Here’s an invented scenario to frame the topic. A fictional mid-size property and casualty carrier, unconnected to any real company, wants an Agentforce agent to answer policyholder questions and take first notice of loss. Its Salesforce org is 9 years old, has been touched by 4 consulting firms, and holds 3 overlapping ways to record a household. Before anyone writes an agent instruction, the carrier’s architect asks what has to be true of the org, the data, the integrations and the compliance program for the agent to behave. This post turns that question into a checklist.

Scope and caveats come first. I read 14 pages. Salesforce supplied 7: setup documentation, a Trailhead unit, an admin blog, a pricing blog, an insurance skill page, a Trust Layer page and a service datasheet. The other 7 were 3 law-firm notes on the NAIC AI model bulletin, a vendor’s governance blog, a consultancy’s readiness article, a services company’s insurance blog and an individual blogger’s health check list. I set no pages aside, but several figures in them are unsourced and I say which. None of this is legal advice. Confirm every regulatory point with counsel and your state insurance department, and every product detail with Salesforce.

TL;DR

  • Check the foundation before the agent. Salesforce’s own setup page says Data 360 must be provisioned before generative AI features go on, and its Trailhead unit describes a Health Check score from 0 to 100 against Salesforce’s standard values. Run that, plus an org complexity review and a duplicate-record review, before you design any agent topics.
  • Regulation shapes the checklist for carriers. The NAIC model bulletin expects a written program for AI systems, senior management oversight, documented controls and diligence on third-party vendors. Law-firm notes say 24 states had adopted it by 2025, and a vendor says 25 by July 2026, which I couldn’t verify, so confirm your state’s position.
  • Test, price and gate before launch. Salesforce’s admin blog describes a 5-step test process, and its pricing blog gives $0.10 per action for Flex Credits. In my invented model, 3,600 agent-handled contacts a month at 5 actions each costs $1,800, so cost matters less than testing and approval gates.

The checklist at a glance

The table lists 12 checks in the order this post covers them. The evidence column says where each came from, and the owner column is my suggestion rather than anything a source prescribes. I grouped the checks into 4 areas: the org itself, the data and integrations, the platform prerequisites for the agent layer, and the carrier’s regulatory and cost controls. No source offered a complete health check for a carrier, so this list is my synthesis of Salesforce documentation, law-firm notes and partner articles.

#CheckWhat to look forEvidenceOwner (my suggestion)
1Security settingsHealth Check score and gaps against a baselineTrailhead unitSecurity lead
2Admin accessFew users with powerful permissionsBlogger’s listPlatform admin
3Org complexityObjects, fields, rules and sharing rulesBlogger’s listArchitect
4Duplicate recordsClean, deduplicated customer dataConsultancy articleData steward
5Policyholder recordAccounts, policies, coverage and claims linkedSalesforce skill pageBusiness owner
6Policy and claims integrationsCore systems reachable through APIsSalesforce datasheetIntegration lead
7Platform prerequisitesData 360, Einstein, Trust Layer settingsSalesforce setup pagePlatform admin
8Agent accessWho can use or change an agentConsultancy articleSecurity lead
9Regulatory programWritten AI program, inventory, oversightLaw-firm notesCompliance
10Vendor diligenceAudit rights and cooperation dutiesLaw-firm notesProcurement and legal
11TestingTest cases, human review, sandboxSalesforce admin blogQuality lead
12CostCredits, actions and licensesSalesforce pricing blogFinance

Why a health check comes first for a carrier

An agent is only as dependable as the org beneath it. A consultancy’s readiness article cites a Salesforce 2023 report, as quoted in the article, saying 86% of analytics and IT leaders agree that AI’s outputs are only as good as its underlying data. I couldn’t read that report, so the figure is secondhand. The same article says testing in a controlled setup helps ensure agents don’t unintentionally modify production data. For a carrier, that sentence carries extra weight, because the data in question may include coverage, claims and personal information.

Security settings and access

Run the built-in tool. A Salesforce Trailhead unit describes Health Check as a Setup dashboard that shows an org’s security settings on a single page. The score runs from 0 to 100, with 100 the most secure, and it measures how closely your values match Salesforce’s recommended values, which the unit calls standard values. You find it by typing Health Check in Setup’s Quick Find box. The unit doesn’t list every setting it covers, so ask your administrator for the full export and keep it as evidence.

