The Calder & Marsh file

This file is a case study and a teaching module at the same time. Three devices run through it:

• Terms you can tap. Words like grab, Role Card and debtor days are dotted. Tap one for a plain-language explanation. No prior knowledge of finance or of AI is assumed.
• Figures you can audit. Dotted numbers like £29.0k open the arithmetic behind them: the formula, the inputs, and where each input comes from. This is not decoration. Auditability is the method.
• Method notes. Green-edged boxes step outside the story to explain what the methodology is doing at that point and why. Read them and you can follow the discipline itself, not just this telling of it.

By the end you should be able to answer four questions: how a business finds where its time actually goes, how an AI hire is specified so a manager can hold it to account, how a term is run week by week, and how savings are counted so that nobody has to take anybody's word for anything.

History and ownership

Ray Calder was a toolmaker at a Hunslet engineering works; Brian Marsh sold fasteners for a Sheffield wholesaler. In 1987 they took a unit off Dewsbury Road with £9,000 between them and started supplying the fabrication shops Ray knew by first name. The business grew the unfashionable way: stock what the trade actually runs out of, open at half seven, and never make a customer wait twice. Ray bought the current Leeds site freehold in 1998, which is why the company still trades on thin margins without rent eating it alive. Marsh retired in 2008 and the Calders bought out his share with bank debt that took six years to clear.

The decision that shapes today's cost base came in 2014. A process-industry maintenance contract on Teesside justified a second depot, so they leased one. The contract halved in 2017 when the customer consolidated suppliers, but the depot stayed, partly pride, partly the fifteen accounts it had gathered. It now runs at roughly 60% of its designed throughput and carries £112k a year in premises costs, with a twice-weekly stock shuttle to Leeds that adds days to anything Teesside does not hold.

David Calder, Ray's son, became managing director in 2019 after twenty years in the business, most of them on the road. Ray, now 78, chairs the board and comes in on Tuesdays, walks the warehouse, and still knows more account history than the system does. The firm is good at what it does for an old-fashioned reason: availability and trust. A fabricator who rings at 8am gets the part on the afternoon van, and the counter staff know which grade of bolt the customer means before he finishes the sentence. What it has never been good at is writing any of that down.

People

Four people matter most to what follows. David Calder, 52, knows every top-40 account personally and closes on relationships and instinct; he is privately aware the business runs on four people's memories, including his. Margaret Whitaker, 63, joined in 1996, built every spreadsheet herself, and retires in April 2027; she is precise, trusted by the bank, and the only person who fully understands the month-end. Gary Pearson, 48, has twenty-two years in and carries the pricing logic for about 300 accounts in his head; first in at 7:15 to triage the shared inbox, last out most nights. Dawn Hargreaves, 44, is nominally the full-time credit controller; in practice she also administers fourteen supplier rebate agreements and covers the switchboard at lunchtimes. Her salary is £28k, squarely the Leeds market rate, a loaded cost of £32.1k. Her actual chasing time is about five hours a week, on Friday afternoons.

The trading engine

Product lines by revenue share: fasteners and fixings 34%, hand and power tools 22%, PPE and workwear 16%, abrasives and consumables 14%, janitorial and site supplies 8%, lifting gear and miscellaneous 6%. Customers: fabricators and small manufacturers around 45% of revenue, maintenance and facilities teams 30%, contractors 17%, and cash trade at the counters the rest.

In mid-2023 a national facilities-management group consolidated MRO supply for 30 of its sites onto Calder & Marsh, worth about £1.1m a year. It was won on service and price, carries 4 to 5 points less gross margin than the book average, demands monthly OTIF reporting, and pays slowly, as large FM and manufacturing customers typically do.

Systems: the honest mess

Sage 200 Professional, on premise, 2021 version, 14 user licences, runs the ledgers, stock and order processing. It is not the problem; what surrounds it is. "Pricing Master.xlsx", Gary's file, 41 tabs, is the real price list for special terms. "Stock Check.xlsx" layers Ian's reorder judgement over Sage's report. "Rebates 2025.xlsx", Dawn's, is the only record of fourteen supplier rebate agreements. Quotes are assembled in Outlook from a saved template, priced by looking across Sage, the Pricing Master and supplier PDFs; a won quote exists nowhere except the sent-items folder until someone re-keys it as an order. There is no CRM: account knowledge lives in the reps' notebooks and in Gary's and David's heads. The sales@ and accounts@ inboxes are worked first-to-open, with no ownership and no service standard. In the warehouse, pick notes print in batches and are confirmed by pen, and the archive is eleven filing cabinets. Teesside reaches Sage over a VPN that drops several times a week.

