Category: Agentforce and AI

  • Your AI agent is only as good as your duplicates: a data readiness guide

    Your AI agent is only as good as your duplicates: a data readiness guide

    The short answer: Agentforce data readiness comes down to four things an agent needs from your records: one record per customer, fields that mean what their labels say, a history of changes, and enough written context to ground an answer. You can test each with reports and Setup pages you already have. Clean up only what the first use case touches.

    A new hire who finds two accounts for the same company walks over to the account owner and asks which one is real. An agent doesn’t walk anywhere. It picks one, writes a reply with your company’s name on it, and moves on to the next conversation. Multiply that by every duplicate, every half-used field and every contact owned by someone who left in March, and you have the real reason people say “get the data ready first”. It isn’t a slogan. It’s the difference between a person noticing a problem and a process repeating it.

    Why is an agent more exposed to bad data than a person?

    An agent has no memory of why your data looks the way it does. A person knows that “Region” was repurposed in 2022 and that the top Acme account is the one with the contracts. An agent knows what the records and field descriptions say, and acts on it at speed.

    Two things make that worse. An agent working as a service agent runs as its own agent user, and as reported by Salesforce Ben, verifying a customer’s identity tells it who the customer is, not what it may read. So it works with the access its user has, which is often broader than anyone planned. And its mistakes repeat. A person’s slip is one slip. An agent’s slip is every conversation that hits the same record. Salesforce’s own readiness advice says the same thing in fewer words: old or incorrect data leads to unreliable outcomes, including made-up answers.

    Is there one record per customer?

    You need one record per customer because an agent can’t tell which duplicate is the live one. It may answer from the thin copy, update the wrong record, or update both.

    Here is how to look with native tools, cheapest first:

    1. A grouped report. Build an Accounts report, group by Account Name, and sort by Record Count, highest first. It only catches exact name matches, but that is where the worst offenders live.
    2. Matching and duplicate rules. Salesforce ships standard matching rules for business accounts, contacts and leads, and person accounts, and a matching rule does nothing until a duplicate rule uses it (Trailhead). They stop new duplicates. They don’t find the ones you already have.
    3. Duplicate jobs. To scan existing records, Setup has Duplicate Jobs. As reported by Salesforce Ben, they’re limited to Performance and Unlimited editions, to the standard Lead, Contact and Account objects, and can fail in orgs with many records. Check your edition first.
    4. A duplicate-bypass report. Trailhead describes a custom report type with Duplicate Record Items that shows records saved after someone clicked past the warning. That report is your list of where people routinely override the rules.

    Merging is a separate decision, and a wrong merge can’t be taken back by the platform. Our guide to merging duplicate accounts covers it.

    Do your fields mean what they say?

    An agent trusts the label and the description, so a field that means something other than its name is a trap. Salesforce’s own Trailhead guidance on agent-ready flows says agents use a flow’s description and the descriptions on its variables, and that without a solid name and description the agent can’t reliably decide what to do with the data.

    That guidance is about flows, but the habit carries over to the fields an agent reads. Pick the ten fields the first use case depends on, then:

    • Open each in Object Manager and read the Description and Help Text boxes. If they’re empty, write one sentence each.
    • Report on each picklist, grouped by value. Two spellings of one value, and a big “Other”, are what an agent can’t interpret.
    • Ask two people what each status means. If you get two answers, write a third, official one down.

    This is slow work and nothing in the platform shortcuts it, because only a human knows what a field was meant to hold.

    Does the org keep a history of changes?

    You want history because an agent that can update records will eventually update one wrongly, and “what was it before?” is the first question afterward. Field History Tracking answers it for the fields you turn it on for.

    As reported by Salesforce Ben, native tracking covers 20 fields per standard or custom object, keeps changes for up to 18 months in the UI and 24 months through the API, and starts only when you select a field, with nothing recorded retroactively. To turn it on: Object Manager, the object, Fields & Relationships, Set History Tracking. For a custom object, tick Track Field History in its Details section first. Our guide to field history tracking limits covers what to do when 20 isn’t enough.

    For readiness, the rule is plain. List the fields the agent might change, and make sure each one is tracked before the pilot, not after the first surprise.

    Is there enough text on the record to ground an answer?

    An agent can only ground an answer in what’s written down. A case with a blank description and a one-word subject gives it nothing to quote, so it fills the gap with something plausible.

    You can’t audit this with a formula, so sample it. Pull 20 recent records of the kind the agent will handle and read them as if you’d never seen the customer. Can you tell what happened and what’s next? Count the ones where you can’t. If the long-text fields hold pasted email chains or “see call notes”, the agent will quote those. Opinion: this is the check most teams skip, because there’s no report for it.

    How much of the cleanup does the first use case need?

    Only the part the first use case touches. Salesforce’s readiness guidance says not to clean all your data: define the business problem, identify only the data the agent needs for it, and get “enough good data to get the agent started”. An agent that answers order-status questions needs reliable orders and contacts. Your 2019 lead data can wait.

    So choose the use case first (our post on choosing your first Agentforce use case helps), then run the four checks against only those objects and fields. That turns “fix our data” into a list you can finish.

    What to do Monday:

    – Write down the first agent use case in one sentence and list the objects and ten fields it would read.

    – Run the grouped Accounts report and note the top ten duplicate names.

    – Open those ten fields in Object Manager and add a Description wherever it’s empty.

    – Turn on field history for any field the agent might change.

    – Download the Agent Readiness Checklist and answer the 25 questions with someone who knows the org.

    Where we come in

    Sounding, by GTM Cloud, is a read-only data quality assessment that scores duplicates, completeness, staleness, ownership and agent readiness inside your org. It’s in pre-release and in pilot, not publicly listed yet. It measures and recommends, and it won’t merge or fix anything. If you’d rather have a person do the first read, we also run the Agentforce Readiness Sprint.

    Questions people ask

    How clean does my data need to be for Agentforce?

