Salesforce and Generative AI: How We Use AI Tools to Ship Faster and Deliver Better Results

Three years ago, building a moderately complex Salesforce flow meant a developer or admin spent the better part of a day clicking through Flow Builder, writing variable assignments, and testing every branch. Today, the same flow can be scaffolded in minutes — described in plain English, generated by an AI assistant, then refined by a human who actually understands the business.

That shift is not theoretical. It is happening right now in our delivery practice at SaaSKool, and it is changing the economics of what NZ and AU businesses can expect from a Salesforce partner. Faster builds. Cleaner code. Better documentation. More time spent on the things that actually require judgement, and less time spent on the boilerplate that does not.

But it is also easy to get this wrong. Generative AI is a force multiplier — it amplifies whatever discipline you bring to it. Without guardrails, it will happily produce confident-sounding code that breaks in production, or design a data model that looks elegant but ignores critical edge cases. The teams getting real value out of GenAI are the ones treating it as a sharp tool, not a magic wand.

Here is what is actually working for us, what is not, and what NZ businesses should expect from a modern Salesforce partner in 2026.

The Salesforce AI Landscape in 2026

Before talking about the tools we use to build, it is worth understanding what Salesforce itself has shipped. The platform has moved aggressively into generative AI over the past two years, and the surface area is genuinely useful — provided you know which pieces matter for your situation.

Agentforce

Agentforce is Salesforce's flagship platform for building autonomous agents — AI workers that can take actions on records, route cases, qualify leads, and handle service interactions. It is the successor to the original Einstein Copilot, and it is the piece most NZ businesses are now asking about.

A practical example: a property management company we work with uses an Agentforce service agent to handle the first response on tenant maintenance requests. The agent reads the message, classifies the issue, checks the lease record for relevant context, and either resolves the request directly or routes it to the right human with a complete summary. What used to take a coordinator 10 minutes per ticket now takes 30 seconds, and the audit trail is better than what humans were producing.

The catch is that Agentforce is not plug-and-play. It needs a clean data foundation, well-defined topics and actions, and clear guardrails about what the agent is and is not allowed to do. This is where a lot of organisations underestimate the work — the AI is impressive, but the implementation discipline is identical to any other Salesforce build.

Einstein Trust Layer

Every conversation about Salesforce AI in regulated industries hits the same question: where does our data go? The Einstein Trust Layer is Salesforce's answer. It masks personally identifiable information before prompts are sent to large language models, enforces zero data retention with model providers, and keeps audit logs of every AI interaction.

For NZ businesses subject to the Privacy Act 2020, this matters. You can demonstrate to your privacy officer, your board, and ultimately the Privacy Commissioner that customer data is not being used to train third-party models or sitting in some external vendor's logs. We always start AI-related projects with a Trust Layer review — it is the foundation that makes everything else defensible.

Prompt Builder and Model Builder

Prompt Builder lets admins create reusable, grounded prompt templates directly in Salesforce. Instead of asking an AI to "write a follow-up email," you build a template that pulls in the actual account context, recent activities, and product information, then generates a tailored draft. It is the most underrated feature in the Einstein suite.

Model Builder takes this further by letting you bring your own model — whether that is OpenAI, Anthropic, Google Vertex, or a model hosted on Amazon Bedrock — and use it inside Salesforce while still benefiting from the Trust Layer. For organisations with specific compliance or cost preferences, this matters.

Einstein for Developers

This is the one that has changed our day-to-day work the most. Einstein for Developers is an AI assistant for Apex and Lightning Web Component development, embedded in VS Code. It generates code from natural language, suggests test classes, and explains existing Apex you have never seen before. We will come back to how we use this in a moment.

The Tools Our Team Uses Every Day

Salesforce's native AI is one part of the picture. The other part is the broader ecosystem of generative AI tools that have become essential to how a modern consulting team operates. Here is what is actually in our toolkit.

Claude and ChatGPT for Discovery and Design

The earliest phase of any Salesforce project is understanding what the client actually needs — which is rarely what they initially describe. We use Claude heavily during discovery to:

• Synthesise meeting transcripts into structured requirements documents
• Identify gaps and ambiguities in business processes a client has described
• Generate process diagrams from narrative descriptions
• Draft data models and entity relationship diagrams for review
• Compare multiple solution approaches with their tradeoffs spelled out

The key here is treating the AI as a thinking partner, not a writer. The output always goes through a human review before reaching the client, but the speed at which we can iterate on ideas has roughly tripled. A discovery workshop that used to take a week to write up now takes a day.

