We keep it running
after it goes live.

Automation does not break loudly. A supplier changes their invoice layout, an API is updated, December volume doubles, and accuracy slides for six weeks before anyone notices. We watch it every day, fix it when it slips, and send you a report each month showing exactly what it did.

This is for you if

Services

What we do every day, week and month

The work splits into three. Watching it, fixing it, and telling you honestly how it is doing.

01

Watch it every day, automatically

Checks that run on their own: how much it processed, how much it got right, how many items are waiting for a person, how many failed outright. If any of those moves outside its normal range, we are told before you are.

02

Watch for accuracy sliding rather than breaking

The dangerous failure is not a crash. It is one supplier changing their layout so that a field is read wrong nine times out of ten while everything looks fine. We track accuracy per document type and per field, so a drop in one is visible even when the overall number barely moves.

03

Clear the exception queue

Someone on our side works through the items the system could not decide, so they do not pile up on your team. Anything needing a business decision goes to you named and explained, not as a raw error.

04

Fix it when a document changes

Suppliers redesign their paperwork without warning. We adjust the extraction, test it against the saved examples, and put it back. This is the most common piece of work in any month.

05

Keep your rules current

Tolerances change, approval limits change, a new company gets added. We make those changes and test them, rather than leaving your team to work around a rule that is now wrong.

06

Keep up with everything around it

API versions retired, ERP upgrades, certificates expiring, libraries with security holes. Most outages come from something changing underneath the automation, not from the automation itself.

07

Answer the phone when it stops

Agreed response times by severity, a named person, and an escalation path that ends with someone who can actually change the code. Not a ticket queue.

08

Report every month, including the bad months

What it processed, what it got right, what needed a person, what broke and why, and what it cost to run. Written so you can forward it to your finance director without translating it.

09

Keep the audit trail defensible

Access reviews, log retention, evidence that a person approved what a person is supposed to approve. When your auditor asks, the answer already exists.

10

Tell you what to improve next

Each quarter, a short list of what would make the biggest difference. Usually a rule change or a supplier conversation rather than more software, and we will say so when that is the case.

If something is already live, we will look at it and tell you what state it is in.

The monthly report

What lands in your inbox each month

One page. The numbers first, then what happened and what we did about it, in plain words.

Invoice processing, monthly reportExample month
3,184Documents processed
91.4%Went through with no person involved
274Needed a person to decide
2Failed outright, both recovered

Volume was 22% above normal. December always is. Nothing needed to change, but the review queue was slower on the 18th and 19th because two of your approvers were on leave at the same time.

One supplier changed their invoice layout on the 9th. Accuracy on their tax field dropped to 40% for eleven documents before the alert fired. We fixed the extraction the same day and reprocessed all eleven. None reached your ERP wrong.

Both failures were the same cause. Your ERP was down for maintenance on the 22nd and the automation stopped rather than retrying into a closed system. Both were posted automatically once it came back.

Suggested for next quarter. Sixty percent of the items sent to a person came from three suppliers who all send a photograph rather than a PDF. One email to those three would remove most of the manual work in this process. That is worth more than anything we could change in the software.

An example report with invented figures, to show the shape and the level of detail. Yours will have your own numbers.

Our stack

The tools, standards and methods we use

You do not need to know any of these. They are listed so you can see this is run the way any serious software service is run, with alerting, agreed response times and a record of every change.

Watching the systemDashboards and automatic checks, so a problem is noticed by a machine at 3am rather than by your team at 9am
GrafanaPrometheusDatadogOpenTelemetryUptime and synthetic checks
Watching the accuracyDifferent from watching the system. This catches the automation quietly getting things wrong while running perfectly
Drift detectionConfidence trackingAccuracy per fieldSampled human review
Getting hold of a personAlerts go to a rota with an escalation path, not to an inbox somebody may open in the morning
PagerDutyOpsgenieSeverity levelsOn-call rota
What we commit to in writingResponse and fix times by severity, agreed before we start, and reported against every month
Service level agreementResponse and resolution targetsMonthly availabilityError budget
Where the tickets liveYours if you have one, ours if you do not, so there is one list rather than two
Jira Service ManagementServiceNowFreshserviceYour own helpdesk
Changing something safelyChanges go out in stages and can be switched back within minutes if the numbers move the wrong way
Staged rolloutFeature flagsRollbackRegression testsChange approval
Watching what it costsTracked per document, so the running cost is a number rather than a surprise on a cloud bill
Cost per documentToken usage trackingBudget alertsModel cost comparison
Keeping the auditors satisfiedThe evidence is collected as it happens, so nobody has to reconstruct it later
Logs that cannot be editedAccess reviewsRetention policyEU AI Act record keeping

Example

The kind of failure nobody notices

Example

Take the supplier who changed their layout on the 9th in the report above. Nothing broke. No alert about the system being down, no failed jobs, no angry email. The automation carried on processing invoices all day and the overall accuracy number moved from 94% to 93.6%, which nobody would look at twice.

What actually happened was that this one supplier had moved their tax number into a different box. For their invoices, and only theirs, that field was now being read wrong nine times out of ten. They are about eleven invoices a month out of three thousand, so the overall figure barely moved.

It was caught because accuracy is tracked per supplier and per field, not just overall. That alert fired on the second document. Without it, the realistic outcome is that this is found at the quarter end by an accountant reconciling tax, three months and a hundred and twenty invoices later.

An example built from the kind of problem we see most often. The numbers are invented.

FAQ

Questions we get asked before starting

Will you look after something you did not build?

Yes, and a good part of this work is exactly that. We start by reviewing what is there and telling you honestly what state it is in. Occasionally the answer is that it is cheaper to rebuild than to maintain, and we will say so.

What if we want to bring it in house later?

Then we hand it over. Everything is documented as we go for that reason. Plenty of clients use this for a year while they hire, and that is a normal way to use it.

Do you need access to our live systems?

We need enough to see what the automation is doing and to fix it. That is usually read access to the logs and the ability to release a change through your own approval process. It is agreed in writing and reviewed.

What counts as urgent?

We agree that with you before starting. Normally: it has stopped, it is producing wrong output, or the queue is growing faster than anyone can clear it. Each gets its own response time.

Will you tell us if we are paying for something we do not need?

Yes. If volume has fallen or a process changed so the automation is no longer earning its keep, that goes in the report. It is a short term loss for us and the alternative is you finding out on your own and wondering what else we did not mention.

How long is the commitment?

We work in agreed periods with a notice period rather than a long lock in. If it is not working you should be able to leave, and take everything with you.

Next step

Tell us what is live and who is watching it.

What it does, roughly how much it handles, and who gets called when it stops. If the honest answer to the last one is nobody, that is the most useful thing you can tell us on the call.

Ahmedabad, India. We work with teams in the US, UK, Europe, Singapore and the Gulf, and we are used to the time difference.