Why CRM deployments fail

The industry stat everyone quotes is “55% of CRM deployments fail.” It is true enough to repeat, and it is also the most useless number in the conversation, because the failures it describes are heterogeneous. A deployment that never got past the procurement stage is in the 55%. So is a deployment that runs fine for two years before quietly being replaced by a spreadsheet on someone’s laptop. So is a rollout that hits 30% adoption, plateaus, and stays there for the lifetime of the contract. Lumping them together produces a headline statistic that explains nothing about which CRM your team is going to lose, or why, or how to notice it before the renewal.

The CRM Graveyard campaign sits on a different question. Since February of this year, the /crm-graveyard landing page has been running a free interactive diagnostic — five guided questions that ask a team’s IT lead, sales operator, or revenue owner what is actually broken about the CRM they have today. The diagnostic terminates in one of three named archetypes: the bloated corpse (CRM is too complex for the team to actually use), the island (CRM works, but lives disconnected from the rest of the business), or vendor abandonment (the implementation partner closed the ticket and went silent). It is not a benchmark survey. It is an instrument.

Over ninety days, one thousand and five self-selecting sessions ran the instrument end-to-end. Forty-four finished a full assessment and got a typed diagnosis back. This article describes what those forty-four diagnoses — and the nine hundred and sixty-one shadow sessions that surrounded them — look like as a dataset. It is not the failure rate. It is the failure shape.

The data are funnel events from the production /crm-graveyard campaign, written to a MongoDB collection named CrmGraveyardFunnelEvent by the POST /api/crm-graveyard/funnel endpoint. Each event carries a session identifier, an event type, a UTC timestamp, and (for completion events) an optional archetype and severity. Session identifiers are stored in the browser’s sessionStorage for the /crm-graveyard route only and do not overlap with experiments on other pages.

The window is February 20, 2026 through May 21, 2026 — ninety calendar days. The aggregation is performed by a single re-runnable script, scripts/rfe-seo/crm-graveyard-extract.ts, which counts unique sessions reaching each funnel stage. Counting sessions rather than events deduplicates the user who refreshes the landing page or re-takes the assessment, and matches how the internal admin dashboard at /admin/crm-graveyard-funnel computes conversion rates. Sessions excluded by the IP-exclusion ruleset (internal traffic, known bots, the office VPN) are filtered out at write time and never enter the extract.

Before this article published we reconciled four ingestion anomalies surfaced during the first extract pass on May 21 (tracked internally as DS1–DS4). The assessment_diagnosed intermediate event fires for only 2 of 109 starts in this window and is excluded from the charted funnel. The pdf_download event fires once because the current diagnosis report renders inline rather than as a separate downloadable PDF; that count is reported as a diagnostic-only footnote and not as a funnel stage. The archetype-attribution sparsity (1 attributed complete out of 44) is reported as a finding in §4 rather than corrected with imputation. The time-to-complete aggregation now requires the landing-to-complete delta to be positive and at least 30 seconds, and we hold the headline number until the median is not dominated by returning-session re-entry events.

We do not claim this is a representative sample of all CRM-owning companies. The population is people who searched for symptoms of a failing CRM and ran a five-question diagnostic to get a typed answer. That is a useful sample, and an honest one. It is not the universe.

Figure 1 shows the full funnel. The shape is the finding.

The cliff is the hero CTA, not the assessment. Eight-seven point five-six percent of landings never click the “Run the diagnostic” button at all. They read the headline, skim the archetype copy, possibly scroll past one section, and leave. This is consistent across the window; the rate does not vary meaningfully week to week within the noise envelope.

Once a session crosses the CTA, the engagement curve flattens dramatically. Of the 125 sessions that click in, 109 (87.20%) start the first assessment question, and 44 (40.37% of starters) push through to the typed diagnosis. The classical product-instinct on a five-question form is “most starts will abandon mid-form” — that is not what we see. The population that clicks the CTA mostly intends to finish.

