The 10 Commandments of AI in the Salesforce EcosystemTen rules from 44 months, six platforms, and thousands of conversations. Almost none of it was about choosing the right tool.

Six ecosystems, every milestone and hard lesson

My AI learning curve, December 2022 to August 2026. Click or hover any milestone to see the detail; toggle platforms in the legend.

Be a good human.

Before I automated a single workflow, I looked at the people it would touch. Who does this work today? What happens to them when the agent takes over?

I'll be honest about the part most AI posts skip: sometimes the answer is that a role goes away. I've built Flows that triggered their own downstream Flows and did work an admin used to own. Pretending that doesn't happen isn't kindness — it's cost-cutting with better PR. Being a good human means naming it early and building the plan for those people into the project, not bolting it on after: what they retrain into, where they land in the new layer, how you move them up the stack instead of out the door.

Across twenty PwC sessions, the technology was never the hard part. The human questions were. The Salesforce ecosystem has always been a people ecosystem, and AI raises the stakes. If your strategy has no plan for the humans it touches, it isn't a strategy. It's a cost-cutting exercise with a better name.

Agentforce is your home base, not your entire operation.

Most teams resist this the longest.

For my first year I was running multiple AI tools without admitting it to myself. ChatGPT for operations. Claude for precise writing and cleaning up emails at 11pm — roughly 150 conversations a month in the early going. Gemini for research. Salesforce AI for enterprise work. MidJourney for images. All at once, routed by instinct instead of on purpose. That gap cost me months.

Agentforce isn't the only AI your team runs, either. Your developers are in Claude Code or Codex. Marketing is in ChatGPT. Data is in Gemini. So I built a Multi-AI Playground that fires the same prompt at Claude, GPT-4o, and Gemini 2.5 Pro at once, with a judge model comparing the answers. It turned into production infrastructure: contextual model routing. One client's voice pipeline dropped 50% in cost per conversation against its single-model baseline. A Slack-integrated MCP workflow, 82%. Not cheaper models. Smarter routing.

Each model has a cognitive fingerprint. Claude thinks carefully. ChatGPT iterates fast. Gemini cross-references. Salesforce AI lives in the enterprise data layer. Codex audits at scale. Agentforce is your home base, not your whole operation.

Embed in the system of record. Don't fight it.

I've operated on this for a decade. In 2016 I was a foundational client during Workato's early build, shaping tables, permissions, and reporting through real feedback. Before that: a CRM build for Chabad.org, financial integrations at Sallie Mae and Nucleus. Integration has been the spine of my practice since before I ever touched a large language model.

So when clients asked me to rip out Salesforce and replace it with AI, I already knew the answer. The CRM holds ten years of data, custom objects, validation rules, approval chains, compliance baked into every layer, and 200 trained people. You're not replacing that. Not in six months. Not ever.

Wrap around it instead. Build an AI layer that reads from it, writes back to it, and handles what it was never designed for. My first AI SDR, in May 2025, didn't replace anyone — it sat inside the sales and CRM workflow and handled the conversations the team had no time for. AI Operations the next month, same pattern. Voice agents in 2026, same pattern.

That institutional knowledge is your moat. No vendor demo replicates it. Most AI implementations fail because they try to be the new system instead of the new layer.

AI will lie to you with confidence. Catch it before it ships.

This is what separates people who use AI from people who use AI well.

I caught an AI-generated strategy deck with the wrong unit count — 9,500 where it should have read 95,000 — sitting next to invented assumptions about APIs that didn't exist. Polished deck. Wrong numbers. If I don't catch it, it goes to a real client with real money on the line.

That was the easy version. The hard one: I was reviewing an AI's analysis of medical call-center data. Real calls, real patients, real metrics. It looked thorough. So I ran the same data through Claude as an audit, and Claude caught inflated percentages, hallucinated citations, and conclusions that didn't follow from the numbers. AI auditing AI. The best check is often a second model challenging the first — then your judgment on the disagreement.

I pushed my team to adopt AI early, and then I had to read what came back: AI garbage submitted as finished work. Confidence is not accuracy. Every output that touches a client or a record gets caught before it ships, not after.

The interface is disappearing. Don't build on top of it.

October 2025, deep in a financial-services project, it clicked: the interface itself is temporary.

Every company was pouring money into dashboards and polished front-ends for AI. Every one of those bets is exposed, because AI is moving from "a tool you operate through a screen" to "a layer that runs underneath the workflow." Ask the honest question about any feature you're building: where is this in twelve months, and will anyone still open it?

Parker Harris said the quiet part out loud at TDX 2026: why should you ever log into Salesforce again? Headless 360 — no browser required. Every workflow, business rule, and data object exposed as an API, an MCP tool, or a CLI command. Your agents get live access to the whole org from inside Slack, from Claude Code, from Codex. The vendor that built the interface is now telling you the seat is optional.

A leader once told me to quadruple my output because "we have AI now." Not better work — more work. Same mistake: optimizing the visible layer instead of rethinking the workflow underneath it. The companies that win aren't building better interfaces for AI. They're removing the need for the interface at all.

