Software for agents

Most “AI” you’ve seen in 2025 was an assistant bolted onto a CRM built in 2014. The CRM won.

That sounds harsh, so let me soften it: those CRMs were brilliant. The contact record, the deal stage, the pipeline view, the activity log — every screen designed for a human rep with a coffee, a cursor, and the patience to open a contact, scroll to the right section, and update twelve fields before lunch. They served their decade beautifully. The problem isn't that they were bad. The problem is who's filling them in now.

Your new coworker reads a contact record in milliseconds and writes in tokens. She doesn't need a tooltip. She doesn't scroll through tabs. She doesn't care that the “Save” button is top-right because she has no eyes. Drop her into a CRM built for human eyeballs and she spends ninety percent of her day translating between what she wants to do and what the contact form will accept.

CRM forms were built for humans. Now humans are training their replacements to fill them out.

That's the trade nobody put a price tag on. Every “AI feature” announced this year is, underneath, an agent paying a tax to use software that was never designed for her. The tax is paid in tokens, in turns, in latency, and in the small humiliation of an LLM trying to figure out which dropdown maps to “lifecycle stage” — or whether the phone number lives on the contact record or the company record.

You can patch it. You can wrapper it. You can give her a copilot that watches her navigate the CRM and offers helpful suggestions, like a parent watching a teenager assemble IKEA furniture. The software still wasn't built for her. The patches still cost a turn each. And the contact record still has forty-three fields she'll never touch.

There's a tell when an industry is about to flip. It's when the new thing tries to live as a feature inside the old thing — and the old thing's bones start cracking under the weight. Browsers cracked first. The bet was that an agent useful enough to actually book your flight couldn't live inside a browser designed for human scrolling. So someone built a browser for agents. The reviews were quiet. The implication wasn't.

The category isn't a smarter CRM. The category is a CRM built for the people who never open it.

CRMs are next. Not because the existing ones did anything wrong — they did something extraordinary, for a buyer that no longer exists. The buyer of business software in 2026 isn't a sales rep with thirty open tabs. It's a sales rep and the workforce they manage, and the workforce reads contact records, writes deal notes, and updates pipeline stages differently than any user the old software was designed around. The form-shaped data entry, the keyword search that returns “page 1 of 5,” the audit log nobody reads, the field validation that holds up for one human but not for thirty agents running parallel enrichment — these aren't bugs. They're correct, for the original user. They're catastrophic, for the new one.

What does software-for-agents look like on the inside? Different in ways the human user never sees. Contact records have memory — the agent doesn't re-ask what it already knows. Search returns a fact, not a paginated list. Every change to a deal stage is an event the next agent can subscribe to. Coordination between agents doesn't leak into Slack at 2 a.m., because the coordination is built into the substrate. The whole thing assumes the next person touching a record might not have hands.

If your AI is great and your CRM is mid, you don't have a great AI. You have an expensive translation layer.

The teams that get this early get a compounding gift. The substrate is the slow part. The model swap is the fast part. When the next model lands — and one is always landing — the team running on agent-native software absorbs the upgrade in an afternoon. The team running on bolted-on AI ports their patches forward, again, for the fifth time this quarter. One of those teams is shipping next week's feature. The other is shipping next quarter's compatibility fix.

What that compounding gift looks like in practice is mundane. A workforce needs to reason about relationships — who this person is, what's happened between us, what should happen next — so the substrate gives it structured memory underneath every contact, every deal, every conversation. It connects to whatever CRM the human team already runs. Salesforce, HubSpot, anything with an API. The humans can keep living there. But something quiet happens over the months. They stop opening it to decide what to do next.

Data entry was the tax on doing business. We just stopped charging it.

Every call captured. Every meeting logged. Every signal turned into structured knowledge — without a single field filled in by hand. The humans kept the judgment calls. The workforce kept everything else. The CRM didn't get replaced. It just stopped being where the work happens.

The CRM was always trying to be a record of the conversation. We finally gave the conversation somewhere better to live.