Insights · Economics
The model is rarely the expensive part.
A recurring document job costs what it takes to get legitimate access, map fields, check the output, host the run, and keep it alive when a source changes. Those line items dwarf token spend on work of this shape. This page is a purchasing frame, not a quote. Figures below are labeled where they are illustrative.
Where the money actually goes
Access
Accounts, MFA, allowlists, and the time to prove you may retrieve the files. If this step is unclear, do not start implementation. You will pay for a job that cannot run.
Mapping
Each source has its own layout. Someone has to say which fields matter, what “empty” means, and what the destination columns are. That work is specific to the circuit. It does not vanish if you switch models.
Verification
Deterministic checks on the extraction — required fields, types, counts, cross-source agreement — plus a path for failures that must not be sent. In the healthcare collection case, this is ordinary automation sitting after the model, not a person rereading every row.
Hosting
Compute, storage, secrets, and the environment the contract requires. For sensitive work this may be the client’s cloud, a dedicated tenancy under a BAA, or hardware they control. See security and data handling.
Maintenance
Portals change. The first quiet week is not the cost of ownership. Someone has to notice a failed run, repair the collector, and confirm the checks still mean what they meant last month. If nobody is on the hook, the team returns to the manual circuit.
An illustrative daily job, not a quote
Take a job like the healthcare case: several external portals, a daily cadence, a spreadsheet the agents already use, mixed extraction plus checks, and an operator on the run. The build is dominated by access, mapping, and verification. Ongoing cost is dominated by failed logins, layout changes, and exception handling. Model inference is a line item, not the reason to do or skip the work.
We are not publishing dollar amounts, hours saved, or a payback period for that engagement. Those figures were not part of the public case, and an invented average would be worse than silence. If you need a number, it has to come from your volume, your failure cost, and a scoped assessment of this circuit — not from a marketing page.
When not to automate
Some circuits should stay manual.
Skip a custom job when a native connector already covers the fields; when the work is a one-time migration rather than a cadence; when volume is too low for maintenance to beat a person on a checklist; or when a missed or delayed file is cheaper than owning a brittle collector.
Also skip it when every source would need a different system. Repeatable leftover work is the point. A new architecture per portal is a services sink, not an operation you can price.
How we price the conversation
We start with a paid workflow assessment of one bounded circuit: sources, destination, frequency, what you already bought, and what remains manual. Implementation is that circuit, not a platform. Ongoing operation is a separate line because that is where the cost lives after go-live.
Approach choices — API versus browser automation versus mixed — are in the companion note. If the assessment says buy software instead, that is a successful assessment.
Contact
Bring the circuit, not a budget guess.
Frequency, volume, destination, and what a bad file costs you. We will say if automation is the wrong spend.
Start a conversation