AI EngineeringAI Product Engineering
Document Intelligence Sprint
One family of documents (for example, supplier invoices) flowing from where it arrives today into the client's system of record through a production extraction pipeline, with accuracy measured field by field against a labelled test set and every low-confidence field sent to a person instead of into the system.
Document Intelligence Sprint
$30K–$45K
- Duration: 4 weeks
- Tier: Build
- Format: Fixed scope
You leave withProduction IDP pipeline with measured extraction accuracy per field
Book a Free SessionSound familiar?
Three people spend all day keying numbers off invoices and statements. It's slow, mistakes slip through, and every time volume spikes we bring in temps. We tried an OCR tool. It read the text fine but couldn't tell which number was the total, so nobody trusts it and someone checks every document anyway.
What you leave with
Extraction pipeline
Format: Code in client repo; deployed in client cloud
How we know it’s acceptableProcesses the full test set end to end in production
Field specification
Format: Document
How we know it’s acceptableSigned by the process owner in week 1
Labelled test set and eval script
Format: In client repo
How we know it’s acceptableClient's engineer re-runs it and reproduces the reported accuracy without us
Per-field accuracy report
Format: Document
How we know it’s acceptableEvery field has a measured accuracy and threshold; signed off by the process owner
Review queue
Format: Configured in agreed tool
How we know it’s acceptableA low-confidence field from the test set lands in the queue as expected
Monitoring view
Format: Client's tooling
How we know it’s acceptableShows volume, straight-through rate and cost per document
Runbook
Format: In client repo
How we know it’s acceptableClient's engineer changes a threshold and redeploys during handover, unaided
Also included
- Integration with one downstream system
How it runs
4 weeks
Week 1
Document census, field specification, labelling the test set, agreeing accuracy thresholds per field.
OutSigned field spec and labelled test set.
Week 2
Build ingest, classification and extraction; first accuracy run.
OutBaseline per-field accuracy.
Week 3
Validation rules, confidence thresholds, review queue, integration with the downstream system.
OutPipeline working end to end in staging.
Week 4
Production deployment, monitoring, cost per document, handover.
OutAccuracy report, runbook, paired handover.
Is this right for you?
This is for you if
- Buyerthe person who owns the process the documents feed: accounts payable, claims, onboarding or back-office operations. Budget usually sits with the COO or CFO.
- Also has to say yesthe owner of the system that receives the data (ERP, CRM or core platform), because we write into it. Whoever signs off data protection, since these documents usually carry financial or personal data. Finance, if the price is above the buyer's sign-off limit.
This is the wrong answer if
- Volume is too low. If the client processes a few hundred documents a month, manual entry or an off-the-shelf OCR tool is cheaper than a pipeline and the return will not cover the engineering.
- The data is already structured. If documents arrive as e-invoices, EDI or through an API, this is an integration job, not an extraction job.
- Nobody can say what 'correct' is. If the client cannot tell us which value is right for each field, we cannot label a test set, and without one we cannot measure accuracy. The process needs defining first.
- They want the whole archive at once. Ten years of mixed invoices, bank statements and drawings is several document families. This sprint covers one; the rest are follow-on sprints.
- The documents need interpretation, not extraction. Engineering drawings, where geometry and annotations must be read together, or heavy handwriting, are R&D and should not be sold at a fixed price.
What’s included, and what isn’t
In scope
- one document family
- up to five major layouts and about 25 fields
- one intake channel (an inbox, a folder or an upload)
- one downstream integration
- a review queue using an existing tool or a simple screen
- one production environment
Out of scope
- processing the historical archive
- additional document families
- cleaning data already in the downstream system
- changes to the client's ERP or other systems
- a full review application
- drawing or handwriting interpretation
- support after handover, which is Managed Services (Essential or Growth) and priced separately
Before we start
Before day one
- Sample documents pulled at random from real recent volume, not a cleaned subset, with permission to use them.
- The field list agreed with the process owner.
- Access to a cloud account and to a sandbox of the downstream system.
- A named expert who can confirm correct values for the test set in week 1.
- Data-processing approval for any model provider we will use.
What we’ll need from your team
- The process owner or a domain expert: four hours in week 1 for the field spec and labelling, then two hours a week to confirm accuracy.
- A platform or IT engineer: three hours a week for access and the integration.
- A decision-maker available within 24 hours for threshold decisions.
- The people who will work the review queue: two hours in week 4 for training.
Pricing
What we need to give you a quote
- 30 to 50 real sample documents, including the worst ones.
- The list of fields they need and the system they go into.
- Monthly volume and number of layouts.
- Hosting and data-residency constraints.
- A named person who can confirm the correct values for the test set.
After we hand over
Who owns it after we leave
The process owner (for example, the AP lead) owns the thresholds and the review queue. A named engineer on the client's team owns the pipeline and the test set. At handover we state the review effort from the review rate measured in week 4 (for example, how many documents a week need a person, and roughly how many hours that takes) so it is planned into someone's week, not discovered.
What we measure
- Before (measured by the client in week 1): minutes per document and current error rate.
- After (reported at handover): straight-through rate, per-field accuracy, minutes per document including review, cost per document.
What comes next
- Another sprint for the next document family, and a scoped backfill of the historical archive.
- Managed Services Essential or Growthonce the pipeline is live.
- The next phase is acting on the extracted data, for example reconciliation, which leads to the Agent Prototype Sprint.
Questions buyers ask us
For standard invoices from a handful of suppliers, it may. If your current tool hits the accuracy you need, keep it. We come in when layouts vary, when the tool can't tell which number is the total, or when values need checking against your own records.
Not before we've measured on your documents. What we guarantee is that you'll know the accuracy of every field, and that anything below the agreed threshold goes to a person rather than into your system.
Reading is fast. Knowing when the reading is wrong takes a labelled test set, validation rules and a review queue. That's most of the four weeks, and it's what makes the output trustworthy.
The pipeline runs in your cloud, using models whose data terms you approve. If you need open-weight models on your own infrastructure, we can do that, but it moves the price.
We start with one document family and prove the numbers. Once the pipeline exists, the archive is mostly throughput, scoped as a follow-on.