SI and AI deployment · Claude · ChatGPT · Agentforce · Breeze
SI and AI deployment with the controls a bank would require
Vantage Point deploys Super Intelligence (SI) and AI agents and assistants into the Salesforce and HubSpot workflows of banks, wealth management firms and insurers, and rolls out Claude for leadership teams at manufacturers, consumer brands and service companies. We cover the whole sequence: data readiness, agent connections, workflow redesign, human review, evals and upkeep when models change.
Short answer
Vantage Point deploys Super Intelligence (SI) and AI agents and assistants into the Salesforce and HubSpot workflows of banks, wealth management firms, insurers and other regulated businesses, and rolls out Claude for leadership teams in any industry. The work covers data and system readiness, agent connections through MCP and APIs, workflow redesign, human-in-the-loop design, evals and upkeep when models change. Work starts with a $7,500 two-week AI Readiness Assessment or a complimentary data readiness score, and every package has a published scope and price.
Picking a model is the quick part. Most of the effort sits in the systems, data, process and controls around the agent, and in regulated firms each of those has to stand up to an examiner. We apply the same standard for every client. We're model-neutral and work with Claude, ChatGPT, Agentforce and Breeze, so the recommendation follows the work.
The deployment sequence
| Step | What it settles | Typical length |
|---|---|---|
| 1. Data and system readiness | Can the agent reach and trust the data it needs? | About 2 weeks to score |
| 2. Agent connections | What can the agent see and do, and as whom? | 2 weeks for a first connector |
| 3. Workflow redesign | Which part of the job does the agent take on? | Part of the build |
| 4. Human-in-the-loop design | Who reviews what, and what gets recorded? | Part of the build |
| 5. Evals | Is the output right often enough to go live? | Before launch and on every change |
| 6. Model-change upkeep | Does it stay right as models and data change? | Monthly, after launch |
Agent builds usually take 4 to 8 weeks per agent once steps 1 and 2 are done. The step lengths come from our published packages.
Step 1: Data and system readiness
An agent is only as good as what it reads. Three versions of the same household, owners who left last year, empty fields and client data sitting in a custodial feed, core banking system or policy platform the CRM never sees all lead to confident wrong answers. So does an agent that can read data it shouldn't.
We score the data your first use case depends on, across duplicates, completeness, ownership, relationships, activity, siloed systems and access. The score tells you whether to build now, fix first or migrate. When the old CRM is the problem, we move you to Salesforce Financial Services Cloud or HubSpot with a field-by-field map.
Services for this step:
- AI Data Readiness (complimentary score)
- AI Readiness Assessment ($7,500, two weeks, credited against your first build)
- Data Foundations accelerator
Step 2: Agent connections through MCP and APIs
The connection decides what the AI can see and do, and whose permissions it runs under. We pick the path per use case, write it down and start read-only.
| Path | Runs as | Typical use |
|---|---|---|
| Salesforce hosted MCP servers | The signed-in user | Claude or other MCP clients reading and updating Salesforce under existing sharing rules |
| Salesforce in Claude (Claudeforce) | The signed-in user | Prebuilt sales skills inside Claude |
| Claude in Agentforce | Agentforce, through Amazon Bedrock in the Salesforce trust boundary | Agents inside Salesforce using Claude for reasoning |
| HubSpot connector and remote MCP server | The connecting HubSpot user | Claude reading and updating HubSpot records |
| ChatGPT connectors | The connecting user | ChatGPT business plans searching CRM data |
| Claude or OpenAI API with tool use | Your application | Back-office automation and custom apps |
| RecordIQ | The Salesforce user | AI actions on record pages without Agentforce licenses |
Writes come later, one action at a time, each behind an approval.
Services for this step:
- Claude and CRM integration (MCP) (from $15,000)
- Claudeforce explained
- Claude Enterprise and ChatGPT Enterprise rollout (from $12,500, up to 25 users)
- RecordIQ
Step 3: Workflow redesign
Dropping an agent into a process built for people usually gives you the same process with an extra step. We map the job as it runs today, decide which part the agent takes on, and redesign the steps either side of it so the hand-offs work.