    Clean enough for the first use case, not perfect everywhere. Salesforce’s own guidance says to identify only the data the agent needs for its task and start there. If the agent reads orders and contacts, those two objects need to be reliable. Everything else can wait for the second use case.

    Can Agentforce handle duplicate records?

    Don’t count on it. An agent works with the records it can see, and nothing in the sources we opened says it picks the right duplicate for you. Salesforce’s own readiness advice is to merge duplicates into one profile per customer before the agent runs, not after it has picked a copy.

    What is the first thing to check for AI readiness in Salesforce?

    Access, then duplicates. Find out what the agent’s user can see and change, because an agent works within that. Then run a grouped report for duplicate account names. Both take under an hour, and both often surprise people. The checklist covers the rest in a sensible order.

    Sources

  • AIforce, Claudeforce, Slackforce: a plain-English guide to the new names

    AIforce, Claudeforce, Slackforce: a plain-English guide to the new names

    The short answer: AIforce Salesforce is an umbrella name for reaching Salesforce data, workflows and agents from outside the Salesforce screens. Claudeforce is the Claude version, Slackforce the Slack version, and Agentforce Coworker the Lightning version. Koa is a model. Guardian and Agent Fabric are governance. Status ranges from generally available to pilot, and some of it is still unclear.

    Your CIO sent a one-line email: “Are we doing Claudeforce?” You opened three recaps and found seven new capitalized nouns, two of which seem to mean the same thing and one of which only appears in half the articles. Nobody has said which ones appear in Setup.

    This is a glossary, in the order you’re likely to be asked about them. Where Salesforce’s own pages give a status, we use it. Where only recaps do, we say “reported”. Features vary by edition and license, so check the release notes for your org before you rely on any of this.

    What is AIforce?

    AIforce is Salesforce’s umbrella name, announced at Dreamforce 2026 (September 15 to 17), for using Salesforce from other surfaces. Salesforce’s newsroom puts it as: “now with AIforce, all those capabilities are available to agents through any AI interface.” As reported by Salesforce Ben, it bundles Claudeforce, Slackforce and Agentforce Coworker, and Salesforce frames the idea as AI replacing the UI.

    The plumbing underneath is a developer layer that exposes Salesforce through APIs, MCP servers, plug-ins and skills. Sources name it differently: Salesforce Ben calls it the Headless Toolkit, and Forrester and Salesforce’s developer pages say Headless 360. We’ve seen no status for AIforce as a whole. It isn’t one switch, and the pieces below each have their own.

    What are Claudeforce and Slackforce?

    Claudeforce means Salesforce inside Claude. Salesforce’s own Dreamforce list labels it beta. As reported, it includes a prebuilt Salesforce MCP server with 37 sales skills, plus a Salesforce plug-in for Claude Code aimed at developers. Reports say Claude works within the connecting user’s existing permissions. A full disclosure: we use Claude in our own work, so weigh our opinions accordingly.

    Slackforce means Salesforce context inside Slack. It covers Slackforce Surfaces (live, interactive views that Slackbot builds on request, “built under existing permissions”), Slackbot in Lightning Experience and Slack Code. Salesforce’s list says GA in October. Slack’s own post says Surfaces, Slackbot in Lightning and Slack Code are “available now”, but warns that features may not show up in your workspace immediately, depending on plan and license. Slack also lists some Slackbot features as early pilot.

    What is Agentforce Coworker?

    Agentforce Coworker is an assistant built into Salesforce that can call your Agentforce agents, Flows and other tools in one conversation. Per Salesforce’s August 4 blog, an admin turns it on by searching Setup for “Agentforce Coworker”. Salesforce says it inherits the permissions and security controls already in your org.

    The status wording is inconsistent. That same post is headed “Generally Available” and also says Coworker “is in beta for Salesforce today”, with web, Microsoft Teams, ChatGPT, Claude and desktop surfaces coming later this year. Salesforce’s Dreamforce list says GA. Our reading is that the core Salesforce experience is generally available while other surfaces are still arriving. Confirm that with your account team.

    What is Koa?

    Koa is a model, not an app. Salesforce calls it its first purpose-built CRM reasoning model, built with NVIDIA on the Nemotron 3 Super architecture and trained on simulated CRM workflows, not private customer data. Its status is pilot. Salesforce reports internal benchmark gains in context retention, answer relevance and tool-call precision, but those are Salesforce’s own tests and we haven’t seen independent ones.

    Salesforce’s Dreamforce list says GA is “expected Winter 2026”, which can’t be right as written, so the real date is unclear. For an admin, Koa is a feature of the platform you’ll eventually see in a model picker, not a project for this quarter.

    What are Guardian and Agent Fabric?

    Guardian and Agent Fabric are Salesforce’s governance names, and they are the least precisely documented of the seven. Salesforce’s Winter ’27 Trailhead module lists MCP Risk Scores, Security Mesh connectors and Global Consent Manager under a “Salesforce Guardian” heading, without defining Guardian as a whole. A September 11 press release on Salesforce’s trusted-AI foundation names Guardian and Agent Fabric among the technologies behind it.

    Agent Fabric is better documented, because it predates Dreamforce. A Salesforce story from April 15, 2026 describes it as a control plane that discovers agents and MCP servers, standardizes model access and limits which agents and tools teams can use. It listed Agent Governance as generally available and Agent Broker as beta, with general availability planned for June 2026. Forrester notes the name has dropped “MuleSoft”.

    That release says many foundations “are available today” while new capabilities and a unified experience are planned from early fiscal FY28. For a concrete Guardian feature you can use now, read MCP risk scores in Security Center.

    What does all this mean if you run an org?

    Mostly it means more doors into the same house. Every surface above reaches your data through your existing permissions, so the quality of your sharing model and permission sets decides how safe each door is.

