Two developments in 2026 made org readiness a harder problem than it used to be. Salesforce shipped Headless 360 in April, which lets external agents reach the platform through APIs and MCP tools without navigating a browser. In August, Claudeforce put Salesforce capability directly inside Claude with 37 prebuilt sales skills.
Both inherit the permissions of the user they act for. Salesforce's own documentation puts it plainly: when an agent acts on behalf of a user, it can do anything that user is permitted to do. That turns permission accuracy from housekeeping into a security control.
Pillar 1: Data quality
Agents cannot tell current data from stale data. They will read your pipeline as it stands and produce a confident answer built on whatever is there.
Run these five reports before you enable anything:
| Check | Acceptable level | Impact if higher |
|---|---|---|
| Duplicate account and contact rate | Under 2% | Account planning, customer health, stakeholder mapping |
| Open opportunities with past close dates | Under 5% | Forecast narratives become unreliable |
| Open opportunities with no 30-day activity | Under 15% | Deal health scoring, pipeline review |
| Opportunities with an empty next step field | Under 10% | Close planning, daily briefings |
| Contacts with no role assigned | Under 20% | Stakeholder mapping, meeting prep |
The fix is usually validation rules and process change, not a one-time cleanup. Without that, the same numbers come back within two quarters.
The field that lies
Plenty of orgs treat a field as required in process but leave it optional in the schema. When Competitor or Loss Reason sits 60% blank, an agent reading the populated 40% will report a selection-biased conclusion as an industry trend, and it will sound authoritative.
Audit every field the business treats as mandatory and make the schema enforce it.
Pillar 2: Permissions
This is where most orgs are exposed, and it is the item most often skipped under deadline pressure.
Audit these:
- Every View All or Modify All grant on standard objects, across profiles and permission sets
- Report export permissions granted to users
- Sharing rules that give broader access than the role hierarchy suggests
- Integration users with elevated permissions and no named owner
- Users still holding permissions from a role they no longer occupy
Most audits turn up at least one profile that got View All during the original implementation as a deadline fix and was never revisited. When an agent can run natural-language queries across every accessible record without the user opening Salesforce, that old shortcut becomes a live exposure.
Remediation order:
- Inventory the over-provisioned profiles
- Check login and report-export history to see what access is actually used
- Move the legitimate needs into narrowly scoped permission sets
- Remove the blanket grants
- Retest the affected workflows before enabling agents
Budget three to four weeks for this at a mid-sized org. It is the single item that turns an eight-week pilot into a full quarter.
Pillar 3: Business logic
When work starts in Claude or Slack and lands in Salesforce, the page layout stops being your control surface.
Anything you enforced through the interface, conditional field visibility, user-restricted buttons, required fields per record type, has to be re-implemented in validation rules, flows and sharing models that an agent will actually hit.
- Document every rule enforced only by a page layout or Lightning component
- Confirm each one has a validation rule or flow equivalent
- Flag flows that need a human: screens, confirmations, waits for input
- Test automations that fire on field update for loop behaviour under rapid writes
Automation chains built for human-paced data entry behave differently when something writes 40 records a minute. Governor limits will catch some of it, but hitting them mid-workflow leaves partial states.
Pillar 4: Process
Two questions decide whether a pilot produces a repeatable result or a one-off anecdote.
Which single workflow are you piloting? Pick one skill with a measurable time cost, not broad platform enablement. Meeting prep and pipeline review are the strongest candidates because the hours are easy to count.
How will you measure it? Set the baseline before you switch anything on:
- Weekly minutes on the task for five reps across one baseline week
- The same measurement after four weeks of agent use
- An accuracy check by a sales manager across twenty sample outputs
- Confirmation that reps are not quietly redoing the work by hand alongside the agent, which is the failure nobody reports
Pillar 5: Measurement and governance
Adoption metrics change meaning. Once sellers start their work in Claude or Slack, Salesforce login counts stop describing usage. Agree the replacement metric before Q1.
Training shifts. The skill is prompting and output verification, not navigation.
Consumption needs an owner. Agentforce billing runs on Flex Credits that expire at order end with no rollover. Someone has to watch the monthly burn.
If you are in a regulated industry
Get the data flow diagram for each deployment path. A single request can cross the model provider, Amazon Bedrock and Salesforce infrastructure. Establish where data rests and where it is processed at each hop. Salesforce's Trust Boundary covers this for Agentforce by running inference inside its own VPC, but external plugin paths can differ.
Confirm logging and retention, then compare the answers against your own retention policy rather than assuming they line up.
How long it actually takes
| Phase | Duration | Owner |
|---|---|---|
| Data quality assessment | 3 to 5 days | Admin or analyst, reporting only |
| Permission audit | 1 week to assess, 3 to 4 weeks to remediate | Admin plus security review |
| Business logic review | 1 to 2 weeks | Developer or architect |
| Pilot design and baseline | 1 week | RevOps |
| Pilot run | 4 weeks | One team |
Eight to ten weeks for a typical mid-sized org. Orgs with a disciplined permission model can be ready in about five. Permission remediation is the longest phase and the one most often cut, and its failure mode is data exposure rather than a disappointing demo.
Originally published at ashapurasoftech.com.
Top comments (0)