Your Help Center Has 340 Articles and Six Were Updated This Quarter

It is Monday, 9:14 AM. Your Head of Support opens the Zendesk knowledge base before the week-start sync. 340 articles live. The newest reads 11 days old, a billing-portal walkthrough the senior rep rewrote after the Stripe migration. The next most recent reads 94 days. The one after that reads 141 days. Six articles updated this quarter. 94 flagged "contradicted by ticket resolution" by the deflection report nobody opens.

She opens the ticket queue. 187 open tickets. 71 of them are repeat questions the help center already covers. 42 of those 71 cite an article in the auto-reply, and the customer still replied "this does not match what I see." The reps answer the ticket from scratch each time. The deflection rate on the Monday report reads 14 percent against an industry median closer to 32 percent for a company this size. The help-center search log shows 2,400 queries last week with 940 ending in a support-ticket submit button click.

Pull the quarter. 11,400 tickets handled across 14 reps. 4,100 of them repeat questions tied to a stale or missing article. 320 tickets escalated to engineering because the article the rep sent contradicts the current product. The CS lead flagged four churned accounts in the Q3 review that cited "docs out of date" in the exit survey. The function nobody staffed is costing the business 4,100 avoidable tickets a quarter, 320 engineering escalations, and four accounts the exit survey tagged on documentation alone.

Knowledge-base maintenance is a function. Most Series B and C support teams staffed it with a lead rep who writes the article after the ticket, a knowledge manager who owns the taxonomy, a technical writer who ships the launch pages, and a Head of Support who runs the Thursday review. The function lives in the gap between the reps who close the ticket, the writer who owns the voice, the PM who ships the feature, and the customer who hits the search bar. On the org chart it reads Support Ops. In practice it reads a Zendesk Guide folder with a last-edited column nobody sorts.

The 94-article decay math

Pull every ticket closed in the last 90 days. Tag each one with the help-center article the macro cited, the resolution the rep wrote, and the delta between the two. Count articles cited where the resolution reads "ignore the doc, do X instead." Count articles cited where the customer replied "this is not what I see." Count articles never cited where the ticket volume says they should have been. Most Series B support teams past 10,000 quarterly tickets find 25 to 40 percent of KB citations contradicted by the resolution, 30 to 50 percent of articles untouched past 180 days, and 40 to 60 percent of top-50 ticket topics with no article at all.

Walk one article. The "reset two-factor authentication" article was written in March 2025 against the pre-migration auth stack. The product shipped a new 2FA flow in July 2026. The rep closes 11 2FA tickets a week with a Slack-pinned screenshot the senior rep drew in Figma. The article still ranks first on the help-center search for "2fa reset." Customers read it, try the old flow, hit a dead screen, open a ticket. The ticket closes with the Figma screenshot. The article sits.

The team that should own this knows it is broken. The lead rep writes "we should fix the 2fa article" in the retro every month. The knowledge manager is on parental leave and the backfill owns three other queues. The technical writer ships launch pages for product marketing and clears a backlog of nine. The Head of Support runs the Thursday review and reports deflection rate in the Monday staff meeting. The 94-article backlog is the question nobody charged against a cadence.

Why Zendesk Guide and Intercom Articles do not answer a question

You bought Zendesk Guide at $84 to $115 per agent per month or Intercom Articles at $99 to $139 per seat. You bought Scribe or Guidde at $29 to $89 a seat for the walkthrough capture. You bought Algolia or the native search at $0 to $14K a year for the search layer. Zendesk renders the article the writer saves. Scribe captures the clicks the writer records. Algolia ranks the article against the query. None of them read Monday's ticket log, score the 71 repeat questions against the current article set, flag the 42 articles the resolution contradicted, draft the rewrite in the brand voice, and ship it to the writer for a 10-minute review.

Zendesk reports the "2fa reset" article was viewed 1,140 times last month with a 14 percent helpful-rating score. It does not know the article contradicts the July auth release. Intercom reports the article resolution rate at 22 percent. It does not score whether the resolution the rep wrote in the ticket renders the article wrong. Algolia ranks the article first on the query. It does not flag that the first result lands the customer on a dead screen. The CMS layer is a CMS layer. The search layer is a search layer. Neither is a function.

