Copilot
Copilot Studio
Agents built over an organization's own knowledge and processes, published into Teams, a website or a phone line.
Public Sector · Data, AI & Integration
Forty thousand web pages, answered in seconds
A major metropolitan city government, Canada. Citizens looking for accurate information had to navigate roughly forty thousand web pages. Frontline staff absorbed the resulting calls, and the answers were not always consistent.
The situation
What the organization was dealing with
Citizens looking for accurate information had to navigate roughly forty thousand web pages. Frontline staff absorbed the resulting calls, and the answers were not always consistent.
A large municipal website is a genuine information retrieval problem. Forty thousand pages is more than any navigation structure can make tractable, and every failed search becomes a phone call that a person has to answer.
What was done
The work, and the part that was actually hard
An AI assistant built with Copilot Studio was deployed over the city's own published content, letting citizens ask questions in natural language and get accurate answers directly.
Grounding an assistant in the city's own published content is what keeps this defensible: the answers are traceable to something the city has already said publicly, which matters when the subject is entitlements or obligations.
Results
What was published
These are the figures exactly as reported in the source. Nothing has been rounded, extrapolated or restated.
Web pages covered40,000
To an accurate answerSeconds
Pressure on frontline staffReduced
What changed
- Citizens finding accurate information without navigating the site structure
- Answers grounded in the city's own published content
- Pressure eased on frontline staff handling routine enquiries
- Public service delivered faster without additional headcount
Platforms involved
What each product was doing here
What transfers
If you were to attempt this
This is the most directly copyable story in the set. Any organization with a large public content estate and a call centre absorbing navigation failures has the same opportunity.
Where it usually gets harder than expected: Accuracy depends on the underlying content being current. An assistant over a site with stale pages will surface stale answers faster and more confidently than the search box did.
How we would take it on
Our approach to Data, AI & Integration work
1
Agree the definitions first
The least technical and most important step. If two functions cannot agree what a customer is, no architecture will reconcile their reports.
2
Deliver one workload end to end
Ingestion through to a report somebody actually uses, before the second workload starts. Platform builds that produce nothing for nine months lose their sponsor by month six.
3
Prove lineage
Every published figure traceable back to its source transaction. In a regulated setting that is a requirement; everywhere else it is what makes the number trusted.
4
Watch the consumption cost
Capacity-based compute behaves differently from licensed software. Monitoring goes in with the first workload, not after the first invoice.
Recognise the problem?
If any of the above describes your organization, tell us where it hurts most. We will tell you what the same platforms could realistically do in your environment, what we would measure, and whether we think it is worth doing at all.
Talk to our teamMore Data, AI & Integration storiesWe reply to every message within one business day.
Related reports
Others you may want to read
Fifteen data sources into one lakehouse, and models that finishAlso in Data, AI & IntegrationForty siloed systems into one citizen recordAlso in Public SectorMunicipal finance modernized, and the savings reinvestedAlso in Public SectorFive hundred officer hours a month, returned to the streetAlso in Public Sector