    NameWhat it isStatus per sourcesAdmin action
    AIforceUmbrella nameNo status of its ownNone; read the pieces
    ClaudeforceSalesforce inside ClaudeBetaDecide who may connect
    SlackforceSalesforce context in SlackGA in October per Salesforce; “available now” per SlackCheck Slack plan and permissions
    Agentforce CoworkerAssistant in LightningGA, with surfaces still arrivingTest in a sandbox first
    KoaCRM reasoning modelPilotWait
    GuardianGovernance layer, MCP Risk ScoresMixed; see aboveUse Security Center now
    Agent FabricAgent control planeParts GA, parts betaRead before you need it

    For a view of what was announced overall, see our Dreamforce 2026 recap for admins.

    What to do Monday:

    – Search Setup for “Agentforce Coworker” and write down whether it’s on in each org.

    – List who in your company has already connected Claude or another AI tool to Salesforce data, and how.

    – Pick one name from the table to understand properly this month, and ignore the rest.

    – Ask your account team, in writing, for the status of any item you plan to use.

    Where we come in

    If you want a second pair of eyes before any of this reaches your org, our Agentforce Readiness Sprint looks at your data and permissions the way an agent would, then picks use cases and designs guardrails. It’s a plan, not a build, and it isn’t a licensing review. It also won’t promise what a beta will do: we plan around what’s generally available. See also our page on agent readiness.

    Questions people ask

    Is Claudeforce the same as Agentforce?

    No. Agentforce is Salesforce’s platform for building and running agents inside Salesforce. Claudeforce is the route the other way: Claude, running outside Salesforce, reaching Salesforce data and skills through an MCP server. Both can touch the same records, which is why permissions matter for each.

    Do I need Slack to use any of this?

    Only for Slackforce. Agentforce Coworker is switched on in Setup and works inside Salesforce, and Claudeforce runs in Claude. Slack’s own post says Slackbot is included in certain Slack plans and reachable another way for others, so check your Slack licensing rather than assuming.

    Is any of this safe to turn on today?

    Some of it is generally available, and Salesforce and press coverage say the pieces covered here work within existing permissions. That makes your permissions the real test. Try it in a sandbox with an ordinary user, not an administrator, and read what comes back. Our guide to choosing a first Agentforce use case helps you start small.

    Sources

  • How to tell what an AI agent changed in Salesforce

    How to tell what an AI agent changed in Salesforce

    The short answer: To audit AI agent changes in Salesforce, use three sources together. Field history shows old and new values for the fields you track. Setup Audit Trail shows configuration changes, including permission changes to the agent’s user. Agentforce session tracing shows what the agent did and why. None covers everything, so switch them on before the pilot, not after.

    Say the pilot report lands on Monday and reads “41 opportunities updated over the weekend.” By Tuesday someone in sales leadership wants to know which 41, what the Amount said before, and whether a person approved any of it. You know the agent did something. You don’t yet know where to look for the details.

    The good news is that you already own most of the tools. The awkward news is that each one answers a different question, and the gaps between them are exactly where the interesting questions fall.

    Features vary by edition and license. Check the release notes for your org before you rely on any of this.

    Does an AI agent run as a user in Salesforce?

    Yes, but which user depends on the kind of agent. That decides whose name your audit trail shows.

    As reported by Salesforce Ben, a service agent working with an unidentified or merely verified customer runs as its own agent user, which is created with the Einstein Agent license and Einstein Agent User profile and starts with minimal permissions. Only when the customer is authenticated, such as a logged-in portal user, does it run as that person. Salesforce’s Trailhead material says employee agents are logged in as the person they’re helping, using that user’s permissions and org security settings.

    So two pilots can leave two different trails. A service agent’s work should point at a user you can name. An employee agent’s work points at whichever employee was chatting. Neither of the pages we read says how every agent action appears in a record’s Last Modified By field, so test it in a sandbox before you assume.

    What does field history show about agent changes?

    Field history shows who changed a tracked field, when, and what it was before and after. It’s the only native source that holds the “before” value, which is why it’s the first thing to switch on.

    The limits matter. As reported by Salesforce Ben, you can track 20 fields per standard or custom object, tracking starts when you select a field and isn’t retroactive, and changes are kept for up to 18 months in the UI and 24 months through the API. For long text and multi-select picklists you see who and when, but not the values. Our guide to field history tracking limits covers what to do when 20 isn’t enough.

    Field history also doesn’t know the change came from an agent. It records a user. If the agent has its own user, you filter on that. If it runs as an employee, an agent change and a human change look the same.

    What does Setup Audit Trail show?

    Setup Audit Trail shows changes to configuration, not to records. It’s the right place to answer “who changed the permissions of the agent’s user?” and the wrong place to answer “what did the agent do to opportunity 0061?”

    As reported by Salesforce Ben, it records metadata changes with the user who made them, including permission and profile changes. The Setup page lists the last 20 changes, and a CSV download covers the last 180 days. Page layout and record page changes show only that something changed, not what. If you want evidence beyond 180 days, download it on a schedule and keep the file.

    One Winter ’27 note: as reported by Salesforce Ben, access to the Setup Audit Trail is moving to a dedicated View Setup Audit Trail permission, as a release update. Existing access is kept automatically, so check who holds it and whether they should.

    What does Agentforce tracing show?

    Agentforce session tracing shows the conversation and the steps the agent took, stored as records in Data 360. It explains what the agent did and why. It’s the closest thing to a flight recorder.

    Salesforce’s architecture guidance says the Session Tracing Data Model records a session as a hierarchy of interactions, steps and messages, and that action steps “record every action an agent takes including API calls, data modifications.” That’s the part that helps most with “what did it change?” Salesforce’s July 21, 2026 Observability post says Observability is included for all Agentforce customers at no additional Data Cloud cost, and describes tracing of planner decisions, prompt and action flows, errors and final output. That post doesn’t use the term “GA”, and it mentions a rollout in August. The Winter ’27 platform blog marks Observability Center, the console view, as beta.

    Two prerequisites to check. Per the architecture guide, audit data is captured in Data 360 only when Audit and Feedback is turned on, on the Einstein Audit, Analytics, and Monitoring Setup page, and it’s on by default when Agentforce is turned on in an org with a single dataspace. And none of the Salesforce pages we read says tracing stores the previous value of each changed field. Don’t assume it does.