Be careful with Fix Risks. The same unit says the Fix Risks button changes all settings to the recommended values at once, and it warns that this could affect integrations or remove some users’ access. It recommends changing settings individually and testing access changes in a sandbox first. For a carrier with policy and claims integrations, that advice is practical: a changed session or password setting can break an integration user at 2 a.m. Record each change, who approved it and what you tested.

Use a custom baseline. The unit says you can import your own baseline as an XML file and suggests this for highly regulated industries, after aligning with IT or compliance. Insurance qualifies. If your information security policy sets stricter values than Salesforce’s defaults, encode them so the score measures what your policy requires. It also names Security Center as a separate add-on for multiple orgs, which reports multi-factor authentication use and admin-level permissions.

Limit admin power and agent access. An individual blogger’s 20-step health check says to find profiles with powerful permissions such as Modify All Data and View Setup and limit them to a few users. A consultancy readiness article says to restrict agent access to qualified users, require multi-factor authentication and define clear permissions for each agent’s actions. I found no source that says how many admins is too many, so set your own number and record the reasoning.

Org complexity and technical debt, scored

A complicated org slows every change and hides risk. The table summarizes thresholds from an individual blogger’s 2022 health check post. They’re a single practitioner’s rules of thumb, not Salesforce guidance, and the limits he cites may have changed since the post was written, so confirm each limit in current Salesforce documentation before you use it. I include them because they turn a vague worry about a messy org into numbers you can count in an afternoon.

CheckThreshold in the postHow I’d use it (my suggestion)
Total custom objectsCites a 3,000 limit; over 1,500 hints at a complex orgCount them and list the owners
Fields per objectObjects using 80% of field capacityReview the busiest objects first
Sharing rulesLimit of 300 per object; simplify beyond 150Check policy and claim objects
Workflow and approval rulesLimit of 500 per objectList the active ones
Inactive rulesClean up inactive validation, workflow and process rulesDelete or document them
Unused reportsNot run in the last yearArchive before launch
CustomizationsAbove 70% means a complex orgFlag it, then plan to simplify
Old automation and componentsMove older process tools to flows and older components to newer onesInventory what remains

Data quality and the policyholder record

An agent answers from whatever data it can reach. The carrier’s data should be clean, connected and defined before an agent touches it. The list below covers the data checks I’d run, drawing on the consultancy’s readiness article and Salesforce’s insurance skill page. Treat the list as a starting point, because no source gave a carrier-specific data standard or a pass mark for any of these checks. Add your own targets where a source gave none.

  • Duplicate and outdated records. The consultancy article says to remove outdated or duplicate records and to validate existing data structures. In insurance, a duplicate can mean 2 households for a single family, 2 policyholder records with different addresses, or a claim attached to the wrong party. Count duplicates by object, set a target, and agree who resolves a conflict between 2 records before the agent starts reading them. Write down the rule you chose.
  • Linked objects. Salesforce’s insurance service assistant page lists the objects it draws on: account, quote, quote line item, product, policyholder, coverage and claims. It says nothing about how those objects connect to policy administration or claims platforms. Check whether your org’s objects are populated, related correctly and current, since an agent that summarizes a policyholder from half-empty records will produce confident but incomplete answers. Sample 50 policyholder records by hand to see how complete they look.
  • Verified knowledge. A services company’s insurance blog says retrieval-augmented answers need verified carrier knowledge bases in Data Cloud. Gather the documents an agent would cite, such as coverage explanations, claims process descriptions and complaint procedures, and confirm that each is approved, dated and owned. Remove superseded versions. The blog says coverage explanations can’t tolerate hallucinations, and a stale document is how they happen. Name an owner who signs off each document before the agent may cite it.
  • Sensitive fields. Salesforce’s Trust Layer page says it masks sensitive data such as social security numbers, and it names no other fields. List the personal, health and financial fields your org holds, mark which an agent could read, and ask Salesforce which are masked. Confirm retention and privacy duties with counsel, because carriers hold data under insurance, privacy and sometimes health rules. Review the list with your privacy officer before building any agent instructions.