The practical read for a team running a similar diagnostic is that the leverage is on the landing page, not on the form. Cutting form abandonment in half on this funnel would move 44 completes to roughly 62 — a 40% relative gain on a stage that is already converting well. Cutting hero-CTA abandonment by even ten percentage points would move 1,005 landings into roughly 226 starts and, at today’s downstream rates, roughly 80 completes — an 81% relative gain on the headline outcome.

The funnel completes include the typed archetype the diagnostic returned to each respondent — that is the entire point of the assessment. In the database, however, the archetype label is currently attached to the completion event for only one of the 44 completes in the window. The other 43 record the completion event itself, plus the lead-form attribution if the respondent went on to ask for a conversation, but not the typed-archetype value the diagnostic returned to them in the UI.

The underlying cause is wiring rather than missing data: the diagnostic returns the typed archetype to the respondent in the rendered UI, but the analytics event that fires alongside the render path does not always carry the archetype payload. We tracked this internally as DS3, attempted a backfill against the lead-attribution table, and recovered exactly one additional attributed complete. That recovery is too sparse to publish a distribution. The honest read at n = 44 is “the instrument is qualifying people; we do not yet know which archetype they are coming in with.”

We will hold the archetype mix until the instrument records enough attributed completes to make a claim. The next time this article ships an update will be when the attributed-complete count crosses 200 — the threshold at which a four-way split (Bloated Corpse / Island / Vendor Abandonment / Other) has enough mass per bucket to publish without a sample-size caveat that eats the headline. Until then, the typed diagnoses are reaching respondents in the UI and the instrument is working as designed; the analytics pipeline catches up next.

Of the 44 sessions that finished the assessment, 13 submitted the lead form on the post-diagnosis page asking Weaver to review the diagnosis with a human. That is a 29.55% rate from typed diagnosis to sales conversation.

For comparison, a standard B2B product’s “Contact us / Request a demo” form converts in the 1 to 5 percent range against the same kind of organic traffic. The Weaver diagnostic is converting at roughly six to thirty times that rate from the equivalent surface — not because the form is doing anything clever, but because the five-question instrument that sits in front of the form has already done the qualification work. A respondent who finished the assessment has named what is broken about their CRM, received a typed answer, and has chosen to talk to a human about it. That is a different population from “a visitor who clicked the contact link in the footer.”

The practical read is that the instrument is the qualifier — not a layer in front of one. The lead-form submit is a downstream artifact of a respondent having self-administered a meaningful diagnostic, not a separate qualification step the sales motion has to do. Teams running an outbound motion against a population that has already self-diagnosed start a conversation roughly five to thirty times more efficiently than teams converting cold traffic on an undifferentiated contact form.

Three honest constraints on what these numbers can be asked to do:

• It is not a population failure rate. The sample is people Googling symptoms of a failing CRM. By construction it over-indexes on companies whose CRM is broken right now. The Johnny Grow, Forrester, Gartner, and Forbes Advisor numbers cited at the top of this article are the industry baseline; this dataset complements them with a shape, not a rate.
• It is not yet enough to claim an archetype distribution. Forty-four completes with 1 attributed is too sparse to publish a four-way split. The pipeline fix is in front of the n threshold — we will not publish a distribution before either is met.
• It is not generalizable to non-diagnostic funnels. The 29.55% diagnostic-to-conversation rate is a fact about this instrument on this landing page; the underlying mechanism (a five-question diagnostic that returns a typed answer) is what produces the rate, and a generic contact form will not reproduce it.

A T1.3 follow-up scheduled for May 27 will extend this article with a “how to read your own funnel” section using the same script, so any team running a similar diagnostic can compute these numbers against their own collection.

Drafted with AI assistance (Anthropic Claude), reviewed and edited by Rafi Seddiqi on May 24, 2026. The underlying dataset and the aggregation script are original to Weaver. The funnel-chart visualization is an original component (src/components/charts/CrmFunnelChart.tsx), not a stock or third-party asset. The named industry benchmarks in §1 and §6 are externally sourced and cited at the bottom of this page. Methodology is reproducible: any reader with access to a similar funnel-event collection can re-run the extract script against their own data. See Weaver’s editorial policy for the full AI-usage and citation standards that govern this and every other editorial page on weaver.work.