Public and Funded Programs
Eligibility, claims and acquittal are checked by people reading documents against an agreement. We move the checking into a system that cites the clause behind every answer, and leave every decision about money with a named officer.
What it costs today
Every one of these is work an officer is already doing, in a way that leaves too little behind to be relied on later. Six patterns turn up in almost every program we map.
An officer opens the application, opens the guidelines, and works through them. The assessment is only as consistent as how tired the officer was, and two applicants with the same facts can get different answers.
A decision you cannot show was made the same way as the last one.Invoices, attendance records, delivery reports and receipts, each one opened, read against a schedule and ticked off on a spreadsheet that lives beside the case rather than inside it.
A processing time set by how many files an officer can open.The claim says a milestone was delivered and the proof is somewhere in a folder of 29 documents. Six months later, at audit, somebody has to establish which document proved which line.
An audit finding about record keeping rather than about spending.The registration document arrived with the application and is requested again at the first claim, because the two live in different systems and neither knows what the other holds.
Delivery partners who stop applying.Uptake, spend against profile, outcomes by cohort and time to decision are all knowable and none of them are visible while the program is still running and could still be changed.
Underspend discovered at the point it cannot be redeployed.US federal agencies reported 186 billion dollars of improper payments in FY2025. The GAO attributes much of that to eligibility and documentation rather than to fraud, which means most of it was checkable before the payment rather than recoverable after it.
Recovery costs more than the check would have.The improper payments figure is there to size the problem rather than to accuse anybody. The GAO puts most of it on eligibility and documentation, which means most of it was checkable before the money moved. That is a process finding, and it is a much more useful one than a fraud finding would be.
The operating pattern
Controlled delegation is the whole design, and everything below this band is an application of it. Reading and checking against the agreement go to the system. What happens to the money goes to an officer, on every claim, without exception.
Applications, claims, invoices, attendance records, milestone reports and variation requests. The same shapes over and over, each one measured against the same agreement.
Eligibility, expense period, category and ceiling, evidence against milestone, participants against register, and the cumulative position against the profile. Each answer carries the clause it came from.
Approve, reduce, return, escalate or grant a retrospective approval. Every one of these reaches a named officer with the assessment done, the rule cited and the evidence attached.
See it run
The funding agreement, the delivery records, the participant list and the invoices get reconciled against each other, and what the evidence supports is separated from what it does not.
Connected systems
Claim Claim 00982 Wrenfield Training Trust
Result
Completed automatically
Requires a person
Whether the claim is paid.
Program officer
Connected systems
Claim Claim 01147 Beacon Community Works
Result
Completed automatically
Requires a person
Whether to part-pay now or hold the whole claim.
Program officer
Connected systems
Claim Claim 01203 Redgate Enterprise Hub
Result
Completed automatically
Requires a person
Whether any discretion applies, and what the recipient is told.
Program officer
This demo uses fictional data. The workflow gets configured around your systems. Build a workflow like this
The work
Claim assessment usually goes first because it is where the volume and the exposure both sit. Each card names what the system handles and where an officer stays in control.
Every line checked against the agreement, with the clause attached
Reads the claim and its attachments, checks eligibility, period, category and ceiling, matches evidence to the milestone it supports, verifies participants against the register, and produces a recommended payable amount with the rule behind every reduction. It does not pay anything.
Data: claims, agreements and delivery records. Human: a program officer decides on every claim.
The same questions asked the same way of every applicant
Completeness checked against the requirement list, eligibility assessed against the published criteria, duplicates and related entities identified, and everything that cannot be determined from the submission named as a question rather than resolved by assumption.
Data: applications and published criteria. Human: an assessor decides eligibility and any discretion.
The agreement, the claims and the evidence on one file
The application, the agreement, every claim, every piece of evidence, every variation and the final acquittal on one record. This is what makes an audit a retrieval exercise rather than a reconstruction, and it is what the two workflows above stand on.
Data: your program, recipient and evidence records. Human: officers own every decision on the file.
Uptake, spend and outcomes during the program rather than after it
Spend against profile, uptake by region and cohort, time to decision, and outcomes against the objectives the program was funded to achieve, visible during delivery. Underspend found in month eight can be redeployed. Underspend found at acquittal is a return.
Data: your own program records. Human: program leadership decides what to change.