    What does each source miss?

    SourceAnswersDoesn’t answer
    Field historyWhat a tracked field was, and became, and who changed itUntracked fields; whether an agent was behind it; long text values
    Setup Audit TrailWho changed permissions or configurationAnything about record data
    Agentforce session tracingWhat the agent did, in what order, and whyConfirmed before values; changes made outside the agent
    Record owner and Last Modified ByWhich user touched the record lastAnything earlier than the last touch

    The honest summary: field history gives you the numbers, tracing gives you the story, and Setup Audit Trail tells you whether someone widened the agent’s access along the way.

    What should you set up before an agent pilot?

    Set up four things, in this order. Each takes under an hour.

    1. Give the agent its own named user wherever the agent type allows it, and note the user’s name in your pilot document. Shared logins make the trail unreadable.
    2. List the fields the agent might change and turn on field history for each. If the list is longer than 20 for one object, shorten the pilot’s scope.
    3. Confirm Audit and Feedback is on, and name who can query the tracing data. A trace nobody can read is a decoration.
    4. Rehearse one question in a sandbox. Have the agent change one record, then try to answer “what, when, who, before and after” using only these sources. Where you get stuck is your gap.

    What to do Monday:

    – Find out which kind of agent your pilot uses and which user it will run as.

    – Turn on field history for the five fields you’d least like to lose track of.

    – Open Setup and check whether Audit and Feedback is on for the agent’s org.

    – Run the sandbox rehearsal above and write down what you couldn’t answer.

    For the wider checklist behind a pilot, our post on Agentforce data readiness and the Agentforce Readiness Sprint cover the data and access side.

    Where we come in

    Wake, by GTM Cloud, is in early access. It logs changes to the fields and objects you choose, beyond the native 20-field limit, with old value, new value, who and when, and lets you restore a value. A dedicated agent view is planned, not built. Tiller, an agent action ledger, is in design and possibly never built: we’re first checking what the platform already records. Read the Wake page for the limits.

    Questions people ask

    Does Salesforce log what Agentforce changes?

    Partly. Agentforce session tracing records the agent’s steps, including data modification actions, in Data 360. Field history records old and new values for tracked fields. No single log covers both, and Salesforce’s pages we read don’t say tracing keeps before values.

    Can I see why an agent made a change?

    Yes, that’s what session tracing is for. It records planner decisions, prompt and action flows and errors, so you can follow the reasoning path. It needs Audit and Feedback turned on in Data 360. Field history can’t give you the why, only the what.

    Is there an undo for an agent’s changes?

    We found nothing in the Salesforce pages we read that offers a one-click undo for an agent run. Field history lets you read the old value and re-enter it by hand, record by record. Anything faster needs a tool that stores before values and can restore them, or a data load you build from an export. Decide that before the pilot.

    Sources

  • Dreamforce 2026 for people who weren’t there: what matters for your org

    Dreamforce 2026 for people who weren’t there: what matters for your org

    The short answer: This Dreamforce 2026 recap comes down to one idea under one new name. AIforce makes your Salesforce data and agents reachable from Claude, Slack and Lightning. Salesforce’s own list mixes generally available, beta and pilot items. At a mid-sized company the useful response is small: check who can see what, decide who may connect what, and pick one use case.

    Somebody forwarded you the keynote clip with the subject line “thoughts?” Then the recaps arrived, each with a different set of capitalized nouns, and by Thursday your VP was asking whether you were “on Claudeforce yet”. You weren’t there. You have a sandbox refresh to finish. You’d like to know which of this is real, which is a roadmap slide, and whether any of it changes your week.

    It changes a little of it. Here’s the calm version.

    Features vary by edition and license. Check the release notes for your org before you rely on any of this.

    What did Salesforce announce at Dreamforce 2026?

    Salesforce announced AIforce, an umbrella for using Salesforce data, workflows and agents from outside the Salesforce screens. As reported by Salesforce Ben, it bundles Claudeforce (Salesforce inside Claude), Slackforce (Salesforce context in Slack) and Agentforce Coworker (an assistant inside Lightning). A developer layer called the Headless Toolkit sits underneath.

    Around that sat a second group: seven job-specific agents (Piper and Hunter on the sales side), Koa (a CRM reasoning model built with NVIDIA), Agent Optimizer for tuning agents after launch, and governance pieces named Salesforce Guardian and Agent Fabric. Our plain-English guide to the new names takes each one in turn.

    What is generally available now, and what is beta or pilot?

    Salesforce’s own post-event list, “21 Things We Announced at Dreamforce 2026”, labels each item. Here is the part that matters most for an org, with the labels as Salesforce wrote them.

    ItemStatus per SalesforceNote
    Agentforce CoworkerGASalesforce’s August post also says some surfaces are still beta
    Claudeforce (Salesforce in Claude)BetaReported as open beta
    SlackforceGA in OctoberSlack’s own post says Slackforce Surfaces is available now, with licensing caveats
    Job-ready agentsMost GA nowHunter is listed as GA in November
    Long-horizon agentsPilot now, GA November ’26
    KoaPilot nowGA wording is ambiguous, see below
    Agent OptimizerGA in OctoberCheck your org, October has passed
    MCP Risk Scores (Guardian)GA with Winter ’27From the Winter ’27 platform blog

    Two cautions. Salesforce’s list doesn’t mention Guardian or Agent Fabric at all, so their status comes from other pages. And the Dreamforce list says Koa is “GA expected Winter 2026”, which can’t be right as written for a date that has passed. Ask your account team what it means.

    What about Guardian and Agent Fabric?

    Both are governance products, and the sources are thinner. A Salesforce press release dated September 11, 2026 introduces a trusted enterprise AI foundation that draws on Salesforce Guardian and Agent Fabric, says many foundational technologies “are available today”, and says new capabilities and a unified experience are planned to begin rolling out in early fiscal FY28.