GitHub Copilot for Apex and LWC

For our developers, GitHub Copilot is the constant background helper. It is particularly strong at:

• Boilerplate code (trigger handlers, test data factories, REST endpoint scaffolding)
• Repetitive patterns (similar SOQL queries with small variations)
• Common Apex idioms (bulkification patterns, error handling)
• Lightning Web Component markup and JavaScript glue code

We do not let Copilot write business logic unsupervised. Every suggestion is reviewed, and anything that touches actual rules of the business gets written by a human or refactored heavily. But for the 60% of any codebase that is mechanical scaffolding, the productivity gain is significant.

Einstein for Developers for Salesforce-Specific Work

For Salesforce-specific patterns — things like Apex governor limit awareness, proper use of platform events, or generating SOQL that respects sharing rules — Einstein for Developers is purpose-built and consistently better than generic coding assistants. We use it for:

• Generating Apex test classes that actually cover the right branches
• Writing SOQL queries with proper field-level security checks
• Building trigger handlers that follow Salesforce best practices for bulk operations
• Explaining inherited or legacy Apex code that has no documentation

The test class generation alone has been a major win. Writing tests is the part most developers procrastinate on, and having an assistant that produces a credible first draft means coverage stays high without becoming a chore.

Flow Builder Copilot

Inside Salesforce itself, the Flow Builder now supports natural language flow creation. You describe what you want — "when a high-value opportunity reaches negotiation stage, notify the account owner's manager and create a task for executive review" — and it builds the flow structure for you to refine.

This is genuinely changing how we approach declarative work. Our admins use it to draft initial flows that they then tune, rather than building from a blank canvas. The first version is rarely right, but it is much faster to edit than to create.

Claude for Documentation and Knowledge Transfer

One of the perennial complaints about Salesforce consulting is that documentation often lags behind what was actually built. We have largely solved this internally by using Claude to:

• Generate end-user training material from configuration screenshots and notes
• Produce admin runbooks for the systems we hand over
• Maintain up-to-date data dictionaries as objects evolve
• Write release notes from git commit history and JIRA tickets

The output is consistent, comprehensive, and importantly — it actually gets produced. Documentation that previously got skipped because nobody had time now ships with every project.

AI-Assisted Code Review

Every pull request in our development workflow gets reviewed by both a human and an AI assistant. The AI catches:

• Missing null checks and edge cases
• Inefficient SOQL patterns that would have broken at scale
• Test classes that hit coverage thresholds but do not actually assert meaningful behaviour
• Inconsistencies between code and the corresponding documentation

This is not a replacement for human code review — context, intent, and architectural judgement still need experienced eyes. But the AI is excellent at catching the mechanical issues that humans miss when they are tired or hurried. The combination of both is materially better than either alone.

How This Changes a Real Project

Let us walk through a recent engagement to make this concrete. An NZ-based wholesale distributor came to us needing a Salesforce Service Cloud implementation with custom case routing, integration to their ERP, and an Experience Cloud portal for their B2B customers. Three years ago, this would have been a 14-week build. We delivered it in seven.

Here is what AI changed at each stage.

Discovery and design. Five client workshops over two weeks, with transcripts processed by Claude into a structured requirements document the same day. Every requirement was linked back to a verbatim client quote so we could verify nothing had been invented or misinterpreted. Process flows were drafted by AI and reviewed in client sessions, cutting the design iteration cycle from days to hours.

Build phase. Apex development was 40% faster with Einstein for Developers handling the boilerplate. Flow Builder Copilot drafted initial flow structures that our admins refined. Lightning Web Components for the customer portal were scaffolded by Copilot, with our developers focusing on the actual interaction logic.

Testing. Test classes were generated by AI as a first pass, then reviewed and extended by developers. We achieved 92% code coverage with meaningful assertions, not the 75% rubber-stamp coverage common in rushed implementations.

Documentation. A complete admin guide, end-user training material, and integration documentation produced alongside the build, not bolted on at the end. The client got a higher-quality knowledge handover than most enterprise projects we have seen.

