Contoso Manufacturing: Out-of-the-Box Copilot vs. Custom Deal-Risk Scoring
Contoso Manufacturing (4,200 employees, EU-based) wants Dynamics 365 Sales Copilot to draft account summaries and follow-up emails for 350 sellers. Leadership has approved a nine-month programme and expects payback within two quarters. Discovery found that 38% of account records have no industry code, 1,100 duplicates exist across business units, and roughly 60% of meeting notes sit in a shared mailbox outside Dataverse. The works council requires that no seller performance data be processed by AI, and legal insists all inference stays in the EU. The sales operations lead is pushing for a custom agent in phase one that also scores deal risk from an on-premises SQL profitability table. The sponsor wants to know whether to buy 350 more Copilot licences now or wait for the pilot.
1. The sponsor asks whether custom deal-risk scoring should be funded in phase one. What must the team prove first, and what evidence would justify stopping?
Read the worked answer
Separate the two value streams before spending anything. Account summaries and email drafts are native Copilot capabilities grounded in Dataverse, so the pilot can test them with no custom build. Deal-risk scoring is a different workload: it depends on an on-premises SQL table that no Copilot capability reaches without a data gateway, a DLP-approved connector, and a governed sync into Dataverse. Even then, 38% missing industry codes and 1,100 duplicates make the risk features unreliable, so a model trained on today's data would encode the data quality problem.
Run a six-week pilot with 40 sellers on the out-of-box capabilities only, with pre-agreed gates: (1) groundedness of at least 85% on 200 human-reviewed drafts; (2) seller edit rate below 50% of generated text; (3) measurable cycle-time reduction against a baseline captured before launch; (4) written confirmation from legal that inference stays in the EU region; (5) a works council-approved field-exclusion list enforced by role and DLP, not by prompt wording.
Stopping evidence: if groundedness is below 85% or edit rates exceed 50%, the data is not ready and funding a custom risk model would compound the defect. Redirect the budget to dedupe and industry backfill. Also insist that deal-risk scoring be scoped as its own governed model with its own evaluation and explainability requirements, not smuggled in as a Copilot extension.
2. How should the plan be sequenced to give a credible two-quarter payback path without breaching the EU and works council constraints?
Read the worked answer
Quantify the quick win first so payback is measured, not asserted. If 350 sellers produce three summaries and three follow-up emails a week and each takes 12 minutes manually, that is 1,050 artefacts a week and 210 hours a week of drafting effort. Even a conservative 30% time reduction is a defensible return, and it needs no custom model, no on-premises data, and no new data flow.
Sequence: quarter one, deploy out-of-box capabilities to a 40-seller cohort, instrument usage telemetry, and publish the baseline-to-result comparison. In parallel, remediate the data that blocks everything else: merge duplicates, backfill industry codes, and move meeting notes into Dataverse with consent capture. Quarter two, extend to 350 seats only if the gates held, and only if the EU data boundary and works council exclusions are documented and tested.
Defer deal-risk scoring to quarter three, after the SQL table is synced through an approved gateway and the security model on profitability data is reviewed. Do not let the on-premises requirement pull the phase-one date; it is the single item most likely to trigger legal review and a works council challenge. Publish the exclusion list (no performance scoring, no individual ranking) as a design constraint in the backlog so makers cannot quietly build against it.