What a fractional AI knowledge-base function does

Hand the Zendesk ticket history, the Intercom conversation log, the help-center article set, the product release notes, the engineering changelog, the Slack retro threads, the search-query log, the Figma screenshots the reps pin, and the brand style guide to a fractional AI agent. The agent does the work a knowledge manager, a lead rep, and a technical writer would do together. The cadence is per-ticket on contradiction detection, per-release on article refresh, per-week on the gap digest, and per-quarter on the taxonomy sweep.

Every ticket resolution scored against the cited article. The rep closes the 2FA ticket Tuesday at 11:04 AM with the Figma screenshot. By 11:20 the agent has flagged the delta between the article and the resolution, drafted the rewrite, pulled the current screenshot from the product, and routed the draft to the lead rep for a 10-minute review. The article updates by lunch. The next 2FA ticket renders a deflection.

Every product release triggers an article audit inside two hours. Engineering ships the new 2FA flow Thursday at 2:14 PM. By 4:00 the agent has read the release notes, scored the 11 articles that reference the auth stack, drafted rewrites against the five that broke, flagged the two that need a new screenshot, and shipped the drafts to the writer. The articles land live before the Friday ticket spike hits.

Every search query ending in a ticket gets logged against the gap list. The Monday search log reads 940 queries that ended in a ticket submit. The agent clusters them into 14 topics, scores each cluster against the existing article set, flags the four topics with zero coverage, drafts the article for each, and routes the drafts to the lead rep. Four new articles ship by Friday and 320 tickets a month route to the help center instead of the queue.

Every stale article gets pulled or merged on a cadence. The quarterly sweep reads the 340 articles against ticket citation volume, product release history, and search ranking. 47 articles pull, 22 merge, 11 split. The help center drops to 282 articles with a maintained freshness score and the search ranking improves on the retained set.

The unit economics of a 4,100-ticket repeat queue

A Series B at $14M ARR running 11,400 quarterly tickets with 4,100 repeat questions is burning three lines. The 14 reps clear 22 to 34 hours a week on repeat questions a maintained help center would deflect, against a loaded hour of $38 to $62. That is $20K to $34K a month of frontline time on answers the help center should carry. The reps get nine to fourteen hours a week back on the complex tickets that compound CSAT.

The engineering line is the second. Pulling the 320 quarterly engineering escalations tied to stale docs saves 80 to 140 engineer hours a quarter against a loaded hour of $140 to $220. That is $12K to $30K a quarter of senior engineering time no longer burned on docs the writer should own. The release train stops colliding with a doc backlog.

The churn line is the third. The four Q3 churned accounts that cited "docs out of date" ran $340K to $820K in annual contract value between them. Catching the article decay before the exit survey writes it down is the single lever on the renewal line the support function controls without a product change. On a full year of a cleaned help center, the retention swing lands in the mid six figures for a company this size.

A 14-day sprint to stand up the agent runs in the low to mid five figures. Ongoing cost lands at $4K to $8K a month on API spend, Zendesk read access, and tooling plus a fractional support operator at $5K to $9K a month who owns the Thursday review and the release-day audit. Ticket-to-article contradiction detection ships in week one. Release audits and the gap digest ship in week two. The 94-article backlog clears inside the sprint.

What changes after the sprint

Picture the same Monday, 9:14 AM moment, thirty days after the sprint ships. Your Head of Support opens the help center. 282 articles live. The newest reads 2 days old, a billing-portal rewrite the agent drafted off a Thursday ticket cluster. 94 articles updated this month. The ticket queue reads 112 open. The repeat-question ratio reads 18 percent, down from 38. The engineering escalation count on the Monday report reads one.

By Thursday the Head of Support reads a gap digest that names the five topics with no article, the two articles contradicted by the week's resolutions, and the writer who owns the next step. The CFO reads a weekly deflection digest that names the deflection rate climbing from 14 to 29 percent. The reps run the complex-ticket queue instead of answering the same 2FA question 11 times a week. The release engineer stops writing an apology to the writer every Thursday afternoon.

If your help center currently reads 340 articles with six updated this quarter and 94 flagged outdated, the version where every ticket resolution refreshes the article it cited and every release triggers an audit inside two hours is fourteen days away. Knowledge-base maintenance is a function. You can hire a dedicated knowledge manager, you can expand the fractional support retainer, or you can scope a sprint and have it running this month. The work is the same. The backlog is not.