Integrations with policy and claims systems

Most carriers keep policy administration, billing and claims in core systems outside Salesforce. A Salesforce datasheet says MuleSoft connects core policy administration systems, claims platforms and legacy data sources to Agentforce through industry-specific APIs and prebuilt templates. The services company’s blog describes a MuleSoft layer bridging Salesforce with legacy core policy and claims databases through API-led connectivity. Both are vendor descriptions, and neither shows a working design or says how long it takes. The Salesforce insurance skill page, by contrast, is silent on how its objects connect to other systems.

The health check question isn’t whether integrations exist but whether they’re ready to feed an agent. For each integration, record the system of record, the direction of flow, how fresh the data is, who gets alerted when a sync fails and what an agent sees when the source system is down. Decide in advance whether an agent may write back, for instance to update an address or open a claim, or only read. My suggestion is to start with reads, since a wrong read is correctable by a person and a wrong write may reach a policy record. Confirm the design with your integration vendor.

Platform prerequisites for the agent layer

Provision Data 360. Salesforce’s setup page says that before you turn on Einstein generative AI features you should ensure Data 360 is provisioned and set up, and that it’s required for the Einstein Trust Layer to function correctly. The consultancy article uses the older name Data Cloud and also lists turning on Einstein, enabling Data Cloud and creating data libraries. Both point the same way. If you haven’t provisioned it, put it on the plan before any other agent work.

Turn on Einstein and configure the Trust Layer. The setup page gives the steps: open Einstein Setup in Setup, turn on Einstein, then configure the Trust Layer to align with your organization’s privacy and security policies, and add agents if you plan to use them. It doesn’t mention required editions or licenses, permission sets, sandboxes, My Domain or Lightning, and the page shows no publication date. Ask Salesforce for the current prerequisites in writing before your project plan assumes any.

Know what the Trust Layer says it does. Salesforce’s Trust Layer page lists grounding in CRM data, masking of sensitive data such as social security numbers, toxicity detection on model output, an audit trail and feedback, and no-retention commitments in Salesforce’s agreements with model providers such as OpenAI. It gives no detail on audit trail contents, language limits or which fields aren’t masked. Treat these as vendor statements to verify against your own privacy review and your contracts. The page doesn’t say who can see audit data or how long it’s kept, so ask Salesforce, since a regulator may ask for exactly that.

Regulatory readiness for carriers

The NAIC adopted its model bulletin on the use of AI by insurers on December 4, 2023. A Sullivan & Cromwell note says it expects licensed insurers to develop, implement and maintain a written program for AI systems that make or support regulated decisions. The bulletin is a model that states adopt individually, so it doesn’t automatically become law anywhere. The list below summarizes what the note says regulators expect, and what each item means for an agent project. It’s a general summary rather than advice on any state’s rules.

  • A written program with senior accountability. The note says senior management accountable to the board or a committee should oversee the program, which should cover the insurance life cycle from product development to fraud detection and include consumer notice that AI is in use. For your agent project, name the accountable executive and the committee that approves launch, and record the approval before go-live. Keep the approval record where an examiner can find it.
  • Documented risk controls. Expected controls include oversight of AI adoption, data quality and bias analysis, model inventories, validation and testing, data protection and record retention. Map each to a check in this post. Your health check results become evidence for data quality and data protection, and your test results become evidence for validation. Write the mapping as a simple table, with the control, the evidence and the owner in separate columns, so a reviewer can follow it in minutes.
  • Third-party diligence. The note says insurers are responsible for diligence on third-party data and model vendors, that contracts should include audit rights and cooperation duties where appropriate, and that insurers should confirm vendor compliance. A Wilson Elser note adds that New York’s department suggests vendors report known or suspected compromise of AI tools. Ask Salesforce and any implementation partner what audit and cooperation terms they’ll accept, and have counsel review them.
  • Records an examiner can request. The note says regulators may request information on AI development, use and outcomes during investigations or market conduct exams, whether or not a formal program exists. Keep an inventory of agents with purpose, risks and safeguards, plus test results, change logs and complaint records. The Wilson Elser note describes similar inventory expectations from New York and Colorado. Decide now who retrieves them when a request arrives, and how fast.
  • State differences. Adoption counts conflict. A Quarles note counted 24 adopting jurisdictions, including the District of Columbia, as of March 2025, and Wilson Elser also said 24 in its January 2026 note. A vendor blog from July 2026 says 25, with 8 more in progress, but cites only a general NAIC page, so I couldn’t verify it. Check your own state’s current position with your department. Re-check the count before every quarterly review.

