Who Owns What Your Company's AI Agents Have Learned?

The AI agents your employees use are learning how your business works. Right now, what they learn is stored in employees' personal accounts and in AI vendors' systems, and your company can't easily get it back out. A sales rep's agent learns which objections stall a deal. A recruiter's agent learns what a hiring manager rejects. If that employee leaves or the company switches AI vendors, the knowledge goes with them. It is also what makes the agents useful to your teams, because an agent that knows how your business works gives better answers than one that doesn't. Make sure to protect that.

A recent video from Nate B Jones, who makes fantastic YouTube videos about AI at work, is a good entrée into the issue. He explains what happens to business software when employees hand routine work to agents, and his first example is DoorDash. On September 30, the company launched an agent that lets U.S. customers order by text message through Apple Messages. The same day, it opened a waitlist for a connector built on MCP, a common standard that lets an AI agent use another company's service directly. Companies can use the connector to have their own agents place office food orders and supply restocks. DoorDash gets the order even if nobody opens its app.

Nate's point, which we share, is that business software is going to go the same way. An employee will ask an agent to pull a customer's history or log a new deal, and the agent will do both inside the CRM without anyone opening that CRM. It still stores the customer records and still runs the process that logs the deal. People just stop using its screens.

A vendor switch would erase one recruiter's weeks of work

Nate tells a true story about a recruiter who used to open a recruiting app every week, search for candidates, and write a shortlist for the hiring manager. An agent took over that routine. She taught it how she screens candidates and what each hiring manager rejects. She also fed the agent the people she hired, so it could spot what they had in common. Her agent still pulled candidate information, such as work histories and job titles, from a data provider like Crustdata, so that company kept getting paid. The recruiting app she once opened every week now sits unused.

Then leadership asked everyone to move to a different AI vendor. These requests often start as a cost conversation. Finance sees the monthly bill, and the recruiter sees weeks of rework on live candidates. Her manager hadn't known how much of her process lived in the agent. Nate expects this fight to reach many teams in 2027, and we think he's right. Finance will compare what each AI vendor charges, the people doing the work will be thinking about the weeks it takes to retrain a new agent, and most companies haven't decided which of those should drive the decision.

Gartner says up to $234 billion of software spending is at risk

In a July 1 press release, Gartner estimated that up to $234 billion of enterprise software spending is at risk by 2030, about 20% of what companies spend on subscription software (SaaS). The risk comes from agents doing work across several products, so fewer employees open each one. Gartner says this breaks the link between a vendor's number of users and its revenue, which is a problem for any vendor that charges per user.

Gartner also tells vendors to keep customer-specific knowledge in addition to customer data. Gartner doesn't define the term, so here is our reading: it means what a vendor learns about how each customer works. The reason is leverage. A vendor that knows how your team qualifies deals and answers pricing objections is hard to leave, because leaving means losing that. The recruiter's screening judgment is the same kind of knowledge, and today it sits in her agent's memory. We'd make sure your company holds a copy, in records you own, before any vendor does.

Large vendors are already letting agents work directly inside their products. On August 26, Salesforce and Anthropic announced Claudeforce, a partnership that lets Anthropic's Claude agent read and act on a company's Salesforce data. Its first product comes with 37 prebuilt sales skills, which are packaged instructions an agent follows for jobs like meeting prep and pipeline reviews. Those skills cover what is stored in Salesforce. Much of what makes a good rep good lives in call recordings and email threads, and an agent that only sees Salesforce data won't capture it.

Nate's advice to employees is to talk openly about how they use AI. We agree, and we'd go one step further. A conversation only covers what people remember to say. Step 3 below is writing the important knowledge down in a place your company owns.

Four steps for the next 90 days

1. Find out what your people actually use. Ask each team lead which agents their people use, what standing instructions they've given those agents (for example, “before every sales call, summarize the account's open deals and recent emails”), and where each agent stores what it has learned, whether that's a personal account or a vendor's cloud. Make it clear nobody is in trouble. In an April 2026 PagerDuty survey of 1,250 office workers at companies with at least $500 million in revenue, a third of those who use AI for work said they would hide it from their managers. If people expect a lecture, you won't get a straight answer.

2. Pick the knowledge you'd regret losing. Start with a revenue team, because the gains show up in numbers like win rate and days to close a deal, which your CFO already tracks. For each person, ask what the team would lose if they left tomorrow. For a sales team, the answers are usually how the best reps decide whether a deal is worth pursuing and how they answer pricing objections.

3. Put it in a context layer you own. A context layer is a set of company-owned records, such as call transcripts, CRM history, deal terms, and approved messaging, stored so that any AI model can read them over MCP. Have your best reps write down how they decide whether a deal is worth pursuing, including the questions they ask on a first call and the answers that make them walk away. Write it once, and every agent reads from it. Switching AI vendors then means pointing the new model at the same layer, and a new rep's agent starts with what the last rep's agent learned.

4. Settle the cost argument before it starts. Bring finance and the CTO in now. Agree on what you'll measure, such as hours saved on deal prep or days to first reply on an inbound lead, and on which budget pays for the AI subscriptions. Nobody had agreed on any of this before the recruiter had to defend her setup.

Two mistakes to avoid: banning tools and depending on one vendor

Banning unapproved tools. A TrustedTech report of UK and US employees found that nearly a third would keep using AI tools their employer hadn't approved, even if rules prohibited it and discipline was possible. A ban moves that use out of your sight. The same report found that 65% of senior decision-makers use unapproved tools, against 31% of employees below that level, so start by asking what your own leadership team uses.

Letting one vendor hold what your team taught its agent. If everything your team has taught its agent lives inside one provider's account, that provider sets your terms at renewal. Ask every vendor what happens to your data and settings if you leave, and what your bill looks like if you use fewer seats. When agents can connect to a vendor's product, the vendor has only earned a place in the conversation. It still has to show why you should keep paying.

Three predictions for the next two years

The timing is the part we're least sure about.

Your software vendors will change how they charge first. Within 18 months, we expect many of your software vendors to propose charging by the amount of work their software does or the results it produces. Gartner's finding on per-user pricing gives them a reason to start that conversation.

Boards will ask where your AI know-how lives. Within a year or two, we expect boards to ask a new question: if your main AI vendor changed tomorrow, what would your teams lose? You'll want the answer ready.

Your job will shift from picking AI tools to owning two things. The first is the company's written knowledge that agents draw on. The second is the rules for what agents may do with company data. Tools will keep changing every few months, and those two stay with you.

If you'd like a quick read on how much of your team's know-how sits outside your own systems, take our Context Moat Score. If you'd rather work through the first 90 days with someone, reach us at rampwell.ai.