The software America runs on is becoming headless

👋 Hey, this is Artem and this is The Blueprint. At WIQ, we spend each day forward deployed at Fortune 500 companies working through the hairy journey of AI transformation.

Salesforce can become less visible without becoming less important.

That’s my read of its Claudeforce launch: sellers would work in Claude, while Salesforce still holds the customer history and checks permissions. I wouldn’t use time spent in the CRM as a measure of how much the business needs it.

I want a team’s instructions for reviewing a deal to remain useful when it switches agents, including the corrections sellers make along the way. Choosing another tool shouldn’t mean rediscovering how the job works.

At Bardeen, we always believed we had to help users discover what was worth automating. Initially, we tied that discovery to our own automation platform. In retrospect, that made less and less sense: customers already had platforms they had invested in, and the AI labs were giving them more ways to run agents.

The work we did to understand a job could be useful inside those other tools. We wanted to turn that understanding into instructions a customer could use in the software they already had.

Atlassian reports more than five million tool calls per working day through the connection it gives AI agents, with nearly a third writing back into its systems. Those writes include creating Jira work items and updating statuses. I expect more enterprise software to work this way, with agents operating it through other companies' AI interfaces.

I see the same separation in Salesforce's announcement. The place employees ask for work can change while the company's records and its knowledge of how to do the job remain useful. Salesforce already holds account history and enforces configured rules; the reasoning a seller uses to interpret a customer's answer may still be in the seller's head.

Salesforce's August 26 announcement introduced 37 prebuilt sales skills, which give Claude instructions for jobs such as meeting prep and pipeline review. The first product was in select pilots at launch.

I’d want to see the cases where the agent asks the seller for help.

Suppose a customer tells a seller they might delay a purchase. One agent adds a warning to the opportunity notes, while another moves the close date into the following quarter. If neither action violates a configured rule, Salesforce can accept either, even though the two updates give the sales team different forecasts.

That is why I wouldn't hand this decision to whichever agent happens to be running the review. I'd ask the person responsible for the forecast what evidence justifies changing the date: does a customer mentioning a possible delay warrant it, or does the account owner need to confirm? If the team hasn't agreed, I'd have the agent flag the risk and leave the date alone until the owner decides.

Having access to all the relevant records still leaves the agent with a judgment to make. I'd put the evidence requirement and escalation rule in its instructions, and keep any existing approval control enforced in Salesforce. Adding an AI interface shouldn't require the company to give up a rule it already relies on.

I'd test the instructions on someone who hasn't already learned the job.

A revenue lead at a publicly traded lender told us he was trying to get roughly 300 people working with AI. His first idea had been to put everyone on GitHub so they could pull down the skills they needed. That wasn't working for the whole team, and people were using different AI tools.

He said they were “literally sending skills around as zip files.” The problem he wanted help with was managing and distributing the instructions people were already using, so the team could keep them current.

A library like Salesforce's gives a team a useful starting point. Once people adapt a skill to their accounts, add an exception, or correct a mistake, the company has to decide which changes belong in the procedure everyone follows.

Standardizing on one AI tool would make distribution easier, as he pointed out. It would still leave the team with revisions to manage: if one rep fixes a mistake in their copy of the pipeline-review instructions, the other reps' agents can go on repeating it. I want the next person running that review to benefit from the correction, including someone working in a different approved agent.

I'd make sales operations own the pipeline-review procedure, with IT responsible for the approved connections and permissions. A seller's correction would go back to that owner for review. I'd configure each approved agent setup to retrieve the current procedure when it starts the job, so a reviewed change reaches the next run without waiting for everyone to replace a downloaded copy.

I'd also record which version the agent used, alongside the evidence behind its recommendation and any approval it received. If a deal moves unexpectedly, I want to know whether the agent followed an outdated rule or made a bad decision with the current one. Those problems require different fixes, and a log that only shows the final CRM update won't tell the team which to make.

Each agent setup would still need testing against recent cases the sales team has reviewed. I'd include both a case where the agent should flag a possible delay and leave the date unchanged, and one where the account owner has approved the new date and the update should go through. An agent that leaves every deal untouched can pass a test for unauthorized changes while failing to help the seller update the pipeline.