Post-go-live. The Agentforce service agent we deployed for first-line case triage is handling around 30% of incoming cases without human intervention, freeing up the support team to focus on complex issues. Six months in, it is still learning from human corrections and getting better.

The total cost to the client was about 35% lower than our original estimate based on traditional delivery methods. More importantly, the quality is higher, the documentation is better, and the system is more maintainable.

What AI Will Not Do — And Where the Risk Lives

It is tempting to read all of this and assume AI is now doing the consulting. It is not, and the places where it falls short are exactly the places where experience matters most.

AI does not understand your business. It can produce plausible-looking output for any prompt, but it has no idea whether your sales process makes sense, whether the data model you described is going to scale, or whether the integration approach you are considering is actually a good idea. Judgement is still entirely human.

AI does not know what it does not know. A confident wrong answer is worse than no answer. We have seen AI-generated Apex that compiles cleanly and passes tests but contains a subtle bulkification bug that only surfaces under production load. A senior developer catches that. An AI does not.

AI cannot navigate stakeholders. Half of consulting is reading the room — understanding when a CFO is sceptical, when a sales manager is feeling threatened by automation, when a board needs reassurance about risk. That work is intensely human and remains so.

AI hallucinations are real. Generated code can reference Apex methods that do not exist. Generated flow logic can call invocable actions that have never been built. Without a human verifying every line, these errors slip into production. Our entire workflow assumes AI output is wrong until a human has confirmed it is right.

Compliance is not negotiable. For NZ organisations under the Privacy Act 2020 and AU organisations under the Privacy Act 1988, you cannot just pipe customer data into a public AI service. The Trust Layer, masked prompts, and strict tool selection are mandatory, not optional.

What This Means for NZ and AU Businesses

If you are considering a Salesforce implementation, an enhancement, or even just a health check in 2026, the right question is no longer "is your partner using AI?" — most are, in some form. The right questions are:

How does your partner use AI responsibly? Ask about their data handling, their use of the Trust Layer, and what happens to client information that flows through AI tools. If the answer is hand-wavy, that is a flag.

What quality controls are in place? Generated code, generated flows, and generated documents all need human review. Ask specifically how that review happens, who does it, and what gets caught.

How are time savings being passed on? A modern AI-augmented delivery model should be cheaper and faster than a traditional one. If a partner is using AI internally but charging the same as if they were not, the value is not being shared with you.

What stays human? The discovery, the architectural judgement, the stakeholder management, the change management — all of this should be clearly led by experienced humans. If your partner is automating that part too, be very cautious.

How SaaSKool Approaches AI-Augmented Delivery

For our clients, AI is a tool we use, not a service we sell. Every engagement starts with the same scoping process: what does success look like, what are the constraints, what is the change management plan, what does the timeline really need to be. AI then shows up in the execution — making us faster, more thorough, and more consistent, but never replacing the human work that actually moves the needle.

A few principles guide how we work:

• Trust Layer or equivalent always on. No client data ever flows through an AI tool without appropriate masking and zero-retention guarantees.
• AI-generated artifacts are always human-reviewed. Code, flows, documents, configurations — nothing reaches a client without a human signature.
• Time savings flow back to the client. Faster delivery should mean lower cost, broader scope for the same budget, or more time spent on the parts that really matter.
• We document what we used and why. Clients receive a transparent account of which AI tools touched their project and how, so they can make informed decisions about ongoing use.

The Next Eighteen Months

The pace of change here is genuinely unusual. Agentforce capabilities that did not exist a year ago are now standard. Einstein for Developers is significantly better than it was six months ago. Tools we evaluated and rejected as immature are now production-ready.

We expect the next 18 months to bring autonomous agents that handle increasingly complex workflows, better integration between Salesforce AI and the broader productivity stack, and far more sophisticated assistance for non-technical admins. The teams that win will be the ones treating AI as a discipline rather than a novelty — investing in skills, processes, and guardrails rather than chasing shiny features.

If you are a NZ or AU business wondering how AI should factor into your Salesforce roadmap — whether that is a first implementation, an upgrade, or an Agentforce pilot — get in touch. We are happy to walk through the practical options and help you separate the genuinely useful from the merely impressive.

The future of Salesforce delivery is here. The question is not whether to use AI, but how to use it well.