These are the starting points we have templates for:
Advisors and relationship managers
- Meeting briefs from household history
- Follow-up drafts for review
- Notes and next steps logged to the CRM
Sales
- Account research and call prep
- Pipeline and deal reviews
- Outreach drafts in your voice
Service
- Suggested case replies
- Case summaries before escalation
- Knowledge article drafts
Compliance
- Review queues for AI-drafted communications
- AI register and evidence for exams
- Sampling and exception reports
Operations
- Document intake into structured fields
- Onboarding checklists and status updates
- Data quality checks
Admins and developers
- Apex, LWC and test drafts with Claude Code
- Flow and metadata documentation
- Org questions answered from metadata
Some steps shouldn't use a model at all. Assignment, approvals and field updates are usually cheaper and easier to audit in Flow or HubSpot workflows, and we'll say so.
Services for this step:
- SI and AI for advisor and sales productivity (from $35,000)
- SI and AI agent development (quoted per agent)
- Claude Code for CRM development (block of hours or retainer)
Step 4: Human-in-the-loop design
Regulators expect existing rules to apply to AI. For each workflow we decide which outputs a person reviews before they take effect, who that person is, what they check and what gets logged, then map it to the obligation behind it: FINRA 2210 and 3110, SEC 17a-4 and FINRA 4511 recordkeeping, Regulation S-P, Federal Reserve SR 26-2 model risk guidance, the NAIC AI model bulletin for insurers, and GDPR and UK GDPR.
| Control | How it works |
|---|---|
| Data boundary | Objects and fields per use case, classified with your security team; approved tools list |
| Least-privilege access | Read-only first, then one write action at a time with a person approving each change |
| Review before release | Every client-facing output stays a draft until a person approves it |
| Records you can produce | Prompt, output, model and approver logged, retention set to your rules, exports to your archive |
We configure the controls. Your compliance team approves them.
Services for this step:
- SI and AI compliance and supervision controls (quote-based)
- AI Governance Pack (policy, data-flow maps and AI register; quoted)
Step 5: Evals
You can't judge an agent from a demo or a handful of conversations. We build an evaluation set from your real records with pass criteria agreed up front, including the awkward cases, and score the workflow against it before launch. Every agent build includes a 30-case test set, run before launch and after every change, with scores tracked over time.
Evals are also your evidence. When compliance, internal audit or an examiner asks how you know the agent works, the scores and the change history answer it.
Services for this step:
- SI and AI agent development
- AI Agent Build package (one production agent, 4 to 8 weeks, quoted)
Step 6: Model-change upkeep
Vendors ship new models every few months, retire features, change connectors and update pricing. Data and permissions change too, and users try requests nobody tested. Without an owner, quality drifts until a bad answer reaches a client.
After launch we stay on as the owner or alongside yours: weekly sample reviews, an eval re-run after any model, prompt or permission change, a monthly vendor release review across Anthropic, OpenAI, Salesforce and HubSpot, a quarterly access review of connectors, and new use cases built from the backlog. Each month you get a short report on what changed, what we tested, what we fixed and what to build next. For the reasoning behind this approach, read AI needs a deployment layer, and regulated firms need it most.
Services for this step:
- SI and AI support and optimization (block of hours, retainer or managed services)
- Managed services (every tier includes monitoring and tuning of AI agents we built)
- Fractional AI Architect (8, 16 or 32 hours a month from $1,800; 3-month minimum)
What changed at Dreamforce 2026
The Dreamforce 2026 announcements put agents in more of the places your staff already work: inside Claude, inside Slack, and in a new teammate that runs across Salesforce. For a regulated firm, the questions for each one are the same: what data it can reach, under whose permissions, with what record, and who reviews the output. Salesforce's own pages answer some of that and leave the rest to the firm doing the deployment.
