Back to the summary
Inside the complete Specimen
Illustrative example. The business, people and figures are invented. This is not a client result.
Illustrative example. The business, people and figures are invented. This is not a client result.
This is the Specimen: a worked example of what a Genesis engagement produces, written out in full so that you can judge the quality of the thinking for yourself. It is one workflow in one invented firm, taken from the problem the audit found to the decision about where a tool belongs and where it does not. You can read all of it here, and it costs you nothing to look.
You are looking at the finished thinking, not a brochure about it. Read it and judge for yourself.
When we decide where AI belongs in a business, we write the decision down in a fixed form. It covers one job, from the problem it starts with to the check that tells you whether it is still earning its place, and it shows you where we stopped the tool and put a person back in. The form is the same for a plumbing firm and a law firm, so you can read how a decision is made whatever your sector.
Every record answers the same questions: the job and how it is done today; the source behind the decision and the date it was checked; who owns it; where the tool must stop; the test it had to pass; the human check that stays in place; and the honest result, including when that result is not to use AI at all.
What you have just read is the method. What follows is one record filled in, for an invented heating and plumbing firm, so you can see the questions carrying real weight. It is a trades and field-service example. When we work with you, we produce one for your own business.
The invented firm. Kestrel Building Services, an eleven-person heating, plumbing and mechanical maintenance firm. The job recorded is checking supplier invoices against what was ordered and what was delivered, before the weekly payment run.
The operating record is how we write a decision down. It covers one job in fourteen fields: the problem and the baseline, the ruling on whether to use AI at all, the boundary the tool does not cross, the data allowed nowhere near it, the named owner, the sources and their dates, the tests and what they returned (including the one the tool was caught failing), the runbook and the fallback, the measure that tells you it is still earning its place, and the next decision already queued. It is the working record, with the test results in it as they came back, not a summary written afterwards to look tidy. The whole of it is below; nothing is held back on another page.
When we decide where AI belongs in a business, we write the decision down in a fixed form we call an operating record. It covers one job, from the problem it starts with to the check that tells you whether it is still earning its place, and it shows you where we stopped the tool and put a person back in. This page is an illustrative example of that method, not a live client record. It shows you the form and one worked example of it, so you can judge how a decision would be made, not evidence that the method is already running inside a business.
This page holds two things. Part One is the record itself, empty: the fourteen questions this method asks of a job, in the same order every time. It is the same for a plumbing firm and a law firm, so you can read how a decision gets made whatever you do. Part Two is one record filled in, for an invented business, so you can see the questions carrying real weight.
Two things this record is not. It is not a claim about what AI will do for you, and it carries no promised result. And it is not a summary written afterwards to look tidy. It is the working record, with the test results in it as they came back.
An operating record covers one job. It has fourteen fields in five groups. We do not treat a job as live until every field is filled and the acceptance tests have a written result.
The business below is invented, and so is everybody in it. There is no real client here, no measured result and no promised one. Every figure belongs to this example and to the assumptions it states about itself.
Four of the five sources in field 6 are real. We read them first-party and re-checked all four on 13 September 2026. They are named with their dates so that you can go and read them yourself. The fifth is illustrative, and it says so.
The business (invented). Kestrel Building Services, an eleven-person heating, plumbing and mechanical maintenance firm. Seven engineers, an office manager, a part-time administrator, a contracts supervisor and a working director. Domestic call-outs and repairs, and planned maintenance contracts on about forty small commercial sites. One trade account with a builders' merchant, four regular suppliers, and two subcontractors on the books.
The job recorded here: checking supplier invoices against what was ordered and what was delivered, before the weekly payment run.
Materials and parts arrive at Kestrel by three routes, and all three produce paper. An engineer walks into the merchant and leaves with a counter ticket. A delivery reaches the site and somebody signs a delivery note, sometimes an apprentice, sometimes the customer. An order placed by phone arrives with a note in the van. At the end of the month the merchant statement lists every one of those lines across every job.
The office manager checks supplier invoices against the order and the delivery note before the Friday payment run. In practice, and this is the firm's own description of its own week, that check happens properly on the large invoices and not on the rest. There is frequently no written order to check against, because the order was a phone call. So the classic three-way match of invoice to order to delivery is, at Kestrel, a one-way match: the invoice is checked against what somebody remembers agreeing.
The baseline, as an illustrative assumption. The firm handles roughly 260 supplier invoices a month, and takes the view that fewer than a third get checked against a document rather than against memory. That is Kestrel's own working figure for this example. It is not a measured study and it is not a statement about anybody else's business.
