The Killer Part: Scoring 7 Pieces of the Unbundled GTM Stack

There's a genre of LinkedIn post doing numbers right now. The format: an "old stack" column — Apollo, a major B2B data suite, Sales Navigator, Apify, a spreadsheet — next to a "new stack" column that's Claude Code plus four or five point APIs nobody had heard of eighteen months ago. The caption says some version of "we used to pay $25,000 a month for this." The comments are engineers trading tool names like baseball cards.

The post is directionally right, even where the headline number is theatrical (the median contract for a major bundled B2B data suite is about $32k a year, per Vendr's purchase data — $25k a month is the top of the multi-seat-plus-intent range). When an agent can orchestrate scraping, enrichment, and outreach from a terminal, the bundled seat-license suite stops making sense; you rent the pieces at API prices. And those prices are collapsing fast: flat-rate "unlimited" B2B data APIs now start at $97/month (MoltSets — whose founder publicly abandoned a $495/mo premium tier for a volume play) and $399/month (BlitzAPI), waterfall enrichment runs $0.06–0.10 per attempt, and email verification bottoms out at $0.00055 per check.

But the interesting question isn't whether the suite unbundles — it's which piece of the unbundled stack is worth owning. Every unbundling in software history left one component that captured most of the value while the rest raced to zero margin. So: in a GTM stack — or any stack that runs on messy company and people data — what's the killer part?

We had a specific reason to ask this week. Two of us have been circling the same decision from opposite ends: one pass took the product vision and audited it against every repo we run (more on that below), and this pass takes the market side. The rule was that both had to show their work.

The seven candidates

Each of these is a component you could build as a standalone product for teams assembling their own data stack. Each is something we run internally today in some form — which matters, because one of our scoring dimensions is how much new tech an MVP actually needs.

1. Entity resolution as a service. "Are these two rows the same company?" Upload a messy CSV, get every row matched, scored, and explained against a canonical record — or resolve a single domain through an alias graph with evidence. The problem every multi-source stack hits first and solves worst.
2. Waterfall enrichment orchestration. Chain enrichment providers in cost order, take the first good answer, handle retries and rate limits. This is the core of what Clay sells.
3. Freshness & data-rot monitoring. Watch a dataset over time: row counts, per-field fill rates, decayed records, dead domains, unexplained jumps. Alert before the send, not after the bounce report.
4. Deterministic scored lists. Pin a transparent multi-dimension scoring model to a list — weights printed, no LLM in the scoring path — so "who do I call first" is reproducible and arguable, not vibes.
5. Agent-native access (MCP for your data). The buyer of data is increasingly an agent, not an analyst. Serve lists, matching, and scores as tools an agent calls directly instead of a dashboard a human exports from.
6. Verification & liveness. Is this email deliverable, is this phone real, is this company even still operating. The insurance layer of the stack.
7. Provenance & audit for AI-derived fields. Every AI-written field carries which model version wrote it and when, with the full decision trace stored. The compliance layer that gets mandatory if AI-enriched data ever meets regulation.

The scoring model

Six dimensions, weights fixed before any candidate was scored, every cell 0–10. This is the same mechanism we ship as pinned scoring models on OnlyData lists — deterministic core, no model in the scoring path, so a disagreement is an argument about a number on the page rather than a shrug.

Why these six: Pain and Marketability decide whether anyone shows up; MVP proximity decides whether a two-person team can prove it this quarter; Moat and Scale ease decide whether it's a product or a consulting gig; Run cost decides whether pricing has room to breathe. Weights favor the first three because at our stage a thing nobody wants, perfectly architected, is still nothing.