The five processes that run the business

Each is described with its numbers, because the numbers are where the engagement will later stand or fall.

Quote to order

About 240 quote requests a week. An emailed request sits in the shared inbox a median of 6.5 working hours before anyone opens it. Triage takes 4 minutes; pricing a typical five-line quote takes 18 minutes across three sources; typing it into the Outlook template takes 9. Around 30% of quotes need a special cost from purchasing, and that request waits a median of 1.2 working days for an answer. Quotes over £2,500 wait a median of 4.8 working hours for Gary's sign-off. Median end-to-end turnaround is 2.6 working days, against a buyer population that increasingly orders from whoever answers first. Roughly one quote in five gets chased, ad hoc, when someone remembers. About £124.8k of work is quoted each week; 42% converts. A won quote is then re-keyed into Sage, 6 minutes and one more chance to mistype.

Order to despatch

Keyed orders release to the warehouse printer in three daily batches at 10:30, 14:00 and 16:30, so an order keyed at 10:45 waits over three hours before a picker sees it. Everything funnels through one despatch checker, where the queue runs 25 to 40 minutes at peak. Carrier collection is 16:45; orders keyed after about 15:50 generally miss next day. A Teesside order needing Leeds stock waits for the twice-weekly shuttle. OTIF, measured only because the FM contract demands it, is 90.8% and has fallen in six of the last seven quarters. Order entry errors run at 25 per 1,000 keyed lines, about 103 lines a week; the checker and pickers catch perhaps two thirds, and the rest reach customers.

Invoice to cash

Invoices generate overnight from despatch confirmations, about 4,770 a month. Terms are 30 days; the book runs at 61 debtor days. Trade debtors stand at £1,772k against credit revenue of £10.6m. £948k of that, 53.5%, is beyond terms, in line with the national picture in which 52% of UK SME invoices are paid late, and £214k is more than 90 days old. Dawn's chasing is Friday afternoons: about five hours and 25 calls, and only accounts past 60 days get a call. There is no dunning sequence, no escalation path and no routine involvement from the reps. Thirty-seven accounts are on stop at any time, and sales override stop an average of six times a week. Bad debt written off in the last twelve months was £31k, against £11k two years earlier.

Purchasing and supplier rebates

Two buyers raise about 85 purchase orders a week from roughly 240 active suppliers. Trade creditors are £1,140k, 52 creditor days, which is to say the company leans on its suppliers to part-fund its debtors; one key fastener importer has already tightened terms after late runs. The fourteen rebate agreements were worth £126k of income last year and are tracked only in Dawn's spreadsheet. In 2024 one claim worth £9.4k lapsed past its claim window unnoticed, and two agreements missed volume thresholds by amounts small enough that a deliberate order-timing decision would likely have closed them, an estimated £7.8k foregone. Rebates appear in nobody's job description.

Customer claims and credit notes

118 credit notes a month, up from 86 two years ago, average value £64. A claim waits a median of 2.1 days for the warehouse to pull the paper pick note, then claims over £150 wait 3.2 days for Margaret's approval. Median end-to-end is 9 working days, during which the customer typically withholds payment on the whole invoice, feeding the debtor book. Internal coding attributes 48% of credits to keying and picking errors, 27% to price mismatches between Sage and the Pricing Master, and 25% to genuine returns and transit damage. Each claim absorbs about 55 minutes of staff time, and about 40% involve a re-delivery.

Performance and measurement

Eight quarters. Revenue, margin, debtors and overdraft come from management accounts. Quote turnaround and conversion were reconstructed for this profile from four two-week samples of sent items, machine-read and matched to Sage order records; the company does not routinely measure either, which is itself a finding. The error baseline is derived from credit-note attribution.

The story the table tells is coherent and quietly worrying. The 2023 FM win lifted volume about 13% over two years, diluted gross margin by 1.1 points through mix, and loaded a manual engine that had no slack. Every service and control metric has degraded in step: quotes slower, conversion lower, errors and credits up, OTIF down, and the extra revenue arriving as debtors rather than cash, so peak borrowing has grown £470k in two years against flat facility headroom.

The board pack each month contains: sales versus last year, gross margin percent, the overdraft balance, total aged debt as a single number, and headcount. It does not contain quote turnaround, conversion, error rates, OTIF, credit note counts, rebate status, or any measure of queue times. Nothing in the pack would have surfaced any trend in the lower six rows of Exhibit 4.

What it costs them

Every figure below is a multiplication of numbers already stated, and deliberately understated.