What that costs, in the firm's own words. A duplicated invoice gets paid twice. A price that does not match the rate agreed with the merchant gets paid at the higher figure. A quantity that was billed and never delivered gets paid in full. None of those announce themselves, and the firm's honest position is that it does not know how often they happen, which is the reason it wanted this looked at.
USE AI to point at disagreements between a supplier invoice, the order behind it and the delivery note, across every invoice rather than the large ones only.
DO NOT USE AI on subcontractor invoices under the Construction Industry Scheme. This is a separate ruling and it is not a caution. The split between labour and materials on a subcontractor's invoice sets the tax the firm deducts. HMRC puts that check on a person: "The contractor must always check and keep records to make sure that the part of the payment for materials supplied is not overstated" (CIS 340, paragraph 3.13). If the materials part is overstated, too little tax is deducted. Paragraph 3.13 places the check, and the consequence of getting it wrong, on the contractor as a person. Here that contractor is Kestrel. That is why these invoices are outside the tool entirely, with the reason written here so that nobody quietly widens the scope in six months.
DO NOT USE AI to decide whether to withhold payment or to open a dispute with a supplier. That is a judgement about a commercial relationship, and about a credit line the firm needs on site next week. Nothing in the invoice tells you any of that.
WAIT on the merchant statement reconciliation. See field 14.
The reason behind the first ruling, and it matters that it is the real one. The reason is not that a general AI tool sometimes gets a number wrong. It is structural, and the UK's audit regulator states it plainly: "LLMs are often not natively strong at computational tasks, as they lack semantic awareness" (Financial Reporting Council, March 2026). So we do not place a tool where the job is arithmetic. Reading a lot of documents and pointing at where two of them disagree, with its working shown, is a different job, and it is the one we placed a tool inside.
The job. For each supplier invoice, the tool reads the invoice alongside the order and the delivery note for the same supply, and reports where they do not agree: a quantity billed above the quantity delivered, a unit price above the rate on the order, a line that appears on the invoice and on no delivery note, or the same invoice number seen twice. Each report cites the document, the page and the line it came from. The office manager then opens each one.
The boundary, in five parts, and every one of them is testable.
Where the tool stops and Kestrel starts. The tool hands over a list of disagreements with citations. Everything after that point is a person: opening the cited line, deciding whether it is a real discrepancy, deciding what to do about it, and paying.
Permitted into the tool: the supplier invoice, the purchase order or written order confirmation, and the delivery note or goods received note for the same supply. The supplier's name and account number. The job reference.
Prohibited, and never entered: bank details, sort codes and account numbers for any supplier; the firm's own banking credentials; any staff personal data, including timesheets, payroll and anything about an individual's health or absence; anything relating to a supplier in an active legal dispute; and subcontractor invoices under the Construction Industry Scheme, which are out of scope under field 2 and are therefore out of scope for the data too.
Owner: the office manager, who runs the payment run and who is accountable for this job and for keeping this record current.
Access: the office manager and the working director, who authorises payments. Two people. The administrator and the contracts supervisor do not run it, because the person who queries a supplier and the person who releases the money have to be people the supplier can be pointed at by name.
Four of these five are real, public and dated, and anybody can read them.
Any of these changing is a trigger to re-check the record. See field 12.
One thing we will not put in this list. No vendor's marketing claim about its own AI appears anywhere in this record, and none was used to reach any decision in it.
Chosen: a business-tier assistant with a data-processing agreement and a setting that excludes the firm's inputs from training, running against documents that have been indexed first so that every statement it makes can cite a line.
Rejected, with the reason in each case:
Before the firm relied on this, we ran it over 200 supplier invoices that had already been settled. Three tests. Every flag the tool raised across all three was opened and checked by hand, which is how the false alarms below were counted. The figures are illustrative and belong to this example.
Test 1: does it catch errors that were known to be there? 6 errors in that set had come to light afterwards by other means, through a credit note or a supplier query. The tool flagged 5 of them. It did not flag the sixth, which was a price above the rate agreed with the supplier, because that rate existed only in an email between the director and the merchant's account manager, and no document in the tool's scope carried it. Result: 5 of 6, and the miss is a records problem the firm owns rather than a tool problem. The agreed rate is now written on the order.
