01
Agentforce
Agents that reason over your data and take action, with subagents for distinct topics. Formerly Einstein Copilot — the rename brought no functional change.
An AI answer is only as good as the data it stands on and the permissions it respects. We build both halves, then put agents where they remove real handling time.
Agentforce, Prompt Builder, Data 360 grounding and the Einstein Trust Layer, configured for your org rather than demonstrated on someone else's. We start by making your data answerable, because that is where AI programmes succeed or quietly fail.
Book a DemoThe names have moved more than once. What matters is which layer you are actually buying, and what it needs underneath it.
Salesforce AI is no longer a single product. It is a stack: Agentforce for agents that take action, Prompt Builder for grounded generative output, Einstein for the predictive scoring that still does quiet, useful work, Data 360 as the data foundation, and the Einstein Trust Layer governing what reaches a model and what comes back.
Most of the difficulty is not in the agent builder. It is in whether your data can answer the question, and whether the permission model behind it is one you would be comfortable defending.
Agents that reason over your data and take action, with subagents for distinct topics. Formerly Einstein Copilot — the rename brought no functional change.
Reusable prompt templates grounded in your records, so generative output cites your business rather than inventing a plausible version of it.
Scoring, forecasting and Prediction Builder. Unfashionable next to agents, still the cheapest reliable win in most orgs.
Unified records to ground on, and the masking, grounding and retention controls that make the whole thing defensible.
Six areas, each one a place where Salesforce AI programmes usually run into trouble.
Choosing the first use case properly: high volume, low risk, measurable on your own numbers. Agent topics and subagents modelled so the agent knows what it is not allowed to attempt.
Prompt templates grounded on the right objects and fields, tested against the awkward questions rather than the demo ones, and versioned so a change can be reviewed.
Unified customer, asset and order records, resolved duplicates, and zero-copy access to systems you already run. The part nobody demos and everybody needs.
Masking checked against the languages your users actually write in, grounding scoped to real permissions, and audit retention in Data 360 configured as a deliberate decision rather than a default.
Prediction Builder, scoring and forecasting where a model beats a rule — and an honest answer when a rule would do the job for a fraction of the cost.
Escalation paths to a human, handling for the cases the agent should refuse, and enablement so your admins can change a prompt without raising a ticket with us.
High-volume, repeatable interactions where an agent removes genuine handling time — and the data work that makes each one possible.
An agent that answers from the orchestration plan rather than a status code, explaining which step is outstanding and what happens next. Grounded on live order and fulfilment data, not a generic FAQ.
Why it works: highest contact volume, lowest risk, and the answer is verifiable.
Assembling charge history, usage and rate card into a plain explanation of why a bill looks the way it does, with resolution inside defined guardrails and a clean handoff when it exceeds them.
Why it works: the data is already there; it is the assembly that takes an agent twenty minutes.
Guided change journeys where the agent checks eligibility, proration and contract term before offering an option — so it never proposes something fulfilment cannot deliver.
Why it works: eligibility rules exist in the catalogue already and can be grounded on.
Agent support for complex B2B quotes: valid configurations, grouping sites by location and serviceability, and flagging a quote likely to fail downstream before it is sent.
Why it works: it reduces middle-office rework rather than just deflecting contacts.
On-site assistance that surfaces the asset history, the entitlement and the technically valid upgrade options, without the technician digging through three systems on a doorstep.
Why it works: the alternative is a phone call to someone at a desk.
Surfacing orders drifting toward a missed commitment while intervention is still possible, with the reason attached rather than just a flag.
Why it works: predictive scoring on data you already hold, no generative component required.
We maintain a working telecom lead-to-cash demo org. Ask for a walkthrough of any of these against your own data model.
The Einstein Trust Layer does real work. It also has edges worth knowing before you promise anything to a compliance team.
Grounding respects the asking user's permissions, which means an agent is a very fast way to discover that your sharing model was looser than you thought. Masking coverage varies by language and region, so it needs checking against the languages your customers actually write in. And zero data retention covers the model provider — the audit trail itself lands in Data 360, where retention is yours to govern.
