Edited isn’t the same as right
In the demo above, both articles look maintained. Only the tickets and the assistant’s ratings show which one still works.
For teams whose AI answers from their knowledge base
Your docs say one thing. Your records say another.
We show where your sources disagree, explain the evidence, and help the responsible owner resolve it. The demo below walks through one example: an internal AI assistant answering from an outdated article.
- Checks what people did, not just when the page changed
- Evidence on every result
- You decide what changes
Demo
Watch a knowledge check
- Articles, tickets and assistant logs, read-only
- Check against how the work is done
- Evidence for every call
- Owner decision and dated record
Internal AI assistants answer employees from the company knowledge base. Here, a sample company’s assistant keeps answering a VPN question from an article that no longer matches how access works.
Two articles explain VPN access with different steps. Ecdotic checks both against the service desk tickets and the ratings employees gave the assistant’s answers. It shows which version is current and why, and the article’s owner approves the correction. The decision is kept as a dated record, and the assistant answers from the current article in the replay. Today, the team makes the article change itself.
Read the replay as text
- Ask. An employee asks Northwind’s internal AI assistant how to get VPN access. It answers from KB0142, written in 2023: email the help desk for a token. The employee rates the answer unhelpful.
- Sources. Ecdotic reads exported articles, 90 days of service desk tickets and the assistant’s logs. Two articles answer the same question with different steps.
- Check. Ecdotic links every ticket and assistant answer to the article it used. All 5 tickets that followed KB0142 were reopened; 0 of the 41 that followed KB0217 were. 38% of the assistant’s KB0142 answers were rated unhelpful.
- Evidence. The finding marks KB0142 outdated and KB0217 current, citing the article lines, the linked records and a hash of each input. KB0305 has only 7 linked records, so Ecdotic makes no call on it.
- Decide. The owner decides. A simulated IT knowledge manager accepts the finding and approves the correction: retire KB0142 and redirect it to KB0217. Nothing changes until they do.
- Record. Every approval becomes a dated record: what Northwind believed, what changed, the evidence, and who approved it. The replay illustrates the assistant answering from KB0217 after the team’s change, with a next check planned for Oct 1.
The problem
Knowledge bases drift quietly
Processes change, people find the new way, and the old article stays up. Now an AI assistant answers from it.
The record already exists
Service desk tickets and assistant ratings show which articles actually helped. Checking them against the knowledge base by hand is slow, so it rarely happens.
Your assistant repeats what’s written
Your AI assistant answers from the knowledge base. When an article goes stale, it gives the old answer to everyone who asks, without hesitating.
How it works
Three steps, and you make the final call
-
01 Input
Share one area of your knowledge base
Articles from one area you choose, plus the service desk tickets or assistant logs that cite them. 100 to 500 articles is plenty as a start. Exports are fine.
-
02 What Ecdotic finds
We check each article against how the work is done
We link each ticket and assistant answer to the article it used, and give each article one of three results:
Current Outdated Not enough evidence
-
03 What you get
Evidence you can check
Each result cites the article lines and the tickets and answers behind it, so your team can check a finding before the article’s owner decides what changes.
Trust
What you can count on
What a check reads, what it reports, and who decides.
Read-only access
A check reads only the articles, tickets, and logs you share. Read-only stays a supported way to use Ecdotic.
Evidence on every result
Every result cites the exact article lines, tickets, and assistant answers it used, with a hash of each input file.
“Unknown” is an answer
If an article has too few linked tickets or answers, the result is Not enough evidence. We don’t guess.
A person approves every change
Ecdotic shows the disagreement. Your team decides which version is right, and nothing changes until the article’s owner approves it.
Your data, and deleting it
Demos use synthetic data only. Before you share anything real, you get our data-handling statement: what we store, where, for how long, and how we delete it.
No Ecdotic tracking
Ecdotic’s own page sets no cookies and loads no third-party scripts, fonts, or analytics. The booking calendar is Calendly’s, shown in its own frame, and may set its own cookies.
From finding to fix
Today Ecdotic reports what it finds, and your team decides what to change and makes the change. Here’s how we plan to take on more of that work, one step at a time.
Today Read-only check
- Ecdotic reports each outdated or conflicting article with its evidence. Your team reviews it.
- This runs today, on synthetic data.
Next Approved fixes
- You approve a correction, and Ecdotic sends it to your knowledge base or through your existing review.
- The evidence and history stay attached to the change.
Later Pre-approved fixes
- Small fixes you’ve pre-approved go through, and everything else goes to the owner.