Test 2: how often does it flag an invoice that turns out to be perfectly correct? This is the test the firm cared most about, and it is the one this tool did worst on. Across the same 200 invoices it raised 31 flags. 24 of them were false alarms. 19 of those 24 came from a single supplier whose delivery notes count in boxes of ten and whose invoices count in units, so a delivered quantity of 12 was compared against an invoiced quantity of 120 and reported as a disagreement every time. The unit conversion now sits in a fixed reference table that the arithmetic runs against, rather than being something the tool is expected to work out. On a re-run the same 200 invoices produced 12 flags, of which 5 were still false alarms. That is more than one flag in three, it is the level the firm now works with, and it is written here rather than smoothed, because the person opening those flags needs to know what to expect.
Test 3: can a person open every citation and find what the flag says is there? 31 flags, 31 citations. 3 of them pointed at the wrong line. The document was right and the line was not, and the person who opened it found it did not say what the flag claimed. Nothing was paid wrongly and nothing left the business, because the check at field 8 is to open the cited line rather than to read the flag and agree with it. That check is the only reason this was caught, and it is the reason the check is mandatory rather than recommended.
What changed because of Test 3. Opening the citation stopped being described as good practice and became the step the job cannot skip. The rule is now written into the runbook at field 10 as step 4, a wrongly cited flag is counted every month at field 11, and the record was changed on the day. It is in the change log at field 13 with that date against it.
Read those three results together, because that is the honest picture. The tool caught 5 of the 6 errors that were there to catch. It also raised 24 alarms about invoices that were perfectly correct, and it pointed at the wrong line 3 times. It is useful because a person opens every flag it raises, and it would be worse than useless if they stopped.
Runbook.
Fallback. If the tool is unavailable, or its output looks wrong, the payment run goes ahead on the previous manual spot-check exactly as before. Losing the tool costs coverage across the small invoices. It never stops Kestrel paying its suppliers, and it never puts the firm in the position of not being able to pay a merchant it needs on Monday morning.
Four numbers, read monthly by the office manager and the director.
These are what Kestrel measures. They are not a claim about what those numbers will say here, and they are certainly not a claim about what they would say in your business.
The queued question: should the same checking run against the monthly merchant statement, which lists every counter ticket and delivery note across every job?
The condition that would open it: not yet, and the reason is not about the tool. Most merchant collections at Kestrel are made without a written order, so there is nothing on the firm's side to compare the statement against. Until counter collections are raised against an order, a tool pointed at the statement would be comparing the merchant's document with the merchant's other document. The question is recorded here so that it is not lost, and the answer today is wait.
Three notes, because the shape of it is deliberate.
The tool is not the actor anywhere in this document. We placed a tool that points at disagreements between documents. Kestrel's office manager reviews. Kestrel's director authorises. When the tool got it wrong, a person caught it, because the job was built so that a person would.
The failure in field 9 is in the record on purpose, and it is one failure, not a catalogue. It sits where it belongs in the sequence, after the method is on the page, because that is where it means what it actually means: the control worked. We have not put it there because being open about AI earns anybody's trust, because as far as we can tell it does not. We have put it there because a record with nothing in it but passes is an advertisement, and because a control you cannot inspect is only a claim.
The honest "do not use AI here" at field 2 is a ruling, not a hedge. An entire class of document, the subcontractor invoices, is outside the tool, with the reason written down. If we could not tell you where the tool has no business being, we would not be much use to you in deciding where it does.
Kestrel Building Services does not exist. Its people, its suppliers, its invoice counts and every number in Part Two were invented for this example and are internally consistent with the assumptions the example states about itself. They are not market claims, they are not a promised result, and they must not be read as either.
Four of the five sources named in field 6 are real, are public, and were re-checked on 13 September 2026. The fifth is the tool provider's terms, which belongs to the example and is labelled that way.
What you see first here is the record itself, which is the same in any sector. The version filled in for a business like yours is one step in. When we work with you, we build one for your own.
Before any decision, there is a diagnosis. The findings-report extract shows what the audit found in the same firm's buying and paying: invoices approved against memory because no written order exists, agreed rates sitting in an email where the person checking cannot see them, and a class of subcontractor invoice that carries a tax judgement the law puts on a person. Each finding says what is happening, the root cause beneath it, what it costs, how sure we are, and the evidence it rests on. It reports what is true before it proposes anything. The whole extract is below.
This is a short excerpt from a Genesis Findings Report. A Findings Report is the document you receive after the audit and before anything is recommended or built. It records what is happening in each part of your business, why it is happening, and what it costs you. It does not yet tell you what to do. That comes later, and only once you agree the diagnosis.