| Announcement | What Salesforce says | What to settle first |
|---|---|---|
| AIforce | The trusted layer that brings all your business data and workflows to any agent through MCP servers, APIs and CLI tools. Salesforce says it makes Salesforce in Claude possible. | Which identity an agent acts under, which objects it can read and which it can change. Salesforce's pages don't describe a log of what an agent read or did, so plan to keep your own record of prompt, output and approver. |
| Claudeforce | Announced August 26, 2026. Launches with Salesforce in Claude, a plugin with 37 prebuilt sales skills. Actions route through Salesforce so business rules are enforced, and an admin manages authentication and permissions centrally. Claude is available through Amazon Bedrock within the Salesforce Trust Boundary, for customers including those in regulated industries. | Your roles, sharing rules and field-level security, because the agent inherits whatever is there today. At announcement it was available to select pilot customers, so confirm current availability with your Salesforce team. |
| Slackforce Surfaces | Lets a person ask Slackbot for a dashboard, report, executive deck or interactive view that stays connected to its source data. Salesforce says Surfaces are built under existing permissions. The live data feature begins rolling out in October. | Channel membership and sharing habits, and how Surface content is retained and supervised. Salesforce's page doesn't address retention or supervision. |
| Agentforce Coworker | An autonomous AI teammate that works across Salesforce, Slack and Teams and indexes 300+ enterprise sources. Salesforce's Claudeforce release adds that Claude powers it by default. | Which sources it may index, which flows it may trigger without a person, and where a human approves before a record changes or anything reaches a client. The product page doesn't describe approval steps or logging. |
Our answer to all four is the six steps above. If you're weighing whether we're the right partner for that work, see whether Vantage Point is right for you. For how Claude fits with Agentforce, read Claudeforce explained.
Pick the platform for the job
We're partners with all four, so the recommendation follows the work.
| The job | Usually the best fit | Why |
|---|---|---|
| Advisor or seller asks questions across the CRM, email and files | Claude or ChatGPT connected through MCP | Runs as the user, respects CRM permissions, works across tools |
| Customer-facing service agent in Salesforce channels | Agentforce (on Claude where supported) | Salesforce topics, actions, trust layer and audit |
| Customer agent or content help for a HubSpot team | Breeze | Built into HubSpot, priced in HubSpot credits |
| Contracts, case histories and transcripts summarized into fields | Claude via API with structured outputs | Long-document reasoning with validated writes |
| Apex, LWC, tests and metadata work | Claude Code | Agentic coding inside your repository and pipeline |
| AI answers on Salesforce record pages without Agentforce licenses | RecordIQ (Vantage accelerator) | Any major model, sharing enforced, audit log on every call |
| Assignment, approvals and field updates | Flow or HubSpot workflows | Deterministic, cheaper and easier to audit than a model |
More: Claude consulting · OpenAI consulting · Agentforce consulting · Breeze consulting · Agentforce vs Claude
What deployment looks like by firm type
Banks and credit unions. Customer data is usually split between the core, the loan origination system and the CRM, so step 1 often means getting core data into Salesforce before any agent reads it. Early workflows tend to be service and banker support: case summaries, referral routing and meeting prep for commercial bankers. Model risk and third-party risk teams join at step 4. See banking.
RIAs and wealth management firms. The strongest early use cases sit around the advisor: meeting briefs from household history, follow-up drafts and notes logged back to the CRM. They depend on clean households and custodial data in the CRM, and client-facing drafts go through the review your compliance team already runs. See wealth management.
Insurers, agencies and IMOs. Service and producer support come first: policy status questions, commission and appointment questions, renewal preparation. The policy system stays the source of truth, and the firm's written AI program under the NAIC bulletin is in place before the first agent goes live. See insurance.
Manufacturers, consumer brands and service companies. Most start with Claude for the leadership team: a governed workspace, training by role, and a pilot that shows where AI saves time before a wider rollout. ERP and operations data usually come next, and our team has worked with legacy manufacturing systems for decades. See Claude for leadership teams.
How an engagement runs
- Assess (step 1). Use cases, data, permissions and compliance, scored in two weeks.
- Set the boundary (steps 2 and 4). Which data, which tool, which connection path and which reviews, agreed in writing.
- Build read-only (steps 2 and 3). Connect, redesign the workflow and build it without write access.
- Score it (step 5). Run the evaluation set and fix until it passes.
- Add writes behind approval (step 4). One at a time, each logged.
- Train and launch. Role-based, on each team's own work, with weekly adoption reporting.
- Run and expand (step 6). Re-test on every change and add the next use case from the data.
In practice
Insurance and payroll brokerage. Our RecordIQ accelerator runs in production on Salesforce Case records, where AI drafts suggested replies that service reps review and send.
Our own firm. Vantage Point runs on the same approach: Claude connected to our Salesforce and HubSpot through MCP for account research, pipeline reviews and content work, read-first with approved writes.
Senior team. 700+ CRM engagements for 175+ clients, with nearly $2 trillion in client assets, led by consultants who have run regulated businesses.
Published SI and AI packages
Fixed scope, published prices. Licenses and usage are billed by the vendors.
SI and AI support and optimization (step 6) runs on a block of hours, a retainer or managed services. A Claude pilot for your leadership team is quoted after a 30-minute call.