None of this is a reason not to deploy. It is a reason to deploy with the configuration reviewed rather than assumed.
Book a DemoSalesforce has renamed much of this stack. Older guides and proposals often use names that no longer match what you will see in Setup.
| You may see | Where it stands now |
|---|---|
| Einstein Copilot | Renamed Agentforce. No functional change; permissions and documentation were updated to match. |
| Einstein GPT | Branding largely absorbed. Generative capability now sits under Agentforce and Prompt Builder. |
| Data Cloud | Renamed Data 360. |
| Topics (in agent config) | Now referred to as subagents. |
| Einstein Analytics / Tableau CRM | Renamed CRM Analytics. Still supported, with Tableau Next positioned as the forward path. |
| Einstein Bots | Not retired, but in maintenance. It shares no authoring objects with Agentforce, so moving across is a rebuild rather than a migration. |
| Einstein Prediction Builder | Still current, still useful, still the right answer for some problems. |
Worth checking any AI proposal you have been given against this list — the naming is a reasonable proxy for how current the thinking is.
Four phases, each ending in something you can review. No phase runs blind into the next.
Data readiness, permission model review, candidate use cases scored on volume and risk, and an honest view of which ones are ready now versus after remediation.
Agent topics and refusal cases, grounding and prompt design, Trust Layer configuration, escalation paths, and the measures you will judge it by.
Configured agents and prompts, grounding tested against awkward real questions, permission testing per persona, and a controlled release rather than a full switch-on.
Containment, handling time and escalation quality measured on your own numbers, then the next use case — once the first one has earned it.
A first agent can ship in weeks once the data foundation is in place. When it is not, that foundation is the timeline, and we would rather say so at assessment than discover it in build.
Because that is where agent projects actually fail. If your records cannot answer the question, no prompt will rescue it.
An agent surfaces whatever its user could already reach. We check that deliberately rather than discovering it in production.
Order status, billing explanation, eligibility and quoting designed against a working lead-to-cash model.
Some problems are a validation rule, a report or a process fix. Saying so costs us a line item and saves you a programme.
150+ implementations across Sales, Service, CPQ and Industries, which is the context most AI work depends on.
Prompt designs, grounding maps, permission matrices and test scenarios come with the build, not as a change request.
Agentforce is the agentic layer: agents that reason over your data and take action, built on the Atlas Reasoning Engine. Einstein now refers mainly to the predictive feature set — scoring, forecasting, Prediction Builder — plus the Einstein Trust Layer. Einstein Copilot was renamed Agentforce with no functional change.
Usually not entirely, and that is normal. Agents need unified customer identity, resolved duplicates and a sharing model someone has deliberately reviewed, because grounding returns only what the asking user is already permitted to see. Our readiness assessment tells you what must be fixed first and what can wait.
The Einstein Trust Layer masks sensitive data before prompts leave, grounds responses on your records within the user's permissions, and Salesforce enforces zero data retention with model providers. Two things need your decision rather than a default: masking coverage for the languages you operate in, and how long audit data is retained in Data 360.
It depends on the problem. Predictive scoring on data you already hold is often faster, cheaper and lower risk than a generative agent, and it is frequently the better first move. We will tell you which one fits rather than defaulting to whichever is more fashionable.
Not directly. Einstein Bots is not retired but is in maintenance, and it shares no authoring objects with Agentforce, so moving across is a rebuild rather than a migration. The upside is that it is a chance to redesign flows that were shaped by the old tool's constraints.
A first narrow agent can ship in weeks once the data foundation is in place. Where it is not, the foundation sets the timeline. Our assessment gives you a defensible estimate in three to four weeks rather than a number chosen to sound attractive.
On your numbers, agreed before build: containment rate, handling time, escalation quality and resolution accuracy for the specific use case. We would rather define a narrow measurable win than claim a broad unmeasurable one.
Yes, and it is often the better arrangement. We will take a workstream, augment your admins and developers, or run the programme and hand it over — with the documentation needed to make the handover real.
Tell us the interaction that costs you the most handling time, and we will tell you honestly whether an agent, a prediction or a process fix is the right answer — and what your data needs before any of them will work.
Book a Demo
Chatbot