People are keying data that already exists
What we do. We extract the fields, write them into the system of record through its own interface, and route the document. The keying stops rather than moving to a review screen that takes just as long.
AISTORSServicesDocument and process automation
Service 02Automation for the work nobody should be doing.
The procedure already exists. It just isn't running itself.
For finance, operations and back-office teams where documents arrive faster than people can key them in. The procedure is already written down, already followed by hand, and already counted somewhere. We automate the keying and the routing, keep a named person on the exceptions, and report against the count you started with.
Book a 30-minute call
Before anything is built
We count what the work costs you today: how many arrive, how long each takes, how often one comes back. Without that number there is nothing to prove afterwards.
01
Who this is for
If the left column is not you, say so on the call and we will tell you what would help instead.
This is for you if
- The procedure is already written down. There is a policy, checklist or template covering what happens to each document, even if nobody follows it exactly.
- Volume runs to hundreds or thousands a month. Enough that keying time is a real cost and accuracy can be scored properly rather than estimated.
- The documents are broadly consistent. Invoices, claims, purchase orders, statements, forms. Variation is expected; total inconsistency is not.
- Somebody can label a sample. A person who can mark fifty real documents right or wrong, so accuracy becomes a number instead of an impression.
- You can live with an exception queue. There will always be one. The measure that matters is whether it shrinks, and that is reported monthly.
This is not for you yet if
- Every document is genuinely unique. One-off contracts with no repeating structure suit a person with an assistant better than an extraction pipeline.
- The volume is small. A few dozen a month rarely repays a build. We would rather say so than sell it to you.
- Nobody owns the exceptions. Automation concentrates the hard cases. Without a named owner for them, the queue simply becomes the new problem.
- The upstream process is broken. If the document arrives wrong, automating the keying only makes wrong data arrive faster. Fix the procedure first.
- You need it to be perfect. Any provider quoting an accuracy figure before seeing your documents is guessing. We commit to a measured figure and a threshold.
02
What we fix, and how
Four things go wrong with document automation. Each has a specific answer, agreed before any build starts.
The last attempt created a worse queue
What we do. A confidence threshold is set before go-live. Below it, work goes to a named person rather than straight through. The exception queue has an owner from day one and its size is reported every month.
Nobody can tell you the accuracy
What we do. Extraction is scored against a labelled sample of your own documents before it ships. You get the figure per field, and the confidence threshold that follows from it.
The vendor will change the screen and it will break
What we do. Where a real interface exists we use it, because it is cheaper to keep working. Where only the screen exists we say so, price the maintenance honestly, and never present it as a one-off build.
03
What you get
Where an API exists we use it. Screen-level automation is a last resort with a maintenance cost attached.
What we build
- Document processing. Invoices, contracts, purchase orders, claims and forms, extracted into the system of record with a confidence threshold and a human queue for anything below it.
- Reconciliation and month-end close. Matching across systems, exception queues that shrink rather than grow, and an audit trail for every automated match.
- Quoting from the rules that actually apply. Including the rules sheet nobody has updated in three years, which the assessment will find and name.
- CRM hygiene and enrichment. Deduplication, normalisation and field completion, with the merge rules agreed before anything is merged.
- Ticket and case handling. Classification, routing, enrichment from history, and drafting for a person to approve.
- Order entry from unstructured input. Emailed purchase orders, scanned documents and attachments, into the order system with the exceptions surfaced.
- Reporting and dashboards. Built once against the source of truth, rather than rebuilt monthly in a spreadsheet.
- Prior authorisation and revenue cycle work. Packaging, submission and exception handling, designed against the CMS deadlines above with the audit record first.
What this is not
- Not a robotic process automation licence. Where an API exists we use it. Screen-level automation is a last resort and is scoped as a maintenance commitment, because the vendor will change the screen.
- Not automation of a broken procedure. Automating a process nobody has fixed makes the wrong output arrive faster. Where the procedure is the problem we will say so.
- Not full autonomy on regulated decisions. The system prepares the file. A named person decides, and that boundary is a contract term.
- Not scoped without a count. If nobody knows how many of these are processed a week, that is the first thing the Diagnostic establishes.
- Not a promise of zero exceptions. There will be an exception queue. The measure that matters is whether it shrinks.
04
How it runs
The count comes before the build, because the work is usually already counted somewhere.
01
Count the work first
Volume, cycle time, error rate and hours on the target procedure. This is the easiest line in the catalogue to baseline, because the work is usually already counted in a system somewhere.
02
Read the procedure as it is actually performed
Not as documented. The gap between the two is where the exceptions live, and the exceptions are what determine whether automation is viable.
03
Integration path confirmed before build
Which systems accept a write, which need a browser-level approach, and what that will cost to maintain. Legacy systems integration is cited by 28 percent of middle-market respondents as a top inhibitor.
04
Build with a confidence threshold and a human queue
Extraction and matching with an explicit threshold. Below it, work routes to a person. Above it, it proceeds with an audit record. The threshold is agreed, not chosen by us.
