Okay, I have to say it because I'm genuinely trying to understand. I see all the announcements about Dext's new AI-powered categorization and smart insights. But after running it for a month on our firm's subscription, I'm not convinced it's anything more than a dressed-up rules engine.
My team's pipeline ingests receipts and invoices from Dext into our data warehouse (Snowflake) via their API. The promise was less manual review. But I'm seeing:
* The "AI" categories often mismatch our detailed chart of accounts, defaulting to vague parent categories we don't use.
* For complex receipts (like a combined software/hardware purchase), it just picks one line item and ignores the rest.
* I can't find any measurable improvement in the reduction of "Needs Review" items in our weekly reconciliation reports. The variance is within our normal noise.
As someone building data pipelines, I look for evidence in the data. So my questions are:
* Has anyone done a proper before/after analysis? Like, tracked the "automatic vs. manual review" rate over a significant sample?
* Are there any actual accuracy metrics published? Or details on what model they're even using?
* Is the "smartness" just better pattern matching on the vendor name? Because my old Python script with a fuzzy-matched vendor dictionary does almost as well 😅
I attached a screenshot of our dashboard showing the flat "auto-matched" trend line after we enabled the new features. It's... underwhelming.
I want to believe! But right now, it feels like a marketing checkbox. I'd love to be proven wrong with some solid data or use cases.
null
Yeah, I've noticed the same thing with vague parent categories. We use pretty specific codes for client reimbursables, and it keeps landing on "Travel" instead of "Meals & Entertainment - Client." Have you tried training it on your chart of accounts? I did a few dozen examples, but the results are still hit or miss.
Your question about a before/after analysis is a good one. I haven't seen any data from Dext either, just case studies. Is it even possible to get those metrics from their API? I'm new to this.
Ask me in a year
That's a really practical way to look at it. Tracking the "automatic vs. manual review" rate is the only way to cut through the marketing. I haven't seen any independent analysis published either.
My hunch is that for firms with very standard, simple transactions, the improvement might be measurable. But for anyone with a detailed chart of accounts or complex multi-item receipts, like your example, the utility drops off fast. The model seems tuned for breadth over depth, which leads to those vague parent categories.
You asked about published accuracy metrics. In my experience, when a vendor doesn't lead with those numbers, it's often because the results are highly dependent on your specific data and aren't impressive in a generalizable way. I'd be curious if their support team can provide any benchmark data for a use case like yours.