Just because you can build it doesn't mean you should.

Spring 2026. AI makes building so cheap it's tempting to build everything. So I did.

I built a custom AI email pipeline — classifier, draft agent, routing logic, the whole hand-rolled stack. Three days. Clean code, working system. Day four I killed it. I'd found an off-the-shelf tool that did 120% of what I needed, maintained by a team, auto-updated, for less than my own infrastructure cost. Three days into the trash. Right call.

Building feels free now. It isn't. The cost was never the code. It's maintenance, edge cases, security patches, and the attention tax of running something yourself when someone else runs it better. The same math governs your org: not every process needs a custom LWC or a hand-rolled agent when an AppExchange package or a native Agentforce action already covers 80% of it.

Every "should I build this?" runs through two filters now. One: is the decision binary or fluffy? Clear rules, automate it; judgment that lives in someone's head, a human keeps it. Two: can I buy something that does 80%? If yes, buy. If no, build. If unsure, wait. Those two filters have saved me more than any tool.

Your AI has no memory. Build it one.

Nobody's talking about this yet, and it's going to bite everyone.

You can run agents, orchestrate Flows, and verify every output, and none of it matters if every conversation starts from zero. No single model remembers enough on its own. Your Einstein agent doesn't know what the customer told your Claude-powered support bot yesterday. ChatGPT doesn't know the decision you logged in Salesforce last week. Every platform is an island, and your context lives in your head and nowhere else.

In March 2026 I built a Second Brain — a personal knowledge base with its own MCP server, about 300 lines of Python — so Claude could carry context across sessions. I was re-explaining the same thing for the fifteenth time and realized that if I didn't build the memory layer, nobody would.

The building blocks already exist: Data Cloud, Knowledge Articles, MCP connections, persistent storage. Most teams just never wire them together. They deploy agents that are technically functional and contextually blind. Your org's data is the memory — but only if your agents can reach it and carry what they learn from one interaction to the next.

Automate the verification, or you'll skip it.

Rule 4 is you catching the lie. Rule 8 is building the system that catches it when you're not looking — because manual review doesn't scale, and the day you're slammed is the day the bad output ships.

So I build self-checking systems: one model writes, another reviews, and disagreements get logged and queued for me before they reach production. The system surfaces its own issues instead of waiting for me to go find them. And you back everything up — I once ran a terminal command I didn't fully understand and lost my work in thirty seconds, with no loop in place to catch what was already gone.

Inside Salesforce this is a monitoring layer on every agent deployment: what did the agent do, what did it decide, where did it deviate from expectation? That's where Flows grow up into governance — checkpoints that validate what an agent produced before it writes a record or sends an email.

Launching is easy. Making it valuable is the hard part.

The AI landscape inside Salesforce moves weekly. Features that didn't exist in January are GA by June. MCP integrations that needed custom middleware last quarter ship native this quarter. The defense is critical thinking, not cynicism.

Be honest about where the work actually is. Turning AI on is easy — that's the 80%. The other 20%, getting it to change how your team operates, is the hardest part and where nearly all the value lives. Most teams do the easy 80%, declare victory, and six months later nothing has changed, because nobody tuned the prompts, built the loops, trained the people, or fixed the data quality the AI just exposed. The easy part is a trap the moment you mistake it for the finish line.

The learning compounds — but only if you're intentional.

In April 2026 I picked up OpenAI Codex, my sixth platform. In 36 days I ran through the same stages that took me years across the other five. Day one: search assistant. Week two: implementation partner. Week five: system audits and architecture calls. Thirty-six days to replay what took years, because the pattern was already installed. Once you've built the instincts — how AI thinks, where it breaks, what it's good and bad at — you run that pattern inside any new tool at close to twenty times the speed.

But it only compounds if the learning was intentional. I spent an entire phase using AI as a ghost-writer, drafting and redrafting for months. The real value was never in the drafting. It was in analysis, discovery, and building things that didn't exist before — and that realization is what tipped me from assistant to operator. Use AI casually, never asking why one approach worked and another didn't, and nothing compounds. You just accumulate hours without skill.

Remember the typo I opened with? From that to shipping production voice agents, maintaining my own AI stack, and having led AI education inside PwC's Salesforce practice. That trajectory isn't normal. But it's repeatable — if you invest in the learning until the learning starts investing in you.

Find yourself on the curve.

• Just using one tool. An Einstein license, maybe Agentforce. It feels like magic, or like nothing happened.
• Using it as an assistant. Daily writing, editing, research. You measure the value in hours saved.
• Starting to orchestrate. Accounts on several platforms — this one for writing, that one for code review.
• Building systems. Agents, automations, feedback loops that run without you typing a prompt. You measure the value in systems built.
• Running operations. AI isn't a tool you open. It's infrastructure. You'd notice within an hour if it went down.

The biggest gap is between assistant and building systems. These ten rules are how you cross it — not by finding a better tool, but by building better judgment about the tools you already have.