This excerpt is the second half of the Specimen. The operating record you can also read shows one place where a tool was later put to work, and the boundary and the checks that hold it. This half shows the step before that: what the audit found in the first place. They are two views of the same illustrative engagement, for the same invented firm.
The firm (invented). Kestrel Building Services, an eleven-person heating, plumbing and mechanical maintenance firm. The same firm as the operating record. There is no real client here, no measured result and no promised one. Every figure belongs to this example and to the assumptions it states about itself.
The part of the business excerpted: buying materials and paying suppliers. In the full report this is one workflow section among several. One section is shown here so that you can read the shape of a finding without reading the whole audit.
Reviewed with (illustrative): the office manager, who runs the weekly payment run, and the working director, who authorises it.
Materials reach Kestrel by three routes, and all three produce paper. An engineer walks into the builders' merchant and leaves with a counter ticket. A delivery reaches the site and somebody signs a delivery note, sometimes an apprentice, sometimes the customer. An order placed by phone arrives with a note in the van. At the end of the month the merchant statement lists every one of those lines across every job.
Before the Friday payment run, the office manager checks supplier invoices against the order and the delivery note. In practice that check is done properly on the large invoices and not on the rest, and there is often no written order to check against, because the order was a phone call. The firm's own account of its week is that the three-way match of invoice to order to delivery is, for most invoices, a one-way match: the invoice is checked against what somebody remembers agreeing.
Baseline (the firm's own working figures, illustrative).
| What we measured | The starting point | How it was captured |
|---|---|---|
| Supplier invoices a month | About 260 | The firm's own count for this example |
| Proportion checked against a document rather than memory | Fewer than one in three | The firm's own estimate, not a measured study |
| Money paid in error each month | Not known | The firm cannot say, which is why it wanted this looked at |
The last row is the important one, and it is left as "not known" on purpose. The firm does not have a number, and the report does not invent one for it.
Four findings are shown from this workflow. Each is presented as it appears in the full report: what is happening, the root cause beneath it, what it costs, how sure we are, and the evidence it rests on.
Finding 1 — Invoices are approved against memory, not against a document.
| Field | Content |
|---|---|
| What is happening | Most supplier invoices are paid after a check against what somebody remembers agreeing, because no written order exists to compare them with. |
| Root cause | Orders are placed by phone or at the merchant's counter, so the price and quantity agreed are never written down before the invoice arrives. The check has nothing to check against. |
| Impact | A duplicated invoice, a price above the rate agreed, or a quantity billed and never delivered can all pass unnoticed. The firm cannot say how often, because there is no record of what was agreed. |
| Priority | Medium-term |
| Confidence | CONFIRMED. Described consistently by the office manager and the director, and visible in the firm's own purchasing paperwork reviewed during the audit. |
| Evidence | The firm's own description of the weekly payment run, and the absence of a written order behind the majority of the month's invoices. |
Finding 2 — The rates agreed with the merchant live in email, not on the order.
| Field | Content |
|---|---|
| What is happening | The prices agreed with the builders' merchant are held in email between the director and the merchant's account manager. They are not written onto the purchase order or held anywhere the person checking the invoice can see them. |
| Root cause | Rate negotiations happen director to account manager, and the outcome is never carried across into the ordering paperwork the office uses. |
| Impact | An invoice charged above the agreed rate looks correct to the person checking it, because the agreed rate is not in front of them. The overcharge is invisible at the point it would be caught. |
| Priority | Quick win |
| Confidence | CONFIRMED. The audit saw directly where the agreed rates are held, in the director's email rather than in any document the office uses to check an invoice. |
| Evidence | The email thread holding the agreed rates, and the ordering paperwork the office works from, which does not carry them. |
Finding 3 — Subcontractor invoices under the Construction Industry Scheme carry a tax judgement that must stay with a person.