The scores

Why each score is what it is

1 · Entity resolution — 7.30, still first

Pain 9: the moment a stack has two data sources, it has a duplicate problem, and the moment it has an agent writing rows, it has a worse one. Every point API returns its own IDs; merging Blitz + Prospeo + MoltSets + your CRM means fuzzy-matching companies and people yourself, and everyone building the "new stack" rediscovers this in week two, usually inside a spreadsheet. The price signal is weaker than we first published it: People Data Labs does charge a premium for identity resolution over enrichment, but against its own published self-serve enrichment rate (~$0.28 a record) that premium is roughly 2×, not the 14× this page originally claimed. See the correction below. MVP 8: this is the part where we're mostly grading existing code — a candidate-accumulation matcher (domain, phone, name+address, fuzzy name with city boost) runs in production today, a resolve API passed 29/29 test cases through an alias graph, and dedup runs with a full reversible merge audit. Marketability 8: "upload 50 messy rows, watch 50 come back matched, scored, and explained" is a demo that sells itself in ninety seconds — and the top comment under the viral old-stack/new-stack post was someone asking about a decision-model startup whose flagship showcased use case is exactly this pairwise matching problem (one we've already benchmarked on 2,000 real decisions and wired into our dedup as an adjudicator). Moat 4 (revised down from 7): the compounding story — every resolved alias makes the graph better — is a target, not a property we have today. It only compounds if customer-supplied aliases can be reused across tenants and held-out match quality measurably improves per batch, and right now uploads gap-fill the shared pool with no isolation, provenance or consent model. 4 is current defensibility; 7 is what it becomes after that work lands. Scale 6 is the honest penalty: our matcher currently lives inside a very large server file and extraction is real work; multi-tenant isolation is designed but not shipped. Cost 8: the core is deterministic string-and-graph work with a small embedding model as adjudicator — cents, not tokens.

2 · Scored lists — 6.75

High MVP proximity (we shipped a pinned-model list publicly this month) and near-zero run cost, because there's deliberately no LLM in the scoring path. The market has a real hole here: predictive scoring is owned by MadKudu at roughly $24–100k/yr — priced for enterprises — while below that, teams score with ad-hoc LLM prompts that are non-deterministic and unauditable. What holds it back: the pain is chronic rather than acute — teams limp along ranking by gut — and willingness to pay at the low end is unproven. Best understood as the upsell you build on top of the killer part.

3 · Freshness / data rot — 6.40

B2B data rots at roughly 22–40% a year, everyone quotes the stat, and the only real remedies today are features locked inside the big suites. Our version already runs: daily snapshots, per-field fill-rate deltas, and anomaly flags on our own datasets, all cheap and stateless. Two things keep it out of first place: it's a smoke detector (the buyer notices the need the day after the fire), and the flat-rate data vendors are already bundling "our records are <30 days old" as a selling point. Pairs naturally with entity resolution — "we keep your records matched and alive" — as the recurring-revenue layer rather than standing alone.

4 · Agent-native access — 5.75

This score dropped two points during research, which is what research is for. The long-game thesis is intact — the buyer of data is increasingly an agent, we serve an 84-tool MCP surface, and we publish an agent-readiness score for everyone else's. But the window for a neutral agent-access product closed this year: Apollo, Clay, Explorium, Lusha, and the big bundled suites all shipped first-party MCP servers in 2026, and a neutral aggregator is a thin wrapper the vendors disintermediate at will. Agent-native access survives as the delivery channel for a data product you own — not as the product.

5 · Provenance / audit — 5.15

We believe in this one and the market, so far, has not agreed. We run model-version registries and full decision traces internally on every enrichment pipeline — and our public experiment in selling exactly this has sat untouched since April with zero external adopters. Pain 3 and Marketability 3 are what the evidence says, not what the thesis says. It becomes a 8+ the day a regulator makes it mandatory; we're not scoring the day after.

6 · Waterfall enrichment — 4.90

Real pain, terrible position. Clay owns the workbench at $185–495/mo self-serve, pay-per-success competitors undercut on price ($0.058–0.10 per attempt), flat-rate unlimited APIs undercut the whole category from below, and the "new stack" post that prompted this page is itself evidence that agents are commoditizing orchestration. Building this in 2026 means fighting Clay and Claude Code at the same time. Pass.