Testing before launch

Salesforce’s admin blog, published in January 2025, describes 5 steps to test Agentforce: identify test scenarios and create test data, review the generated test cases, run tests and evaluate results in Testing Center, perform human validation, and review results and iterate. It says each test case should include an utterance, expected topics, expected actions and an expected outcome, and that tests should run whenever you update topics, actions or prompts. Testing Center evaluates topic selection, action selection and whether the response matches the expected outcome, using model-based interpretation, which means the grader is itself a model and needs human spot checks.

The blog also says Testing Center uses AI to evaluate test cases and so draws down Einstein Requests based on the number of cases, with usage visible in Digital Wallet. It says Testing Center is enabled for Agentforce customers to use in their production sandbox instance, wording I read as a sandbox environment but couldn’t confirm. The article is more than a year old, so confirm the tool names and consumption rules with Salesforce. For a carrier, add tests the blog doesn’t cover: wrong-policy lookups, coverage questions with ambiguous wording, requests that should escalate, attempts to extract another customer’s data and requests that need a licensed person.

Use case selection and human handoff

Start with low-risk reads. Salesforce’s datasheet lists use cases including answering routine policyholder questions, automating first notice of loss, claims intake and triage, handling policy changes and complaints with step-by-step guidance, flagging at-risk policyholders and scheduling field inspectors. The services company’s blog adds mid-term endorsements and renewals with cross-selling. These range widely in risk. My suggestion: begin with questions that read data and return it, such as proof-of-insurance requests and claim status, then move toward actions that change records.

Claims and coverage statements. A statement about what a policy covers can create legal exposure if it’s wrong. The services company’s blog says coverage explanations can’t tolerate hallucinations, and that high-value or bodily injury claims should be flagged and routed to adjusters with a summary. It also describes sending low-complexity claims such as windshield damage to fast-track adjudication workflows. Whether your regulators and policy forms permit that approach is a matter for counsel and your claims leadership, not for this post.

Handoff design. Both the datasheet and the blog stress handoff to people, with the datasheet describing a context-preserving transfer to live agents and the blog saying human-in-the-loop controls remain critical for complex claims, sensitive decisions and regulated workflows. Define the triggers for handoff in writing, such as an injury mention, a complaint, a request to cancel or a legal threat. Test each trigger, and measure how often a person has to repair a handed-off conversation.

Cost and licensing checks

Cost is rarely the main risk, but it needs a line in the plan. Salesforce’s May 2025 pricing blog describes Flex Credits priced per action, at $0.10 an action, after earlier metering at $2 per conversation. It defines actions as functions the agent executes on the platform to get information or perform tasks, and it says interactions without an action carry no charge. The checks below are the questions I’d settle before launch, and the figures are Salesforce’s own as of that post. Confirm current pricing with Salesforce.

  • Credits and actions. The blog’s example is an interaction using 3 to 6 actions, costing $0.30 to $0.60, against $2 under conversation billing. Count the actions your design triggers per interaction, because a lookup, a summary and a record update could each count. Ask Salesforce which steps count as actions in your design, and whether retries and failed actions are charged. Keep a log of real action counts during the pilot so the forecast reflects practice.
  • Unmetered licenses. The blog says a new unmetered per-user license is available for Sales, Customer Service and Industries users with uncapped use of Agentforce. It gives no price for that license or for credit bundles. Ask for both, and model the break-even between credits and per-user licenses at your expected volume. Compare the per-user price with expected credit spend for each role, and revisit the comparison when your volumes or headcount change. Confirm the license terms in writing.