| Field | Content |
|---|---|
| What is happening | Subcontractor invoices are split between labour and materials, and that split affects the tax Kestrel deducts. The materials figure has to be checked before the payment is made, and that check is done by whoever prepares it that month. |
| Root cause | The Construction Industry Scheme places the materials check on the contractor as a named responsibility (CIS 340, paragraph 3.13). It is a legal duty, not an administrative step a tool can take over. |
| Impact | An overstated materials figure means too little tax is deducted, and paragraph 3.13 puts the duty to check it, and so the responsibility for getting it wrong, on the contractor, Kestrel. This is a compliance exposure, not a time saving to be found. |
| Priority | Strategic (a boundary to protect, not a problem to fix) |
| Confidence | CONFIRMED against HMRC guidance, for the materials-check duty paragraph 3.13 establishes. The wider monthly-return and penalty regime is real but sits in other parts of the scheme and is not claimed here from paragraph 3.13. |
| Evidence | HMRC, Construction Industry Scheme: a guide for contractors and subcontractors (CIS 340), paragraph 3.13: "The contractor must always check and keep records to make sure that the part of the payment for materials supplied is not overstated." Published 31 March 2014, last updated 27 August 2026, checked 13 September 2026. |
Finding 4 — Duplicate and overcharged payments are probably occurring, but the firm cannot yet measure them.
| Field | Content |
|---|---|
| What is happening | On the routes where invoices are not checked against a document, the firm believes some invoices are paid twice or paid at the wrong figure. It has no count. |
| Root cause | There is no measurement in place. Errors that are not looked for do not announce themselves, so they are neither caught nor counted. |
| Impact | Unknown by amount. The point of this finding is the not-knowing, which is itself the risk the firm is carrying. |
| Priority | Medium-term |
| Confidence | HYPOTHESIS. This is a reasoned inference from the current state, not a measured fact. It becomes measurable only once a check is running and its results are recorded. |
| Evidence | Findings 1 and 2 above, and the firm's own inability to state a figure for money paid in error. |
Note the labels. Findings 1, 2 and 3 are marked CONFIRMED, because each was evidenced in the audit and cross-checked against a second source or against published guidance. Finding 4 is marked HYPOTHESIS, because it is an inference and not yet a measurement. You are never handed a guess dressed as a fact.
The picture in buying and paying is of a process that works on the large invoices, where attention naturally goes, and runs on memory everywhere else. The single thread beneath the findings is that the firm agrees things it never writes down, so the check at the end of the week has less to check against than it appears to. None of this is a failure of care by the people doing it. It is a records gap that leaves them checking against memory and carrying a cost they cannot size.
A Findings Report stops at the diagnosis. The decision and the build are a separate step, and you agree the diagnosis first. This excerpt shows where these particular findings led only so that you can judge the quality of the thinking, and the other half of the Specimen, the operating record, is that decision written out in full.
The reason to read the two halves together is that the finding you might expect a consultancy to seize on, the one that looks most like a case for putting a new tool in place, is the one where the honest answer was to change a habit and buy nothing. Telling you where a tool does not belong is as much of the work as telling you where it does.
Kestrel Building Services does not exist. Its people, its suppliers, its invoice counts and every number here were invented for this example and are internally consistent with the assumptions the example states about itself. They are not market claims, they are not a promised result, and they must not be read as either.
This excerpt carries no statistic about how common any of this is in real businesses. That is deliberate. The Specimen stands on the quality of one worked diagnosis, not on a claim about a population. The external authority named in Finding 3, HMRC's CIS 340, is real, public and dated, and you can go and read it.
What you see first in the Specimen is the method, the same shape in any sector. The version filled in for a business like yours is one step in. When we work with you, we produce one for your own.
Read the two halves together and one thing stands out. The finding that looks most like a case for putting a new tool in place, the messy invoice checking, is not where the money went. The agreed rates were fixed by writing them onto the order, a change the firm made itself. The subcontractor invoices were placed outside any tool, because the law puts that judgement on a person. Only one narrow job was left for a tool, and even there a person opens every flag it raises. Telling you where AI does not belong is as much of the work as telling you where it does. That is what you are judging when you read the Specimen.
The firm in this Specimen is invented, and so is everybody in it. Its people, its suppliers, its invoice counts and every figure were invented for this example and are internally consistent with the assumptions it states about itself. They are not market claims, they are not a promised result, and they must not be read as either.
The Specimen carries no claim about how common any of this is in real businesses. It stands on the quality of one worked diagnosis and one worked decision. Its external authorities are real, public and dated: HMRC, the Financial Reporting Council, the PCRT bodies and the Harber tribunal decision.
If the Specimen shows you the quality you are looking for, the next step is a free Discovery Call. Forty-five minutes, no pitch. We learn how your business runs and tell you honestly whether there is a real opportunity to help. You leave with a straight read on where you stand, whether or not you ever go further.
Book a Discovery CallFree, forty-five minutes, no pitch.
Email Genesis instead