What Salesforce’s own research says
| Finding | Figure |
|---|---|
| Leading LLM agents, single-turn success on realistic CRM tasks | About 58% |
| The same agents, multi-turn settings | About 35% |
| Inherent confidentiality awareness | Near zero |
| Data and analytics leaders saying their data strategy needs a complete overhaul for AI | 84% |
| Organisational data those leaders estimate is untrustworthy | Over a quarter (26%) |
| Leaders with AI in production who have seen inaccurate or misleading outputs | 89% |
Sources: Salesforce AI Research, CRMArena-Pro (arXiv, 24 May 2025); Salesforce, State of Data and Analytics — 3,800 data and IT leaders and 3,852 business leaders in 18 countries, surveyed June–August 2025 (published 4 November 2025). Checked 5 October 2026.
These are the platform vendor’s own findings, which is why they are worth reading. They describe the gap every agent project has to cross, whoever builds it.
Where Agentforce projects stall
| Where it stalls | What actually happened |
|---|---|
| Data | The demo used tidy sample records. Real accounts have duplicates, empty fields and history nobody cleaned |
| Multi-step work | One question answered well; a case that takes four steps and two conversations does not finish |
| Permissions | Nobody decided what the agent may see and change for each kind of user |
| Systems outside Salesforce | The answer lives in billing, the warehouse or a partner portal — each with its own owner and sign-in |
| Success | Nobody agreed what “working” means, so nobody can show it is |
None of these is solved by a better prompt.
A way forward: one workflow, one number
- Pick the workflow that justified the project — case triage, order status, renewal preparation. One, not five.
- Clean what that workflow reads. Not the whole org; the objects and fields this one workflow depends on.
- Decide what it may see and do. Per kind of user, with a person reviewing where the cost of a wrong answer is high.
- Agree the measure before you build. A fixed set of your own past cases, the result that counts as right, and the level that justifies go-live.
- Measure, then decide whether to widen, change approach or stop.
When the decision must be the same every time
Some steps — an entitlement check, a discount approval, a refund rule — need the same answer for the same input and a trace back to the rule that produced it. Those are better as compiled logic than as a live conversation with a model.
For that, Pricient quotes IOanyT’s Proof Sprint: in three to four weeks, one workflow is compiled — LLMs write and test the logic at build time, and deterministic code runs it, so the same input gives the same answer (what Compiled AI means). It ends with a consistency and accuracy report on your own data and a written go / no-go, you keep the code either way, and the price is fixed and agreed at scoping.
Who signs off
The team that builds the agent should not be the one that certifies it. Independent review exists for exactly that reason.
How Pricient handles a stalled Agentforce project
Describe what stalled and which workflow mattered most. A person reads it; this lane’s own intake is not live yet. You receive the specification and a gap list — data, permissions, integrations, success measure — before anyone quotes a sprint. We do not make you compliant; we help you find out, on your own records, whether the workflow works.