An invented sizing and cost model

Everything in this section is invented, and I recomputed it in code. Assume the fictional carrier receives 18,000 servicing contacts a month, and an agent handles 20% of them, which is 3,600. Assume each handled contact triggers 5 actions. That’s 18,000 actions a month, and at $0.10 an action it comes to $1,800 a month, or $21,600 a year. At the older $2 per conversation, the same 3,600 conversations would cost $7,200 a month, or $86,400 a year. Using the blog’s range of 3 to 6 actions, the monthly cost runs from $1,080 to $2,160.

ItemAmountNote
Contacts per month18,000Invented
Share handled by the agent20%Invented
Agent-handled contacts3,600Computed
Actions per contact5Invented; blog example is 3 to 6
Actions per month18,000Computed
Monthly cost at $0.10 per action$1,800Salesforce’s May 2025 rate
Monthly cost at $2 per conversation$7,200Earlier rate
Test cases30012 intents with 25 variants each, invented
Test runs before launch1,800300 cases, rerun 6 times

The cost is small next to the cost of a single wrong coverage statement or mishandled claim, which is why the sizing model matters less than the gates. If the agent’s share of contacts grew to 30%, the monthly action cost would reach $2,700, still modest. The 1,800 test runs matter because each run consumes credits or requests under rules I couldn’t verify. Nothing I read supports the 20% handling share or the 5-action average, so use your own pilot data to replace them.

What the sources don’t tell you

Several gaps stand out. I found no independent study of how Agentforce performs for insurance carriers, no regulator guidance written specifically about Agentforce and no Salesforce page that gives a carrier-specific readiness checklist. Salesforce’s insurance skill page and datasheet give no verified results. The datasheet lists 3 percentages cited to its own fiscal 2026 customer success metrics without attaching clear labels to them, and the services company’s blog lists 5 figures, such as a 50% reduction in processing time, each pointing to a footnote that has no text. I used none of those numbers. Salesforce’s insurance skill page also warns that its capabilities may not reflect actual product details, and the datasheet’s statements that functions operate within your regulatory requirements are marketing.

Sources also conflict. The count of states that adopted the NAIC bulletin differs between the law-firm notes and the vendor blog, and each count is a snapshot. The consultancy article’s author appears as a business analyst in the text but as a different name in the page metadata. The Salesforce setup page uses the name Data 360 where the consultancy article says Data Cloud. The blogger’s limits come from 2022, and I didn’t check them against current documentation. I also read no Salesforce contract terms, so I can’t speak to liability, audit rights or data handling. Confirm all of it with Salesforce and counsel.

A 90-day preparation plan and how to review it

Days 1 to 30. Run Health Check and export the results. Count objects, fields, rules and sharing rules, and list duplicates by object. Name the accountable executive and an approval committee. Inventory the integrations that would feed an agent, and ask Salesforce for the current prerequisites in writing. My suggestion, not a sourced benchmark: treat this as a discovery month, with findings reviewed by security, compliance and the business owner at the end.

Days 31 to 60. Fix what you found, starting with security settings, admin access and duplicates. Provision Data 360, configure the Trust Layer in a sandbox and build the test set. My suggestion, not a sourced benchmark: expect the cleanup to take longer than planned in an org with many integrations, so keep scope tight and carry unresolved items into a risk register rather than ignoring them. Review the risk register every week with the business owner.

Days 61 to 90. Run tests, collect human review results and rehearse handoffs with the people who will receive them. Draft the records an examiner might ask for: the inventory, the test results, the change log and the approval. My suggestion, not a sourced benchmark: launch to a small internal group or a single low-risk intent first, with a rollback plan and a named owner for each decision. Keep the rehearsal notes for the record.

Go and no-go gates

The gates below turn the checklist into decisions. Each is a question a committee can answer yes or no, with evidence attached. The gates are my suggestion, built from the sources above, and a carrier’s compliance team should add or remove items for its states and lines. If any gate is a no, hold the launch or narrow the scope, and record the decision with the reason. Security and data gates follow the same pattern, using the Health Check export and the duplicate report as evidence.