The rule-writing stage finds things nobody wanted to find, and that is the point. Turning a set of guidelines into rules a system can apply surfaces every place two officers would reasonably reach different answers. Those places are already producing inconsistent decisions, in both directions, and nobody can currently see it. Several programs have taken that finding and never built the workflow, which is a legitimate outcome.
Authority and evidence
These get agreed with program leadership, whoever owns privacy and whoever will answer to the auditor. Between them they decide what the workflow is allowed to be, and the first one is not negotiable.
What counts as working. Accepted output against your own baseline is the proof. The first build is usually tested against decisions your officers have already made, so you can see where the system and the officers disagree before it touches a live claim, and so a disagreement is a finding rather than an incident. How we measure AI work sets out the method.
How it starts
The first build runs on one program and gets tested against assessments your officers completed, which is the only way to find out whether the rules as written match the rules as applied.
Proof
The approach on this page comes out of these engagements. Two sit inside funded programs, from both sides of the table, and two are the design patterns the workflows above are built on.
A funded program with money to deploy and a deadline to deploy it by. A 70,000 dollar multichannel campaign across more than ten channels, reaching the applicants the program existed to reach. The demand side of the same problem this page is about.
A plan written to a funding body's standard, assessed and accepted. Useful here for the opposite reason to usual: it is what a good submission looks like from the inside, written by people who now build the systems that assess them.
Fixed rules produce the classification so the same inputs always give the same answer, and every output waits in an approval queue before anyone sees it. The same two design decisions the claim workflow above is built on.
An escrow exchange built from nothing, holding funds until both sides were satisfied. Designed, built and deployed, with the controls as the feature rather than as a wrapper around it.
Begine Fusion is a Zoho Authorized Partner working across Canada, the United States, the United Kingdom and Nigeria, with a team on the ground in Nigeria since 2013. Every engagement is published in full, including what was measured and what was a design target. The full set is on our work.
Fit
The work pays back when a program runs long enough and at enough volume that consistency between officers is a real exposure. If that is not where you are, the right column says what to do instead.
Where to start
The same five offers every other client buys, scoped to an organization that has to evidence how public money was assessed. Find the line that sounds like your program.
No price is quoted on this page. A sector is not an offer and does not have a price of its own. What a program here pays follows which of those six lines is the real one and how precisely the guidelines are already written. Working out which line is yours is what the first call is for.
Questions
The seven that decide whether a build is worth starting, answered before the proposal rather than inside it.
No, and it is built so that it cannot. It reads the claim, applies the rules in your agreement, cites the clause behind every answer and produces a recommended payable amount. An officer approves, returns, reduces or escalates. There is no auto-approve threshold, including for claims that pass every check, because the value of the control is that a person signed and an exception-only review means nobody looked at most of the money.
That is the normal starting position and it is the most useful finding of the engagement. Turning guidelines into applicable rules surfaces every place where two officers would reasonably reach different answers, and those places are already producing inconsistent decisions today. Writing them down is worth doing whether or not anything gets built on top of it.
The workflow reads what arrives, including a scanned form or a posted document, so the burden of changing sits with the program rather than with the applicant. A program that only works for applicants who can submit cleanly through a portal has narrowed who it serves, and usually not in the direction it intended.
By showing what was checked, which version of which rule produced which answer, what evidence supported it, what the system recommended and which named officer decided. That record is a by-product of how the workflow runs rather than something assembled for the auditor, which is the difference between a retrieval exercise and a reconstruction.
It surfaces things worth looking at: duplicate claims across periods, related entities applying separately, evidence that does not support the line it is attached to. Every one of those has an innocent explanation as well as a guilty one, so the system reports the pattern and the basis for it and never characterises intent. Whether something is fraud is an investigation, and it belongs to people who run investigations.
A baseline is taken before the build exists: current time to assess a claim, the rework rate, how often assessments are overturned on review, and how long a recipient waits for a decision. Afterwards the same work is measured against it, with the officer review burden counted openly.
Roughly eight to twelve weeks from the first session to one workflow running on one program against a baseline you set. Ahead of it sits the mapping and the rule writing, which together are four to five weeks and produce something you can use whether or not we build anything.
Thirty minutes with a program lead, an officer who assesses claims and whoever answers to the auditor is enough to name the first workflow, the rules it needs written down, and the baseline it gets measured against.