    Salesforce’s Winter ’27 materials put one concrete Guardian feature in your hands now: MCP Risk Scores in Security Center, which scores third-party MCP servers Low, Medium or High. We cover it in MCP risk scores in Security Center. The rest of the governance story is a direction of travel, not a Setup page.

    Where do the sources disagree?

    They disagree on status in three places, and it’s worth knowing before you repeat any of them in a meeting.

    • Agentforce Coworker. Salesforce’s August 4 blog is headed “Generally Available” but also says it “is in beta for Salesforce today”, with web, Microsoft Teams, ChatGPT, Claude and a desktop app coming later this year. A November 4 Salesforce story lists Teams, iOS, desktop apps and a Coworker MCP server as GA in November.
    • Slackforce. Salesforce’s list says GA in October. Slack’s post says “available now” but warns that features may not appear in your workspace right away, depending on plan and license.
    • The 30,000 figure. The LeanData recap reports 30,000 customers live on Agentforce, attributed to Salesforce. We couldn’t find that number in Salesforce’s own Dreamforce list, so treat it as reported.

    None of this is sinister. Launch weeks are messy. The point is that “announced at Dreamforce” and “turned on in your org” are different dates.

    What should an admin at a mid-sized company actually do about it?

    Three things, in this order. None needs a budget conversation first.

    1. Look at permissions before you look at agents. Salesforce says Agentforce Coworker inherits the permissions and security controls you already have, and recaps say the same of Claude. Good news, with a catch: an assistant that respects access will happily surface everything a user is allowed to see, including what they were never meant to. Pick three users with the broadest access and read what they can reach. Our guide to running a user access review has the steps.
    2. Decide who is allowed to connect what. Claudeforce is an open beta, and Setup lets an admin turn Coworker on. Both are one click for someone with the right permission. Name an owner, write down which AI clients may reach org data, and test in a sandbox with an ordinary user first. Forrester’s advice for buyers is similar: try to break the permission setup before you open it up.
    3. Pick one first use case and ignore the rest. Most of this week’s list is for later. Choose something frequent, low-risk and easy to check, using our guide to choosing your first Agentforce use case.

    What can wait from the Dreamforce 2026 recap?

    Koa, long-horizon agents and the unified governance experience are pilot or early-rollout items. Read about them, ask your account team for dates in writing, and don’t build a plan around them. If an item has no status in Salesforce’s own material, assume it isn’t ready for production.

    What to do Monday:

    – Open Setup, search “Agentforce Coworker”, and note whether it’s on in each org and sandbox. Don’t change anything yet.

    – Pick three high-access users and list the objects and fields they can see that surprise you.

    – Write one sentence naming who approves a new AI connection to org data.

    – Put the table above into your own notes with your edition next to each row, and mark what you’ve confirmed.

    Where we come in

    We publish an Agent Readiness Checklist: 25 yes/no checks across data, field meaning, access, guardrails and proof. You can answer most of it in an afternoon with someone who knows the org. It’s a checklist, not a verdict, and it says nothing about which Salesforce products you should license. For a full assessment, we also run an Agentforce Readiness Sprint.

    Questions people ask

    Is AIforce a product I can turn on?

    No. It’s an umbrella name for several pieces, each with its own status. As reported, it bundles Claudeforce, Slackforce and Agentforce Coworker over a developer layer. You enable the individual pieces, such as Coworker in Setup, and each has its own edition, license and release status to check.

    Do I need to do anything about Dreamforce before Spring ’27?

    Not urgently. Check permissions, name an owner for AI connections, and watch for the Spring ’27 release notes in January. Anything that reaches your org through a beta or an open Setup toggle is worth deciding on now, because it can arrive before you’ve decided.

    Are the Dreamforce announcements already live in my org?

    Some are, many aren’t. Salesforce labels items GA, beta or pilot, and availability still varies by edition, license and region. Check the release notes for your org and ask your account team to confirm anything you plan to rely on, in writing.

    Sources

  • How to choose your first Agentforce use case

    How to choose your first Agentforce use case

    The short answer: Choose a first Agentforce use case that is frequent, low-risk, easy to check and sits on data you trust. Score each candidate from one to five on those four things, drop anything that scores a one on risk or data, and write the stop condition before you start. Answering questions or drafting for a person to review beats changing records or messaging customers.

    Say three ideas land on your desk in the same week. Sales wants an agent that updates opportunities from call notes. Support wants one that answers order questions. The COO wants one that “handles renewals”. All three are reasonable, and nobody has said which should go first, or what would make you stop.

    The first one sets how much your company trusts the next ten. Here is a way to pick it on purpose.

    What makes a good first Agentforce use case?

    A good first use case has four traits: it happens often, a mistake is cheap, a person can tell quickly whether the output is right, and the data underneath is reliable. Salesforce’s own getting-started advice says much the same: focus on one achievable use case, look for high-volume simple requests, and ground the agent on smaller, clean pieces of data before widening it.

    Here’s what each trait means in practice.

    • Frequent. You need enough real examples to learn from within weeks, not quarters. A task that comes up twice a month will take half a year to tell you anything.
    • Low-risk. If the agent is wrong, nothing irreversible happens: no money moves, no customer is told something binding, no record is silently changed.
    • Easy to check. Someone can read the output and know in seconds whether it’s right. “Is this the correct return window?” is checkable. “Is this a good account strategy?” is a matter of taste.
    • On data you trust. The records and articles the agent reads are current, deduplicated and mean what the fields say. This is the one that sinks most pilots, and it’s the part you can fix before you start.

    How do you score a use case?

    Score each candidate one to five on each trait, add up the four, and apply two rules. Use the table below, which is plain enough to paste into a spreadsheet. The scores are judgment, so have two people score separately.

    CandidateFrequent (1 = rarely, 5 = daily)Low-risk (1 = costly if wrong, 5 = harmless)Easy to check (1 = needs an expert, 5 = obvious)Trusted data (1 = messy, 5 = clean)Total (of 20)
    Example: answer common policy questions from published articles545418
    Example: update opportunities from call notes423211
    Your candidate 1
    Your candidate 2

    Both example rows are illustrative, to show the method. They aren’t measurements.