- We’ll prove this out before we offer it.
Who’s behind it
Aidan Ehrenhalt, Founder
I’m a computer engineering student at Georgia Tech with a background in AI, ML, and data engineering.
Ecdotic started from one question: you don’t know what you don’t know, so why would an organization be any different?
I ran into this firsthand during an internship. The onboarding docs told me to connect to the system with one command. The command worked for the tech lead and others who’d been there a while, because their accounts had permissions that new hires didn’t get. Security rightly declined to grant them, and it took me about a week to find another way in. The docs were right for the people who wrote them, but wrong for the next person who read them.
I rewrote the documentation myself, but fixing one section doesn’t keep the others from drifting. Because of that experience, I’m building Ecdotic to help teams see what people actually do, correct the documents that disagree, and keep them correct, so the next new hire doesn’t hit the same wall.
Evidence
Where the evidence stands
Ecdotic is early. Here’s what’s been shown, what hasn’t, and what’s next.
Shown on synthetic data
- Links service desk tickets and assistant answers to the articles they cite, and marks each article Current, Outdated, or Not enough evidence.
- Keeps article line references and file hashes with each result.
- Keeps the owner’s decision as a dated record: what was believed, what changed, the evidence, and who approved it.
- Checks a runbook’s steps against deploy logs, including their order.
Not yet shown
- Results on a real company’s knowledge base. That’s what the first reports are for.
Testing next
- A time-boxed, read-only report on one area of your knowledge base, with the tickets or assistant logs you can share.
- Someone outside Ecdotic reviews each finding before we call it actionable.
Where this is going
The knowledge base is one case of a wider problem
Direction, not built This is where we’re headed, not what Ecdotic does today. Today it checks knowledge-base articles against tickets and assistant logs, on synthetic data.
Memory
- Read what your organization wrote down and what your people did: articles, tickets, assistant logs, runbooks, and other records.
Truth
- Show where those sources disagree, with the evidence. The responsible owner decides what’s current, and the record keeps what changed, who approved it, and when.
Intelligence
- Once that record holds up, point people and AI agents to the current answer. What they do feeds back into memory.
Four kinds of disagreement we expect to find
Drift
Documents no longer match what people do.
Gaps
People do something that no document describes.
Contradictions
Two sources make claims that can’t both be true.
Emergent knowledge
One team has a practice that others haven’t seen yet.
FAQ
Questions teams ask first
What does Ecdotic need access to?
Article exports from the area you choose, plus the service desk tickets or assistant logs that cite those articles. Exports are fine, and we can work from anonymized data if you prefer.
Will Ecdotic change my articles?
No. Today Ecdotic reports what it finds and the evidence, and your team decides what to change. We’re building the next step: once the owner approves a correction, Ecdotic sends it to your knowledge base, with the evidence attached.
Is a disagreement always a mistake?
No. It means the article and the record disagree. The article may be out of date, people may be working around a policy, or there may be an approved exception nobody wrote down. What happens isn’t automatically what should happen, so the article’s owner decides which side changes.
What happens when there isn’t much data?
Ecdotic reports Not enough evidence for that article and makes no call. It doesn’t turn missing records into a finding.
How is this different from a “last reviewed” date?
A date tells you when a page changed, not whether it still works. Ecdotic checks each article against the tickets and assistant answers that used it.
How is this different from duplicate and conflict detection?
Finding two pages that disagree tells you there’s a problem, not which page is right. Ecdotic checks both against how the work is actually done, so the owner can see which version is current and why.
Is the demo real?
Northwind and its articles, tickets, and assistant logs are synthetic. The counts, verdicts, and file hashes are computed from those records, and the decision record is written by Ecdotic’s record-keeping code. The desktop, the reviewer, and the clicks are illustrative.
How do you handle our data?
See What you can count on: it covers access, approval, and the data-handling statement.
Can we use Ecdotic today?
Ecdotic is early. It runs today on synthetic data, and we’re looking for the first teams to try it on one area of their knowledge base. Book a walkthrough to see whether yours is a fit.
Book a walkthrough
See it on a knowledge base like yours
30 minutes with the founder. I’ll run the demo live, answer your questions about access and review, and hear about an article that no longer matches how your team works.
- A live demo on synthetic data, start to finish.
- Your questions about data access, evidence, and review.
- If it fits, a plan for a time-boxed, read-only report on one area of your knowledge base.
Calendar not loading? Book on Calendly.