7 · Verification / liveness — 4.60

Email verification is a solved, brutally commoditized market: a 15× price spread from $0.008 down to $0.00055 per check that reviewers attribute to brand, not technology, with entry stakes near zero. The one under-served corner is company-liveness — is this business still operating — which is genuinely hard and genuinely ours to do (we crawl and score company surfaces already), but as a wedge it's a feature of freshness monitoring, not a company.

The verdict

The killer part of the unbundled GTM stack is entity resolution — sold as "upload messy, get canonical," kept honest by a freshness layer, and delivered where the agents are: through MCP, not a dashboard.

The top of the table isn't just a winner, it's a shape. Entity resolution is the wedge — the acute pain, the ninety-second demo, and a primitive vendors still price at a premium to plain enrichment while everything around it falls toward $97/month flat. Freshness monitoring is the retention layer that makes it a subscription instead of a one-shot cleanup. Agent-native delivery is the channel, because the customer's "user" is increasingly an agent. And the runner-up, scored lists, is what you build on top of clean, resolved, fresh records — downstream of the killer part, which is exactly why it can't be the wedge.

How we used our own machinery to make this page (and where we didn't)

This section is the part we most wanted on the record, so it's held to the same honesty rule as our eval posts: claims about ourselves link to things that exist, and the failures stay in.

• The scoring mechanism is our shipping product, reused. Fixed dimensions, pinned weights, deterministic totals, printed on the page — this is the same scored-list machinery behind inventory-need-boise, applied to our own strategy. No model chose these numbers; two humans and their agents argued them, and the model just makes the argument reproducible.
• The candidate list came from auditing our own repos, agentically. The same day this page shipped, a separate pass ran a multi-agent survey across the seven repos we operate — capabilities, data model, tenancy, ops — then had four candidate architectures judged from three perspectives (founders, security, first customer). Its conclusion, reached from the code side before this market-side scoring existed: the strongest thing we own is hosted entity matching. We originally called that convergence the strongest signal on this page. It isn't — both passes ran over the same repos, the same team and the same day's context, and neither tested a paying buyer. That same code-side pass scored entity matching 3/10 on "could a stranger get value today." Treat it as a consistent technical hypothesis with customer demand still untested.
• MVP-proximity scores are graded against running code, not ambition. The matcher, the 29/29 resolve API, the reversible merge audit, the daily dataset snapshots and anomaly flags, the 84-tool MCP server, the pinned scored list — all live today. That's why entity resolution's MVP score is 8 and not a hopeful 10: the honest blockers (the matcher needs extracting from a monolith; uploads need tenant isolation before a stranger touches them) are known, listed, and in the two-week plan.
• The eval culture set the scores it didn't flatter. Provenance scored 5.15 because our own attempt to sell it has no external users — the internal audit that surfaced "no commits since April, nobody posts to it" is the same discipline that produced our Jev benchmark and its published failures. Dedup's credibility here rests on a measured result: an embedding adjudicator that vetoed 56 wrong merges the fuzzy rules wanted.
• This page was produced by the two-agent loop we've written about. The ask arrived as a phone message; an agent on a Mac Studio pulled the source conversation and commits, dispatched a market-research subagent, scored, wrote, shipped, and verified this page, then handed it to a second, independent agent — Instinct — for adversarial review. If Instinct's review moves a score, the page gets a dated correction, not a silent edit.
• Where we didn't use ourselves: the market research (pricing anchors, competitor states) came from ordinary web research by a subagent, not from OnlyData data — our datasets don't cover SaaS pricing. Caveats travel with it: vendor-page prices are firm; the bundled incumbent suites, Bombora, and MadKudu are quote-based, so those anchors are reseller and reviewer estimates. And no LLM sat in the scoring path, by design.