Software that uses AI, data and tools to carry out a task. It may have a chat interface or work in the background. Access depends on connection identities and permissions; an agent is not automatically a separate Dataverse user.
Agents
Agents that work inside your system – under your rules.
An AI agent can retrieve information, suggest responses and perform bounded tasks through connected services. We start with the work that needs doing and agree what can happen automatically and what needs a person's approval.
What an agent is, and is not
A magic fix that “repairs the CRM”. An agent on top of messy data produces wrong answers faster. The foundation comes first.
Start with what is available
We assess built-in Dynamics 365 and Microsoft 365 Copilot capabilities before building something new. Availability, language support and cost depend on the product, region, licences and release status. We distinguish generally available features from previews before making production decisions.
Sales meeting preparationAccount history and previous contact summarised for review
New prospectsSuggested qualification and follow-up against agreed criteria
Sales follow-upDraft next communications for sellers to assess and approve
Case overviewA summary of the information a service colleague needs to continue
Enquiry classificationSuggested categories and routing, with a path for ambiguous requests
Case recordingHelp with recording and updating cases within defined permissions
Knowledge sourcesSearches of approved sources and article drafts for their owners
Quality testingTesting Icelandic and English answers, sources and human handoff
Custom agents in Copilot Studio
When built-in capabilities are not enough, we can build in Copilot Studio and connect Dataverse or other systems through appropriate connectors, APIs or MCP, a standard for connecting AI to data and tools. We first assess whether Power Automate or a simpler change in working practices will do. Custom development is an option when the requirements call for it.
Governance
- A named owner and least-privilege access for each connection
- Review of data flows, service providers and approved knowledge sources
- Appropriate action and error logging configured and tested; coverage varies by service
- Running-cost estimates, alerts and supported usage limits
- Icelandic and English tests, including incorrect answers and attempts to exceed permissions
- Approval for consequential actions, human handoff and a way to stop execution
- Agreed follow-up covering quality, costs, outcomes and retesting after changes
Offer
One task. Test its value.
A bounded trial with agreed success criteria. We aim for 4–8 weeks once data, access and scope are established; timing and price are confirmed in a proposal. The outcome may be an agent in production, simpler automation or no change for now.