    The rules, which are our opinion and not Salesforce’s:

    1. A one in low-risk or trusted data disqualifies it as a first pick. A high total can’t buy that back.
    2. Sixteen or more is a reasonable first pilot. Eleven to fifteen means fix the weak column first. Under eleven, leave it for later.

    Also answer one yes or no question outside the score: does a named person own this and want it? A use case without an owner stalls the week after the demo.

    What are examples of good and bad first use cases?

    Good first use cases mostly read and suggest. Bad ones mostly decide and change. Here are generic examples, described by kind and not by company.

    Usually a good first pickWhyUsually a bad first pickWhy
    Answer a common internal question from published articlesFrequent, checkable against the sourceUpdate or merge records without reviewMistakes are silent and hard to find
    Summarize a case or account history for a personA reader spots errors at a glanceSend messages to customers unreviewedErrors reach the customer first
    Draft a reply for a person to approveA human is the checkAnything involving refunds, discounts or pricingMistakes cost money
    Suggest a category for incoming requests, with a person confirmingLow-risk, easy to compareWork that depends on a messy field everyone uses differentlyData trust is a one
    Find the right article for a questionRead-onlyAn open-ended task with no right answerNobody can say if it worked

    The good column starts with answer, summarize, draft or suggest. The bad column starts with update, send or decide. Acting comes second, once the reading and suggesting has proved right.

    What can go wrong even with a good candidate?

    The agent runs as a user, and it believes your data. If three records exist for the same customer, it may answer from the wrong one, confidently. If its user can see more than anyone intended, it can repeat what it finds. A low-risk use case reduces the damage; it doesn’t remove the cause.

    So before you build, check two things. First, look at the data the use case reads: duplicates, empty fields, fields that mean different things to different teams. Our post on Agentforce data readiness goes through it, and the Agent Readiness Checklist turns it into 25 yes/no checks. Second, check what the agent’s user can see and change. If the answer is “everything”, fix that first.

    The platform already helps with the rest. Salesforce’s developer documentation describes a Testing Center and a testing API for agents, so use them in a sandbox with realistic questions, not the clean ones from the demo. Salesforce’s Dreamforce 2026 list includes Agent Health Monitoring (labeled GA), and its Winter ’27 platform blog lists an Observability Center in beta. Check the status for your org first. For afterwards, see how to tell what an AI agent changed.

    How do you set a stop condition for an Agentforce pilot?

    Write down, before the pilot starts, the specific results that end it. A pilot without a stop condition becomes a permanent experiment. Fill in the blanks below with numbers that suit your volume, and agree them with the owner in advance.

    Pilot stop condition (template)

    – End date: The pilot ends on ___ whatever else happens. Extending needs a written decision.

    – Review: A named person reads every response for the first ___ days, then a sample of ___ per week.

    – Stop early if: reviewers mark more than ___ in 10 responses wrong in any week; the agent shows anyone data they shouldn’t see; any record changes that nobody expected; or the reviewer can’t keep up.

    – Continue if: the wrong-answer rate stays under ___ for ___ weeks in a row, and the owner still wants it.

    The blanks are deliberate. Salesforce’s getting-started advice says to decide what you’ll measure before launch, and we can’t know your volumes. If you can’t fill a blank in, you’ve learned the use case isn’t easy to check yet.

    What to do Monday:

    – List five candidate use cases on one sheet. Don’t evaluate yet.

    – Score each one with the table, with a second person scoring separately.

    – Take the top scorer that passes both rules and name its owner.

    – Fill in the stop condition template and send it to the owner before anyone builds anything.

    Where we come in

    The working session in week two of our Agentforce Readiness Sprint does this scoring with the people who do the job today. The Sprint lasts two to four weeks and ends with a use-case short list, guardrails and a pilot plan including a stop condition. It’s a plan, not the build, and it isn’t a licensing review.

    Questions people ask

    Should my first Agentforce use case be customer-facing?

    Usually not. An internal use case lets your own staff catch mistakes before a customer does, and gives you real examples with less at stake. Once an internal agent has run cleanly through its pilot, a customer-facing version with a human approving replies is a sensible second step.

    How long should an Agentforce pilot run?

    Long enough to see ordinary cases and a few odd ones, so volume decides more than the calendar. Pick an end date up front, check in weekly, and extend only by written decision. A pilot that has no end date has no stop condition either.

    Does this scoring only work for Agentforce?

    No. The four traits apply to any AI assistant, from a chat tool to a custom build. We use Agentforce as the example because the platform is where your data already lives, and its permissions apply to the agent. The scoring itself needs no license.

    Sources

  • MCP risk scores in Security Center: what an admin needs to know

    MCP risk scores in Security Center: what an admin needs to know

    The short answer: For MCP Salesforce security, Winter ’27 adds a first filter. Salesforce scans each third-party MCP server when it’s registered, gives it a Low, Medium or High risk score in Security Center, keeps monitoring it for changes, and lets admins block high-risk servers. It doesn’t tell you what a server will do with your data, so permissions and a short list of questions still decide the outcome.

    A developer on your team messages you at 4:30 on a Thursday. They found an MCP server that lets an AI assistant look up Salesforce accounts, update a field and draft a follow-up. It has a tidy README and a few hundred stars. “Can we connect it for a test?” You don’t know who runs it, what it logs, or what happens to the account notes it reads. You do know the assistant will act as a real user.

    Winter ’27 gives you a first signal for that conversation. It isn’t the whole answer. Features vary by edition and license, so check the release notes for your org before you rely on any of this.

    What is MCP, in two sentences?

    MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside data sources, tools and workflows; its own documentation compares it to a USB-C port for AI. An MCP server is the thing on the other end of that port: it lists the actions an AI client may call (look up a record, update a field), and the client calls them on someone’s behalf.

