Almanac
Microsoft/d365csDynamics 365

Consultant-focused KB for Microsoft Dynamics 365 Customer Service: implementation notes, gotchas, and configuration decisions beyond the official docs — across case management, queues and routing, knowledge base, SLAs, omnichannel, agent experience, Copilot, analytics, and administration.

feature-choose-deflection-or-headcount.mdv1 · history
CurrentApplies to AllUpdated 6 days agoSource Microsoft Learn

What it does

Every service business case eventually reduces to the same choice: spend the money on knowledge and self-service so fewer contacts reach a human, or spend it on humans. Dynamics gives you the machinery for the first — knowledge articles with versioning and translation, portal search, external search providers, Copilot answers in the agent desktop, and Copilot Studio agents answering customers directly. What it does not give you is the editorial function that makes any of it work.

Key facts

  • Knowledge articles carry a lifecycle: draft, review and approval, scheduled publish, expiry. Expiry is the setting most organisations skip and most need.
  • Major and minor versioning is supported per article, which matters when a product changes and you need the old answer to stay reachable for customers on the old version.
  • Translations are per-article records against a default authoring language set at system level. Translation is a workflow you have to staff, not a toggle.
  • Search can federate beyond Dataverse. External search providers can index enterprise websites using the site map protocol, and results are ingested into Dataverse alongside native articles.
  • The knowledge analytics dashboard reports article usage, views, shares, and — the useful part — searches with low success rates. That report is the closest thing you get to a content backlog generated for you.
  • Copilot Studio agents can be pointed at the same knowledge for customer-facing self-service, so one well-maintained article can serve the portal, the bot and the agent's Copilot pane.
  • Knowledge exists in both Professional and Enterprise, but the surrounding capability — Copilot answers, external providers, analytics — is where the editions separate.

When to use / skip

The recommendation is unambiguous: do knowledge first, always, and do it before you buy a bot. Not because deflection beats headcount in every case, but because knowledge is the prerequisite for every other option. Copilot answers are only as good as the articles behind them. A Copilot Studio agent with no knowledge base is a very expensive way to say "I'm sorry, I didn't understand". Adding agents to a team with no knowledge base means every new starter takes four months to become useful instead of six weeks. Knowledge is the cheapest lever in the building and it is the one clients consistently defer.

Where deflection genuinely does not pay is when contact volume is dominated by long-tail, judgment-heavy cases. Complex claims, safeguarding, regulated advice, B2B technical escalations — if the top ten contact reasons account for a quarter of volume rather than two thirds, there is nothing to deflect and you should hire. Check this before promising anything. Pull the last twelve months of cases, group by subject, and look at the concentration. High concentration means deflection; a flat distribution means people.

Also skip the deflection story when nobody will own the content. This is the real determinant and it has nothing to do with technology. If there is no named knowledge manager with time in their week, articles go stale within two quarters, search success rates fall, agents stop trusting the results, and the whole investment quietly becomes a folder of out-of-date documents. I would rather deliver a client three hundred well-maintained articles than three thousand nobody has touched since migration.

The opposite call — hire first, deflect later — is right when the service is in crisis. Queues at breaking point, SLA failures visible to the board, staff leaving. You cannot build a knowledge programme in a burning building. Stabilise with contractors, then do the knowledge work from a calm base three months later. Saying this to a client who wants an AI answer is uncomfortable and it is usually correct.

Configuration decisions

  • Who owns knowledge editorially, how much of their week it takes, and whether authoring rights sit with a small central team or with senior agents. Both models work; the undecided model does not.
  • Article expiry policy and review cadence, configured rather than intended. Set expiry dates at creation or nobody will ever set them.
  • Which languages are in scope and how translation is resourced — human, machine-assisted, or not at all for phase one.
  • Whether external search providers are used to index existing intranet or product content instead of migrating it into Dataverse. Often the faster route to coverage.
  • Whether Copilot Studio self-service is in phase one or deferred until the article set has been proven with agents first.

Gotchas

  • Migrating an existing SharePoint or Confluence knowledge base wholesale is the classic mistake. You inherit years of stale content and the search results are immediately untrustworthy. Migrate the top articles by usage and leave the rest behind.
  • Deflection metrics are easy to fake. Portal page views are not deflection. Measure contacts per customer per period against a baseline, or you are reporting on nothing.
  • Knowledge search results depend on how the articles are written, not how the search is configured. Long prose articles with buried answers score badly against Copilot and against customers. Rewriting for the search behaviour is real work.
  • Article expiry silently removes content from search when it fires. If nobody is watching the expiring-soon view, coverage decays without any alert.
  • Publishing internal articles externally without a review pass leaks internal process language, pricing detail and occasionally staff names. Put a review gate on the external channel specifically.

Consultant notes

  • Ask for last year's case volume grouped by subject before the first design workshop. The concentration curve settles this argument faster than any discussion, and it is a five-minute report.
  • Get the knowledge owner named in the project RAID log as a dependency, with the hours committed. If the client will not name one, record it as a risk and repeat it at every steering meeting.
  • Demo Copilot answering from the client's own three worst articles, not from sample data. It makes the content-quality argument for you in about ninety seconds.
  • Set a coverage target for phase one — the top twenty contact reasons, written and reviewed — rather than an article count. Article counts encourage volume and volume is the problem.
  • Before go-live, run the low-success-search report against pilot usage and fix the top ten gaps. That is the cheapest quality improvement available on the whole project.

Worth another look when Copilot's knowledge grounding changes, or if the client's contact mix shifts enough to redo the concentration analysis.

Was this accurate?