Complimentary accelerators: Claude MCP Accelerator · Agentforce Discovery and POC · Breeze AI Discovery and POC · Data Foundations
All SI and AI services
Each service runs on its own or as part of a package. Every page lists what you get, how it comes together, when it's the wrong call and the mistakes we help you avoid.
| Service | Step | Price |
|---|---|---|
| AI Readiness Assessment | 1 | $7,500 |
| SI and AI Data Readiness for Salesforce and HubSpot | 1 | Complimentary |
| Claude + CRM Integration (MCP) | 2 | From $15,000 |
| Claude Enterprise and ChatGPT Enterprise Rollout | 2 | From $12,500 (up to 25 users) |
| Claudeforce: Claude in Salesforce, Agentforce and Slack | 2 | Quote-based per agent |
| SI and AI for Advisor and Sales Productivity | 3 | From $35,000 |
| SI and AI Agent Development | 3 to 5 | Quote-based per agent |
| Claude Code for Salesforce and HubSpot Development | 3 | Block of hours or retainer |
| SI and AI Compliance and Supervision Controls | 4 | Quote-based |
| SI and AI Support and Optimization | 6 | Block of hours, retainer or managed services |
SI and AI deployment: frequently asked questions
What does SI and AI deployment involve for a regulated firm?
SI and AI deployment covers six steps: data and system readiness, agent connections through MCP and APIs, workflow redesign, human-in-the-loop design, evaluation sets, and upkeep as models change. In banks, wealth firms and insurers it also includes mapping each workflow to supervision and recordkeeping rules. Vantage Point runs all six steps on Salesforce and HubSpot.
Do you deploy AI for companies outside financial services?
Yes. Manufacturers, consumer brands and service companies usually start with a Claude pilot for the leadership team, then move to workflows that touch ERP, operations or CRM data. They get the same controls we build for banks: SSO, retention settings, approved connectors, audit logs and human review.
How long does it take to deploy an AI agent?
A readiness score takes about two weeks, and a first CRM connection about two weeks. Building one production agent usually takes 4 to 8 weeks, including a 30-case evaluation set run before launch. The AI + CRM Jumpstart takes a first workflow from pilot to production in 4 to 6 weeks.
What does SI and AI deployment cost?
Vantage Point's AI Readiness Assessment is $7,500 for two weeks and is credited against your first build. A CRM Connector Fast Start starts at $15,000 and the AI + CRM Jumpstart at $35,000. Agent builds and governance work are quoted, and licenses and usage are billed by the vendors.
How do you connect AI agents to Salesforce and HubSpot?
Through Salesforce hosted MCP servers, HubSpot's connector and remote MCP server, ChatGPT connectors, Claude in Agentforce, or the Claude and OpenAI APIs with tool use. We pick the path per use case, start read-only under the user's existing permissions and add write actions one at a time behind an approval.
What is human-in-the-loop design for AI in financial services?
It decides which AI outputs a person reviews before they take effect, who reviews them, what they check and what gets logged. Client-facing communications, record changes and decisions that affect an account or policy usually need review, and the prompt, output, model and approver are kept under the firm's retention rules.
What are AI evals and why do they matter?
Evals are sets of real cases with agreed pass criteria that an AI workflow is scored against. They show whether an agent is accurate enough to launch, catch regressions after a model or prompt change, and give compliance and auditors evidence that the workflow works.
Who keeps AI agents working after launch?
Vantage Point's AI support re-runs evaluation sets after any model, prompt or permission change, reviews vendor releases monthly and reviews connector access quarterly, with a monthly report. It runs on a block of hours, a retainer or managed services, and every managed services tier includes monitoring of agents we built.
Talk to a client before you decide
Ask to speak with one of our clients before you sign anything. Tell us your industry and what you're planning, and we'll introduce you to a client who has done similar work with us, whether that's a financial services firm, a manufacturer, a consumer brand or a services company. You can speak with them on your own, without us on the call.
Put SI and AI to work, with controls
Book a free assessment with a senior consultant. We'll look at your data and setup and tell you which workflow to start with and what has to be true before an agent goes live.
Free, about 3 minutes: take the SI and AI readiness quiz. Answer 10 questions and get an instant readiness score across 10 areas. No email needed to see your results.
Dreamforce details come from Salesforce's own announcements and product pages, and availability can change. Updated October 13, 2026.
