Operating & Building on ERP/CRM Copilots

The platform wave is the strongest evidence in this whole category, and it landed on the people who administer ERP and CRM systems, not on ML teams. SAP has Joule live across 35 solutions (SAP's count) plus Joule Studio on BTP; Salesforce shipped Agentforce with Agent Builder and the Einstein Trust Layer; Dynamics 365's 2026 wave brought autonomous agents and shipped MCP servers for Sales and ERP; Oracle claims 600+ pre-built Fusion agents with a free Agent Studio. All vendor counts — flagged as vendor counts — but the direction is unambiguous: the platform specialist's job now includes operating AI, verifying AI, and building on AI. The pay data follows: Agentforce, Data Cloud and prompt-engineering skills carry a measured $15–20K premium on mid-tier US Salesforce admin salaries, and that premium translates into Indian consultant billing rates as client agent projects ramp.

The scenario — the answer that missed the Z-table

Meghana is an SAP techno-functional consultant (SD module, eleven years) at an India services firm, staffed on a manufacturing client's S/4HANA estate. The client enables Joule, and a sales-ops manager starts asking it order questions directly.

That is the platform premium in one story. Any consultant can enable a copilot; the one who can measure it, scope it, and extend it is the one clients keep.

Operator skill — working the copilots properly

Embedded copilots reward users who know what they are for. Three prompt modes recur across every platform, with different trust profiles:

• Navigational"take me to the app for supplier invoices under review." Lowest risk: the destination is on screen, wrong answers are self-evident. Use freely; teach freely.
• Transactional"create a quotation for this customer from last month's order." Medium risk: the copilot drafts, the document is reviewable before posting — if users know to review. The operator rule: AI drafts the transaction, the human owns the posting, always.
• Analytical"what is this customer's overdue balance?" Highest risk: a figure with no visible working, exactly where the Z-table problem lives. Analytical answers get verification habits — and for anything feeding a decision, a check against the report or table that is the system of record.

Per-solution scope knowledge is part of the skill: Joule's capabilities differ by SAP solution and release; D365 Copilot features are enabled app by app; Fusion agents cover specific flows. "What can it actually do here, this quarter?" is a question the platform person should answer from the release notes, not from vendor keynotes.

The verification protocol — the skill no vendor teaches

Meghana's spot-check generalizes into a protocol you can run on any platform copilot, and it is deliberately boring statistics rather than vibes:

1. 1Sample. 15–30 real questions across the answer types users actually ask, weighted toward business-data answers (the risky kind), including questions that touch known customizations.
2. 2Verify. Each answer checked against direct queries — SQL/CDS views on the SAP side, SOQL on Salesforce, direct Dataverse queries for D365 — written by someone who knows where truth lives in this org's schema, custom objects included.
3. 3Classify. Every miss gets an error class: stale data (grounding lag), wrong scope (the Z-table class — customization invisible to the copilot), wrong aggregation (right rows, wrong math), or invented content (the model filled a gap). The classes matter because the fixes differ — grounding config, extension work, prompt constraints, and rollout limits respectively.
4. 4Gate. Rollout widens only on measured accuracy per question type: green types get enablement and training, amber types get "verify before use" guidance, red types get blocked from the use-case catalog until fixed.

Rollout itself is a change program, not a feature flag: a use-case catalog per role (what to ask, what not to ask), a prompt library tuned to the org's data and vocabulary, adoption metrics beyond login counts (answers acted on, deflected tickets), a feedback loop that routes wrong answers into the verification queue — and config documentation that keeps pace, because copilots ground on metadata and customization, so every transport or config change can silently move copilot behavior. The rollout kit is standard consulting muscle applied to a new object, which is exactly why platform people, not data scientists, end up owning it.

Building custom agents — the four platforms

Operating someone else's copilot is table stakes; building agents on the platform is the premium skill. The builder surfaces differ; the discipline is identical.

The anatomy transfers across all four: an agent is topics (what it handles), instructions (natural-language behavior constraints — these are code now, version them like code), actions (the platform automations it may invoke — the dangerous part), and grounding (what data it may see). Two design rules carry the weight:

• Actions are the risk surface, not the modela wrong sentence embarrasses; a wrong action posts, updates or emails. Scope actions minimally, make state-changing actions require approval (Module 8 formalizes this), and reuse the platform's permission model so the agent can do nothing its runtime user could not.
• Instructions are guardrails, not guarantees"only discuss order status" reduces off-topic behavior; it does not prevent it. Anything that must hold is enforced in action scoping and permissions, never in prose. The same lesson as Module 5's RLS rule, one layer up.

Prompt templates deserve first-class treatment on every platform (Salesforce Prompt Builder, Copilot Studio topics, SAP prompt templates in AI Foundation): record-grounded inputs, an output-format contract, and attached test cases, versioned and reviewed like any other platform asset. A prompt template with no test cases is a config change nobody can regression-test.

Agent testing — evals, or it does not ship

Agentforce Testing Center productizes this on one platform; the discipline is portable to all four with a spreadsheet and a scripting hour. The eval mindset — the same one that gated NL-to-SQL rollout in Module 5 — is the single strongest transfer skill of this track: platform agents are software; test them like software.

The India-services reality

For the services consultant, all of this runs inside someone else's tenant under an MSA — which changes the rules:

• Client tenant, client rulesAI features are enabled by the client's licenses and policies. You operate Joule or Agentforce as configured by their org; recommending is your job, unilaterally enabling is a contract violation.
• The approved-channel hierarchyfor any AI assistance involving client material: the client's own copilot licenses first, then a client-approved gateway, then synthetic stand-in data, then abstain. ABAP and Apex under client IP, master data, and config exports never enter a public AI tool. One paste can breach an MSA.
• Zero-retention literacy is client-facing skillTrust-Layer-class commitments (masking, no training on client data, retention terms) are things consultants must read and explain, because the client's DPO will ask, and "the vendor says it's fine" is not an answer. Module 8 goes deeper on DPDP and sub-processor terms.
• The premium concentrates where verification livesagent-building certifications signal; a portfolio artifact — a working agent, its eval report, a copilot verification memo like Meghana's — closes. Services rates follow demonstrated judgment, and verification artifacts are judgment made visible.

Try it now

1. 1Run a five-answer spot-check. On any copilot you can access — even Copilot in Excel against a known workbook — ask five business-data questions, verify each against the source, and classify any miss: stale, scope, aggregation, invented.
2. 2Draft one agent spec on paper. For a real process you support: topics, instructions, the minimal action list with each action marked read or state-changing, grounding sources, and the runtime user's permissions. Twenty minutes; this is the design skill.
3. 3Write a 15-utterance eval set. For that same agent: ten in-scope phrasings, three off-topic traps, two adversarial attempts — with the expected topic and action for each. You now have the artifact most production agent teams still lack.

Check yourself: A client wants yesterday's Joule pilot expanded to 200 sales users next week, "since the demo went well." What do you put in front of them before agreeing, and what does each artifact prevent?

This module anchors LAB-13 (copilot answer verification and month-end close support) — fifteen seeded copilot answers audited with SQL, error types classified, go/no-go decided. Do it at work this week → pb-37 (working Joule properly), pb-38 (the spot-check protocol), pb-39/pb-40 (Agentforce supervised operation and your first agent), pb-41 (utterance-set testing), pb-42 (Copilot Studio over ERP MCP with approvals), pb-43 (grounded prompt templates), and pb-44 (the rollout kit).