    For an admin, the practical point is that an MCP connection is an integration with a new kind of caller. The caller is software that decides at run time which tool to use, and it runs with whatever access the connecting user has.

    What do MCP risk scores do in Security Center?

    Per Salesforce’s Winter ’27 platform blog (October 13, 2026), Security Center assigns each third-party MCP server a Low, Medium or High risk score. Servers are scanned at registration and then monitored for changes, and admins can set policies and block high-risk servers without investigating each one by hand. The blog says all its features are generally available unless marked otherwise, and only Observability Center is marked beta.

    Salesforce’s Winter ’27 Trailhead module lists the feature under a “Salesforce Guardian” heading and gives the reason: third-party MCP servers “can be exposed to tool poisoning and rug pull attacks”, and the scores are meant to “proactively identify server threats before they compromise agent workflows.” For where Guardian sits among the Dreamforce names, see our plain-English guide to the new names.

    What don’t the sources tell us?

    Quite a lot, and it matters before you build a process on the score. In the sources we could open, Salesforce doesn’t say how the score is calculated, which signals feed it, whether an admin can override it, or which Security Center edition and license you need. Confirm what your org has, in Setup and with your account team.

    Treat the score as a screening step. It’s a signal about the server, not an assessment of your data, your permissions or your use case.

    What are tool poisoning and rug pulls?

    Tool poisoning hides instructions in a tool’s description that the AI follows, though a human reading the tool list would never notice. Salesforce’s January 2026 Agentforce MCP post describes it as a form of indirect prompt injection and recommends vetting each server and tool closely. A rug pull, in our plain-English use of the term, is a server that behaves while you evaluate it and then changes its tools or behavior later.

    That second one is why “monitored for changes” is the useful half of the feature. A review at registration only tells you about registration day.

    What should you ask before an MCP server reaches org data?

    Ask these, in this order, and write the answers down. Salesforce’s docs don’t prescribe them; they’re our opinion of a sensible minimum.

    1. Who runs it? A named company or person, a contact, and where it’s hosted.
    2. What tools does it expose? Read, write or delete? Salesforce’s own post says admins can choose which tools to import from a server, with each tool becoming an action. Import the smallest set that does the job.
    3. What leaves Salesforce? Which fields and records does it receive, where are they stored, and for how long?
    4. What does the risk score say, and when did it last change? A Medium that was Low last month deserves a conversation.
    5. Who is the connecting user? See the next section. This is where most of the risk sits.
    6. How will you see what it did? Decide before connecting how you’ll find out what the server changed.
    7. How do you turn it off? Name the person who can, and how fast.
    8. Who re-reviews it, and when? A server you approved in November isn’t the same server in May.

    How do you give the connecting user least privilege?

    Give the connecting user the narrowest access that lets the use case work, because the server can do whatever that user can do. Salesforce’s hosted MCP documentation describes requests running as the authenticated user, set up through an External Client App, and says you can restrict who may authenticate by requiring a permission set. Confirm the details against the current docs, because the product is still moving.

    In practice:

    • Create a dedicated permission set for the MCP use case. Start from nothing and add the objects and fields the use case reads. Add edit access only when read-only has worked for a while.
    • Don’t connect with an administrator. An assistant running as a System Administrator, or any user with View All Data or Modify All Data, can reach everything. Use an ordinary user’s access, or a purpose-built integration user.
    • Use one External Client App per client type, and require the permission set for pre-authorization so only approved people can connect.
    • Check field-level security, not only object access. A server that can read Contact can read every Contact field the user can.
    • Try it in a sandbox first, as an ordinary user, and read what comes back. Look for the field you hoped was hidden.

    Nothing here needs a new product. Permission sets, field-level security and the Security Center scores are already in the platform. Our guide to running a user access review covers the permission-reading part, and how to tell what an AI agent changed covers the afterwards.

    What to do Monday:

    – Open Security Center and find any third-party MCP servers already registered. Note each score and who registered it.

    – For each server, name an owner and write down the connecting user and their permission sets.

    – Block anything scored High unless someone can say in one sentence why it’s needed.

    – Draft the eight questions above as a one-page form that anyone requesting a new connection must fill in.

    Where we come in

    Our Agentforce Readiness Sprint looks at your data and permissions through an agent’s eyes: what the connecting user can see and change, usually more than anyone intended. It ends with a plan, a guardrail design and a pilot with a stop condition. It doesn’t build the integration, and it isn’t a licensing review.

    Questions people ask

    Does a Low risk score mean an MCP server is safe?

    No. Low means the scan found less to worry about at the time. It doesn’t describe what the server does with your data once access is granted, and it can change. You still need permissions, a named owner and the questions above. Treat Low as “fine to look at”, not “fine to approve”.

    Does the risk score apply to Salesforce’s own MCP servers?

    The sources we found describe scores for third-party servers. They don’t say how Salesforce-hosted servers are treated, so don’t assume. Hosted servers run as the authenticated user, so least privilege applies to them in the same way. Check the documentation for your release.

    Can I block an MCP server in Salesforce?

    According to Salesforce’s Winter ’27 platform blog, yes: admins can set policies and block high-risk servers in Security Center. Separately, admins register third-party servers in Setup and choose which tools to import, so an unregistered server shouldn’t be reachable by your agents. Test that assumption in a sandbox before relying on it.

    Sources

  • Where should a small business start with AI?

    Where should a small business start with AI?

    The short answer: To start using AI in a small business, pick one boring, frequent task where a person can check the output, such as drafting replies to common questions or summarizing call notes. Score a few candidates on value, effort and risk, set rules for what data may go into which tool, and run the first task for about a month. Leave anything customer-facing or involving money for later.

    Say a few people on your team already use an AI assistant. You can tell, because some of their emails got smoother around spring. Somebody in sales has probably pasted a customer list into a public tool to “tidy it up”. Nobody told you, and nobody did it to cause trouble. Meanwhile a vendor emails you every morning with a plan to change your entire business.

    You don’t need a strategy deck. You need to choose one thing, and to decide what never goes into the tools. Both take less time than the vendor’s demo.