05
Measure against the count, then widen
Same four measures as the baseline. The exception rate is reported honestly, including where it is higher than expected, and scope widens only after the first procedure holds.
05
Platform native
Delivered natively on whichever platform you are already on. All four, no reseller relationship on any.
Azure
Document Intelligence, Logic Apps and Functions, Power Automate where the estate is Microsoft-centric, Service Bus for queueing.
AWS
Textract, Step Functions and Lambda, EventBridge, SQS for exception queues, A2I for human review.
Google Cloud
Document AI, Workflows and Cloud Run, Pub/Sub, Cloud Tasks for retry and queueing.
DigitalOcean
Functions and App Platform, managed Postgres for state, managed Kafka or Redis for queueing, open-source OCR where a managed service is not warranted.
Where an API exists we use it, which is more robust and cheaper to maintain. We say which category each integration falls into before you commit.
06
The evidence, if you want it
You do not need these numbers to recognise the problem. They are here because somebody in your approval chain will ask.
1 Jan 2027
is the date by which impacted payers must primarily meet the API requirements of the CMS Interoperability and Prior Authorization Final Rule, having implemented certain other provisions by 1 January 2026.
Source: CMS, CMS-0057-F.
$4.4tn
in illicit financial activity was estimated for 2025, up 1.3 trillion dollars since 2023, with fraud scams and bank fraud causing 579.4 billion dollars in losses.
Source: Nasdaq Verafin, 2026 Global Financial Crime Report.
28%
cite legacy systems integration as a top inhibitor to AI deployment, level with the talent gap.
Source: RSM US Middle Market AI Survey, 2026.
There is no reliable market size for this work, and we are not going to quote one. Named research firms published 2026 base-year figures for intelligent document processing ranging from roughly 3.1 billion to 14.2 billion dollars, a four-and-a-half-fold spread for the same category in the same year. Dated regulatory deadlines are a far better guide to where this work is genuinely being bought.
07
The dates that force the work
Five dates, each verified against the body that issued it.
Artifact / Dated obligations that force document and process work
Verified against the issuing body
1 January 2026
CMS-0057-F
Certain provisions of the CMS Interoperability and Prior Authorization Final Rule required of impacted payers.
Source: CMS
2 December 2026
EU AI Act
Transparency and machine-readable labelling obligations for AI-generated content apply to systems placed on the market before 2 August 2026.
Source: European Parliament, 11 June 2026
1 January 2027
CMS-0057-F
The main API requirements for impacted payers.
Source: CMS
2 December 2027
EU AI Act
High-risk obligations for stand-alone Annex III systems, moved out from 2 August 2026 under the Digital Omnibus on AI.
Source: Council of the EU
2 August 2028
EU AI Act
High-risk obligations for AI embedded in regulated products under Annex I.
Source: Council of the EU
08
Questions we are actually asked
How accurate is document extraction?
It depends on the document, and any number quoted before seeing yours is marketing. What we commit to is a measured accuracy figure on your own documents, scored against a labelled set before it ships, plus a confidence threshold below which work goes to a person rather than through.
What happens to the exceptions?
They go to a named queue with a named owner. There will always be an exception queue, and a provider who tells you otherwise is selling something. The measure that matters is whether the queue shrinks over time, and that is reported monthly.
Our system has no API. Is this still possible?
Usually yes, through browser-level automation, but it is scoped as an ongoing maintenance commitment rather than a one-off build, because the vendor will change the screen eventually. We will price that honestly rather than treat it as free.
Is this the same as robotic process automation?
No. Where an API exists we use it, which is more robust and cheaper to maintain. Screen-level automation is a last resort for systems that give us no other option, and we say which category each integration falls into before you commit.
We are a payer facing the CMS deadlines. Where does this start?
With the audit record and the exception path, not with the extraction. The API requirements land primarily on 1 January 2027, with certain provisions from 1 January 2026, and work designed backwards from the audit obligation is the work that survives a review.
Can you automate a process we have not documented?
We can, but the first step is observing how it is actually performed rather than how it is written down. The gap between the two is where the exceptions live. If that observation shows the procedure itself is broken, we will tell you to fix the procedure first.
09
Next step
Bring the procedure and the count.
Thirty minutes, no obligation. Tell us what is being processed by hand and roughly how much of it there is. If nobody knows the volume, that is the first thing worth establishing.
- Diagnostic duration
- Three to five days. One day for a solo operator or small business.
- Investment
- Scoped on the introductory call.
- If nobody knows the volume
- That is the first thing the Diagnostic establishes: volume, cycle time, error rate and hours on the target procedure.
Book a 30-minute call
Or write to [email protected].
10
Related work
Service 03
Data and retrieval for AI
Answers from your own records. For the cases where the automation needs the record rather than the document in front of it.
Service 01AI agents and assistants
Inside the system, not beside it. For the steps that need judgement rather than a rule that can be written down.
Service 09Managed AI and cloud operations
The same four measures, every month, including whether the exception queue is still shrinking.