft Teams and Microsoft 365 Copilot; escalation tasks for named owners; the monthly gap review
Technologies used
the agent: SharePoint knowledge, permission-trimmed retrieval, in-text citations, ungrounded answers blocked
Awhere employees ask, and where escalations and follow-ups arrive
Athe curated knowledge set: libraries, owner and review-date columns, versions
Azero-rated use for licensed staff; Restricted Content Discovery, access reports and reviews
Asensitivity labels, audit of agent prompts and responses, data security posture for AI
Athe alternative build: permission-inherited indexes over the same libraries, as a Teams app
Athe follow-up record, and the escalation task with owner and response time
Aidentity, group-based access to the knowledge set, a distinct agent identity to review
AIllustrative economic model
The arithmetic is open, so it can be argued with.
Assumptions, plainly: no client was measured, and the three inputs below are the whole model. Twelve minutes is the combined working time per question: the expert's search and reply plus the asker's wait and repeat, averaged across trivial and hard cases. Thirty euro is a fully loaded hourly cost for office and specialist roles in Central Europe. The result is the size of the pool, not a promised saving; how much the agent answers without a person is what the pilot measures.
Run the numbers on your data
An illustrative estimate from your own inputs. It models released capacity; it is not a promise of savings.
Business benefits
- An answer that used to take hours arrives in seconds, including evenings and weekends when shift and field staff actually ask
- Experts keep the questions that need judgement and stop being the routing layer for everything else
- Every answer carries its source, so two people asking on different days get the same rule and the same version
- Questions that are really requests end in a created record rather than another email to the wrong inbox
- The company finally knows what it is asked, which turns document maintenance into a ranked list
The management view
- Knowledge stops being personal property: ownership, review dates and usage become visible properties of a document
- Risk of the wrong answer drops where it costs money, because thresholds and specifications come from the current version
- The monthly gap report ranks which rules are unclear, which forms are hard and which training is missing
- The pattern extends site by site and language by language without new headcount, and an acquired company joins one knowledge set
Board-level KPIs
Security and governance
Control is not an add-on.
- Retrieval is trimmed to the person asking, so the agent cannot show a document the employee could not open in SharePoint. Content under an encrypting sensitivity label is excluded from grounding entirely, a constraint to design around rather than discover later.
- The knowledge set is a deliberate list of sites, not the whole tenant. SharePoint Advanced Management reports show which sites are over-shared or inactive and who still has access.
- Prompts, responses and agent activity are auditable in Microsoft Purview, and the agent has its own identity in Microsoft Entra ID, so it is reviewed like a service account.
- Neither path leaves the EU. On the Microsoft path, content never leaves your tenant. On the UiPath path, indexes inherit SharePoint source permissions, guardrails cover prompt injection and PII, and processing stays in UiPath Automation Cloud's EU region.
- Escalations, follow-ups and rule changes are logged with who asked, who answered and which document was cited, which is what an auditor asks when a threshold is disputed later.
Why now
The trigger for most organisations is the assistant they are already piloting. Broad AI search reads the same estate employees do and quotes a superseded travel policy as confidently as the current one, so the content work happens either way.
The capabilities are documented and generally available: SharePoint knowledge with permission trimming, a setting that blocks any answer without an in-text citation, one publication to Teams and Microsoft 365 Copilot, and UiPath conversational agents as a Teams app since April 2026.
The pool is large enough to fund the work quickly: about €36,000 a month of working time goes into answering questions the documents already answer, against a build whose main cost is content triage you need anyway.
Relevant executive roles
The rules the operation runs on stop depending on who happens to be free to explain them
One governed agent over a curated, permission-aware knowledge set instead of departmental assistants on uncontrolled folders
New joiners and field staff get the same answer as head office, at the hour they actually ask
Thresholds and travel rules are quoted from the current version, and the time spent explaining them becomes measurable
Common questions and objections
Ungrounded responses are switched off, so the agent replies only when it can cite a document from the curated set, and otherwise passes the question to the owner. Before go-live it is scored against a question bank from real history, counting grounded, wrong and refused answers separately.
Then the agent would too, which is why content triage comes first. Contradictions are resolved or one version retired, every document gets an owner and a review date, and the reply shows which document it came from.
Microsoft 365 Copilot searches everything a person can see, drafts and superseded files included, and is licensed per user. This agent is scoped to a curated set, refuses to answer outside it, escalates to a named owner, can trigger a robot, and reaches employees without a Copilot licence.
When this is not the right solution
- The three people who know the rules sit in the same room, or questions stay below a few hundred a month
- The rules are not written down anywhere. An agent cannot cite a document that does not exist, so the first project is writing the content
- The question needs data rather than a document: a leave balance, a ticket status, an order. That is the self-service agent pattern reading HR or service-desk data, covered separately
A question for the next management meeting
What does it cost us, in working days a month, to answer questions whose answers are already written down, and who owns those documents today?
Implementation approach
Delivery runs in stages, so it can be stopped at any point.
We deliver
- Content triage: owners, review dates, duplicates retired, exclusions agreed
- The agent itself, in Microsoft Copilot Studio or as a UiPath conversational agent, published into Microsoft Teams
- Grounding, citation and refusal settings, agent instructions and answer style in each working language
- Escalation routes to named owners, and the first follow-up automations that create a form entry or a ticket
- An evaluation question bank from real history, scored for grounded, wrong and refused answers before go-live
- The monthly gap report and a runbook for whoever owns the knowledge set afterwards
We need from you
- The twenty to fifty questions your experts answer most often, with their current answers
- Access to the SharePoint sites that should be in scope, and a decision on what stays out
- A named owner for each document family: finance, procurement, IT, health and safety, product
- A sponsor who can retire a contradictory document instead of keeping both
Stages
Content triage
Inventory the documents, name an owner and review date for each, retire contradictions
Design
Knowledge scope, permission model, answer style and languages, escalation routes, follow-up actions
Build
The agent, its grounding and citation settings, publication into Teams, the first follow-up robot
Evaluation
Scoring against the question bank, counting grounded, wrong and refused answers separately
Go-live
Two departments first with a weekly gap review, then the rest of the group and further languages
Departmental. Effort is driven by the state of the content rather than the technology: how many documents are in scope, how many owners must be found, how many contradictions resolved, how many languages the answers must work in.
6,000 questions a month, and the answers are already written down.
Send us the twenty questions your experts answer most often and the documents that should contain the answers. We come back with a knowledge-set scope, an owner map and the contradictions we found.
Send us your twenty most-asked questionsThe neighbouring process usually has the same problem
Half of what a project meeting agrees is never written anywhere another person can see it.
View solution IT & servicesProject workspaces provisioned with governance in minutesProject teams and sites built by hand, named differently every time, and never closed.
View solution HR & peopleJoiner, mover, leaver: access on day one, revoked same dayNew starters wait days for access; leavers keep theirs for weeks. Both are the same missing handover.
View solution Sales & marketingTender documents turned into a compliance matrixYour bid team spends its best days reading tender packs, not deciding which ones to win.
View solution Case studyThe HR agent on TeamsAn employee asks on Teams about leave, a certificate or a benefit.
View case study Case studyExecutive assistant 2.0 — Teams and Microsoft 365 CopilotOur executive finance assistant steps beyond Teams: with the bi-directional UiPath–Microsoft Copilot Studio integration it also answers in Outlook and Microsoft 365 Copilot.
View case studyIndustries we deliver this in most oftenManufacturing & industryRetail & e‑commerceServices & IT