If you've spent any time integrating an extraction tool, you know the loop. A new vendor shows up. Their invoice has the same fields as the last twenty — invoice number, date, line items, total — but the layout is just different enough that your template misses three columns. So someone in ops opens the template editor, drags some boxes around, ships it, and prays the next vendor's PDF lines up. It never does.
This post is about why that loop exists, why it's worse than people think, and what Docusift does instead.
Where templates actually fail
Template-based extraction works when documents fit a fixed shape. The problem is that the real world doesn't generate fixed shapes. Three failure modes show up over and over:
1. Layout drift within a vendor. Even a single supplier rotates layouts seasonally — a banner ad slot in Q4, a redesigned header in January, a new line-item table when they switch ERPs. Each shift breaks the template.
2. Cross-vendor variance. "Invoice" is not a layout, it's a class. A 50-vendor portfolio has 50 layouts. Some put the total at the top. Some bury it in the footer. Some print line items across two columns. A template per vendor scales linearly with new customers — and it never stops growing.
3. Quality variance. Scans rotate, get cropped, lose contrast. A template anchored to "the box at coordinates (412, 88)" misses if the page is shifted 20 pixels right.
The cumulative cost isn't just the templates themselves — it's the _operational load_ of maintaining them. Every new vendor is a ticket. Every layout change is a regression. Every onboarding gets blocked on someone with the template editor open.
What "template-free" actually means
Docusift doesn't ship a template editor because we don't use templates. The pipeline looks like this:
1. The document arrives (PDF, JPG, PNG, TIFF). 2. A multimodal model reads the page like a person does — it doesn't need coordinates, just visual + textual context. 3. The model classifies the document type (invoice, receipt, bill of lading, W-2, etc.) and returns a structured JSON payload of the fields that matter for that type. 4. Two confidence numbers come back with every result: classification confidence (how sure are we this is an invoice?) and extraction confidence (how sure are we about the fields we pulled?).
That's it. There's no per-vendor configuration. Onboarding a new supplier is zero clicks.
What you give up
We get this question a lot: _isn't this less precise than a tuned template?_ The honest answer is — for a single vendor whose layout never changes, a perfectly tuned template can be marginally more accurate. But there are very few vendors like that, and the tuning cost over a year of layout drift swamps the marginal accuracy gain.
The more useful framing: templates optimize the easy cases at the cost of the hard ones. Template-free extraction handles the long tail of layouts at the cost of needing a confidence bar for the rare ambiguous case.
The confidence bar is the real interface
Once you accept that no extraction system will be 100% on every document, the question becomes _what do you do with the borderline cases?_ Docusift's answer is a per-workspace auto-approve threshold:
- Above the bar (e.g., extraction confidence above 0.92), documents auto-approve and sync to your accounting tool. - Below the bar, they land in the Review queue with the PDF side-by-side with the editable fields. A human eyeballs it, fixes anything off, and the document moves on.
The goal isn't 100% automation. It's making the last 3-5% of tricky documents cheap to handle.
What replaced the template editor
In the template world, the canonical workflow is "draw boxes." In the template-free world, it's "tune the threshold." If too many docs are landing in Review, raise the bar. If you're seeing wrong fields auto-approve, lower it. One number, applied across every doc type, no per-vendor knobs.
It's not just simpler — it's the right interface. Confidence is the thing you actually care about. Coordinates were always a proxy for it.
Try it
Drop a PDF into the free extraction audit — we'll send back the JSON we extract along with confidence scores, and you can compare it against whatever template-based tool you're running today. No setup, no vendor mapping, no template draws.