    How do you start using AI in a small business?

    Start with the work, not the tool. Write down the tasks your team repeats every week that involve reading or writing, then pick one. The right tool is usually obvious once the task is clear, and it’s often one you already pay for.

    Good first tasks share four traits:

    • Frequent. It happens daily or weekly, so you’ll see results fast.
    • Text-based. Reading, writing, sorting, summarizing and extracting are jobs you can check at a glance.
    • Low-risk. If the output is wrong, the damage is small and caught before it leaves the building.
    • Easy to check. You already know what a good answer looks like, so a person can review it in seconds.

    Typical candidates: drafting first replies to common customer questions, summarizing meetings or calls, sorting incoming requests into categories, pulling fields out of emails or forms, and writing first drafts of routine documents. None of that is exciting. That’s the point. It’s the work nobody wants to do, and the cost of a mistake is a corrected draft.

    How do you choose between candidates?

    Score each one on value, effort and risk, then compare the totals. It takes twenty minutes with the person who does the task today, and it stops the loudest idea from winning by default.

    Use a simple scale. For each candidate give a 1, 2 or 3 on three questions:

    • Value: how much time or irritation does this remove? 3 is a lot.
    • Effort: how hard is it to set up and check? 3 means easy.
    • Risk: how bad is a wrong answer? 3 means low risk.

    Here is an illustrative example. These scores are made up to show the method, not measured in any real business.

    Candidate taskValueEffortRiskTotal
    Draft first replies to common customer questions3227
    Summarize call or meeting notes2338
    Sort incoming requests by type2338
    Pull key fields out of emailed orders3227
    Decide which invoices to pay2114

    Two rules keep the table honest. First, a 1 on risk is a veto, whatever the total says. Second, ties go to the task where checking is fastest. In the example, you’d start with call summaries or sorting requests and leave the invoice decision alone for a good while.

    What data rules should you set on day one?

    Write down what can be shared, with which tool, before anyone starts. The National Cybersecurity Alliance’s guidance for small businesses puts it this way: define in writing what information can and cannot be shared with outside platforms, and assume that anything entered into a public AI tool may be stored or processed by a third party.

    A first set of rules fits on half a page:

    1. Name the approved tools. Anything else needs a yes from one named person. The Cyber Readiness Institute suggests starting with an inventory of what staff already use, approved or not.
    2. Write the never list. Customer personal details, financial records, passwords, and confidential documents stay out of public tools.
    3. Prefer business plans over consumer ones. The National Cybersecurity Alliance notes that business versions often include stronger contractual data protections. Read the terms for your plan, because they differ.
    4. Treat what you paste as shared. The Cyber Readiness Institute’s wording is that anything entered into a public tool should be treated as public disclosure.
    5. A person reviews anything that leaves the building. The Small Business Administration’s guidance also says a person should assess customer messages and outreach before they go out.

    The Cyber Readiness Institute also makes a point worth keeping: governance, not prohibition, is the key. A blanket ban mostly pushes use underground. Rules people can follow work better. Our post on a two-page AI policy goes into the wording.

    How do you run the first task?

    Run it as a small experiment with an end date. Four steps are enough.

    1. Time it today. Ask the person doing the task how long it takes and how often. That’s your baseline, and without it you can’t tell whether the AI helped.
    2. Run it for four weeks with a person checking everything. Keep a note of every output they had to fix.
    3. Decide a stop condition up front. For example: if more than a set share of outputs need rewriting, pause and rethink. Choose the number before you start, not afterward.
    4. Compare, then decide. Did it save time after checking? Did anything slip through? Keep it, change it, or drop it.

    The checking is the work at first. Once the fixes dry up, you can relax the review a little, and only then try something harder.

    What should you not do?

    Skip these, in roughly this order of how often they happen:

    • Don’t start with something customer-facing and automatic. A bot that replies to customers on its own is a second or third project.
    • Don’t buy a tool before you’ve chosen a task. You’ll end up shopping for a use.
    • Don’t paste customer lists into a public tool to tidy them up. It’s an easy leak, and it needs no malice.
    • Don’t promise your board a number. Until you’ve timed it in your own business, any figure is a guess, including ours.
    • Don’t wait for perfect data. Tidy only what the first task reads. If the assistant answers from your policy documents, those documents need to be current. Your whole file server can wait.
    • Don’t treat the output as the final word. The Small Business Administration says to have another person review what these tools produce.

    If you run Salesforce, the same logic applies to agents: pick a narrow job first. Our post on choosing your first Agentforce use case covers that, and Agentforce data readiness covers what the records need.

    What to do Monday:

    – List ten tasks your team repeats weekly, then mark the ones that involve reading or writing.

    – Score your top five on value, effort and risk, with the person who does the work.

    – Ask your team which AI tools they use now, and promise no blame for honest answers.

    – Write the never list: five things that don’t go into a public tool.

    Where we come in

    We do AI consulting for companies of roughly 10 to 500 people, on any stack, with Salesforce not required. The usual starting point is a half-day AI Opportunity Workshop that scores your candidates by value, effort and risk, and ends with a written short list. We’re not a reseller and take no commission from AI vendors. See AI consulting for how it runs and what we won’t do.

    Questions people ask

    Do we need to clean up our data before using AI?

    Only the data the first task touches. An assistant that answers questions from your policies needs current, well-organized policies. It doesn’t need your whole file server sorted out. Pick the task first, then tidy what it reads, and expand the cleanup only when the next task needs more.

    Is it safe to put customer data into AI tools?

    It depends on the tool, the plan and the data. Treat anything entered into a public tool as shared. Business plans often have stronger data protections than consumer ones, but terms differ, so read yours. Keep customer personal details, financial records and passwords out until someone has checked the terms.

    How long should a first AI project run?

    About a month is enough to learn something real. Time the task before you start, check every output with a person, and set your stop condition in advance. That’s our rule of thumb, not a standard. A shorter run teaches you little. A longer one drifts.

    Sources