Your $25 Leads Are 10x Cheaper. Are They 10x More Useless?

Intercom just answered a question most SaaS companies are still ducking. Their outcome-based pricing model charges per resolution, not per seat. When your AI agent handles half the support conversations, per-seat pricing punishes your best customers for adopting the product. Intercom reaccelerated growth after making the switch.

The concept is clean. The execution is harder. You need to define "resolved" in a way that's auditable, defensible, and not gameable. A chatbot that marks every ticket "resolved" after auto-responding hits the metric while delivering nothing. Any buyer with a pulse will notice.

Here's where I keep getting stuck, though. Intercom's model works because support resolutions are measurable. "Did the customer's problem go away?" has a verifiable answer. But what about the rest of us? In consulting, the resolution metric doesn't exist. "Better strategy" isn't a countable outcome. "Improved pipeline quality" takes quarters to prove. Outcome-based pricing assumes the outcome is legible, and for most knowledge work, it isn't.

That's the real question this model surfaces. Every company selling AI-augmented products needs to figure out what their "resolved ticket" equivalent is. Most don't have one yet, which is why they're still charging per seat for work an agent handles. Intercom built the template. The hard part for everyone else is finding their own outcome metric that's honest enough to charge against.

Teams are still debating Claude vs. GPT like it's a meaningful variable. Tomasz Tunguz's latest piece argues they're tuning the wrong dial. His thesis: the richest context wins, regardless of which model runs underneath.

I've been building context engineering systems for the past year. This newsletter is one of them. Specialized agents scan hundreds of sources, structured context flows through debate rounds, and the output is shaped by accumulated editorial memory. The prompt matters far less than the architecture feeding it. When you see "paste a prospect's LinkedIn into ChatGPT and ask for a personalized opener," that's prompting. Context engineering means building pipelines that assemble firmographic data, engagement history, and intent signals before the model ever sees the request.

Tunguz frames proprietary data as the moat. Your structured context becomes the thing no model upgrade can replicate. GPT-5 ships? Doesn't matter if your competitor's pipeline is five layers deeper.

But there's a competing argument worth holding alongside this one. Swyx's Latent Space analysis surfaces a genuine tension: Anthropic rewrites Claude Code's harness from scratch every three to four weeks. Their bet is that the model itself improves fast enough to make thick wrappers disposable. The "Big Model" camp says keep the harness thin, because anything you build around the model will be obsolete by next quarter. The "Big Harness" camp says the orchestration layer is where real value lives.

Both are partially right, and the answer depends on your time horizon. If you're building for this quarter, invest in context pipelines. The model is a commodity and your data assembly is the edge. If you're building for next year, hold the architecture loosely. The rep with twenty years of industry knowledge is running their own context engine. It's biological, it doesn't scale, and it walks out when they leave. But the models are getting better at replicating that judgment faster than most of us expected.

Bill Binch has been carrying a quota for 116 consecutive quarters. Twenty-nine years across market cycles, tech waves, and multiple "this changes everything" moments. His career is a geological record of what compounds in B2B sales.

The obvious read: AI replaces reps. The sharper question Binch's career answers: which reps were doing replaceable work? The ones doing outreach, data entry, and basic qualification are already being automated. Binch's work compounds in a different direction: reading a room, navigating a seven-figure deal through twelve stakeholders who each need a different version of the story, knowing when to push and when to listen. That's judgment. Judgment doesn't automate.

But there's an implication the "humans win" crowd glosses over. If AI handles the commodity work, you need fewer total reps. Not different work for the same number of people. Fewer people doing higher-value work. The math is straightforward: automate research, CRM hygiene, scheduling, and first-draft emails, and you've eliminated the majority of a junior rep's day.

Which raises the harder question. Junior reps historically developed taste and judgment by doing the grunt work. Cold calls taught them to read tone. Data entry forced them to learn the CRM. Basic qualification built their instinct for what a real deal looks like. If AI handles all of that, how do you build the next Bill Binch? The traditional ladder from SDR to AE to enterprise closer assumed you'd spend years accumulating pattern recognition through repetition. That ladder is compressing.

If you already have judgment, this is a great time to be alive. AI clears the busywork and lets you spend more time on the work that closes deals. If you haven't developed that judgment yet, the path to developing it just got less obvious. Sales leaders need to think about this now: how do you build reps with taste when the apprenticeship work is being automated away?