  • Platform gate. Data 360 is provisioned, the Trust Layer is configured to your policies, and Salesforce has answered in writing on prerequisites, audit data and retention. Evidence: the configuration record and Salesforce’s written answers. If the answers are missing, the gate stays closed. Also confirm which sandbox holds the configuration under test and who can promote it to production, so the setting you tested is the setting you launch. Record the date.
  • Compliance gate. The accountable executive and committee are named, the agent is in the AI inventory, the vendor terms have been reviewed by counsel, and your state’s rules have been checked. Evidence: the program document, the inventory entry and counsel’s review note. Also confirm that complaint handling and consumer notice are in place for any customer who talks to the agent. Counsel should sign the review note, and the committee should record who voted.
  • Test gate. Test cases cover the intents, escalation triggers and abuse attempts, humans have reviewed the results, and the failures are fixed or accepted. Evidence: test results, review notes and the retest log. Rerun the tests after any change to topics, actions or prompts. Include a sample of conversations reviewed by licensed staff, and write down the pass mark before the first run so nobody moves it afterward. Keep the transcripts for the file.

FAQs

What should a Salesforce health check cover before an Agentforce rollout?

I’d cover security settings, admin access, org complexity, duplicate records, the policyholder record, integrations, platform prerequisites, agent access, regulatory readiness, vendor diligence, testing and cost. No source gave a full checklist for carriers, so this list is my synthesis of Salesforce documentation, law-firm notes and partner articles. Start with the built-in Health Check score, then add the checks that fit your lines of business, and have compliance review the result.

What’s the Health Check score, and how is it calculated?

A Salesforce Trailhead unit describes it as a Setup dashboard with a score from 0 to 100, where 100 is the most secure, measuring how closely your security settings match Salesforce’s recommended standard values. You can import your own baseline as an XML file, which the unit suggests for highly regulated industries. It doesn’t list every setting covered, so ask your administrator for the full export and keep it dated.

Is Data 360 required for Agentforce?

Salesforce’s setup page says Data 360 should be provisioned and set up before you turn on Einstein generative AI features, and that it’s required for the Trust Layer to function correctly. A consultancy article calls it Data Cloud and lists enabling it as a setup step. The page shows no date and doesn’t discuss editions or licenses, so ask Salesforce for the current requirements for your contract before you plan.

Does the NAIC bulletin apply to my company?

It depends on your state and license type. The bulletin is a model that states adopt individually, and a law-firm note says it applies to licensed insurers, not surplus lines insurers. A pair of law-firm notes counted 24 adopting jurisdictions, while a vendor blog said 25 by July 2026 without a clear source. Check your department’s current position with counsel, and remember that some states have their own rules. Verify the current count first.

What does the Trust Layer do for an insurer?

Salesforce’s page lists grounding in CRM data, masking of sensitive data such as social security numbers, toxicity detection on model output, an audit trail and feedback, and no-retention commitments in agreements with model providers. It names no other masked fields and gives no detail on audit contents or languages. Treat it as a vendor description, and verify it with your privacy review, your contract and a test. Ask Salesforce for the details in writing.

How should I test an agent before launch?

Salesforce’s admin blog describes 5 steps: identify scenarios, review generated test cases, run tests in Testing Center, validate with humans and iterate. Each case should list an utterance, expected topics, expected actions and an expected outcome. The post dates from January 2025, so confirm current tools. Add carrier-specific cases for wrong-policy lookups, ambiguous coverage questions, escalation triggers and attempts to reach another customer’s data. Keep the transcripts and reviewer notes with your records.

How much does an Agentforce agent cost?

Salesforce’s May 2025 pricing blog gives $0.10 per action for Flex Credits, after earlier metering at $2 per conversation, and says an unmetered per-user license exists without giving its price. In my invented model, 3,600 handled contacts a month at 5 actions each cost $1,800. Pricing changes, so get current terms in writing, and ask which steps count as actions in your design. Revisit the forecast after the pilot using your real action counts.

Which use cases should a carrier start with?

I’d start with low-risk reads such as proof-of-insurance requests and claim status, then move to actions that change records. Salesforce’s datasheet lists first notice of loss, claims intake and triage, policy changes and complaints, and the services company’s blog adds endorsements and renewals. Coverage statements and claims decisions carry legal exposure, so define handoff triggers, test them and let counsel and claims leadership decide what an agent may do. Keep the first release small.

By

Leave a Reply

Your email address will not be published. Required fields are marked *