• Cash tied up in debtors. Each debtor day holds £29.0k of cash. Against the company's own 2023 performance of 53 days, today's 61 ties up an extra £232k, costing £21.5k a year at the 9.25% overdraft rate. Against a 45-day level that a trade distributor on 30-day terms can realistically hold, the excess is £464k and £42.9k a year.
• Errors, claims and credits. 1,416 credit notes a year at 55 minutes each, plus re-delivery carriage, is £27.6k of hard, auditable cost. Industry estimates of £40 to £100 per order error, vendor-published and treated with caution, would put the true figure at £64k to £142k. The £27.6k is what is carried.
• Quotes lost to slow turnaround. Conversion has fallen five points while median turnaround rose from 1.5 to 2.6 days. Attributing a conservative three of those five points to speed, the loss is £53.4k of gross margin a year. The full five points would be £88.9k.
• Rebate leakage. £9.4k documented as lapsed in 2024; up to £17.2k including the near-miss thresholds. The £9.4k is what is carried.

Conservative annual total: £112k. Less conservative, using the 45-day benchmark, industry error costs and the full conversion decline: up to £213k. Against profit before tax of £260k, the conservative case is 43% of the bottom line, recurring, before any growth.

The sharper point is resilience rather than profit. September's peak borrowing of £1,080k against the £1,250k facility leaves £170k of headroom, barely more than one month's gross payroll. One large customer slipping from 60 to 90 days would take the company to its facility limit with wages due.

Why now

Margaret Whitaker retires in April 2027. She is the only person who can run the month-end, the only person who fully reads the debtor book, and the author of every spreadsheet that glues Sage to reality. Eighteen months is exactly long enough to capture what she knows into processes that do not depend on her, and exactly short enough that waiting a year makes it impossible. Ray's view, delivered on a Tuesday in his usual register, is that the business never used to borrow to pay wages and he does not see why it should start now; he is not wrong, and the overdraft trend line says so.

What actually rattled David was Hartside Engineering. A £310k-a-year account, solid for a decade, drifted quietly: statements went out, no calls were made because the account had never needed them, and in August 2025 Margaret found £86k of it at 94 days during her facility review for the bank. The money came in, after three awkward weeks and a conversation that damaged a good relationship, but David has done the arithmetic. Nothing in the monthly pack would have caught it, Dawn's Friday afternoons cannot watch 1,100 accounts, and the next Hartside might not pay.

The engagement began with a telephone call in September 2025, a month after the Hartside discovery, on an introduction from the bank's relationship manager. David asked TCEL for a view of the whole business, not of its software. TCEL's position from the first meeting was the one it holds everywhere: AI belongs inside your strategy, not on top of it, and nothing would be bought until the analysis said where intervention pays.

The principal spent three days on site and a week in the management accounts, the sent-items samples and the Sage records. The read, stated as measured findings and ranked by what each leak costs:

1. Working capital is the binding constraint. The book runs at 61 debtor days against 30-day terms, 8 days worse than the company's own 2023 performance. That slide ties up £232k of cash. The deeper finding is resilience: one large customer slipping from 60 to 90 days would take the company to its facility limit with wages due.
2. The revenue engine is saturated. A 98% load on the sales office, visible in Exhibit 3, and its symptoms visible in Exhibit 4: turnaround up from 1.5 to 2.6 days over eight quarters, conversion down five points, a conservative £53.4k a year of gross margin attributed to speed alone.
3. Errors cascade into cash. 25 errors per 1,000 keyed lines; credit notes up from 86 to 118 a month; each disputed invoice typically withheld in full while the nine-day claims process runs, feeding the debtor book.
4. Income nobody owns leaks. Fourteen rebate agreements worth £126k a year live in one spreadsheet, and £9.4k of it lapsed unnoticed last year.
5. The business cannot see itself. Not one of the six degrading rows in Exhibit 4 appears in the board pack.

Three observations a senior operator adds, beyond process. Concentration has changed the shape of the risk: the top 40 accounts take 54% of revenue and hold most of the overdue book, so the risk is now customer-shaped rather than demand-shaped, and the credit control was designed for neither. The knowledge cliff is dated: Margaret retires in April 2027, Gary carries the pricing logic for 300 accounts in his head, and the firm's real asset, availability and trust, has never been written down. The company has pricing knowledge but no pricing control: a 41-tab spreadsheet is the real price file, 27% of credit notes are price mismatches, and margin has drifted 1.1 points. There is margin to recover without raising a single price, but only once there is one source of truth.

The decision logic follows from the split. Invoice to cash is the clearest staffing case, with the fastest cash effect and the lowest change burden, because it runs through one small team. The quote-and-order engine is the second, and relieving it protects the revenue line while the first intervention repairs the cash line. TCEL therefore recommended two interim hires rather than a systems project, and commissioned GRABS to run a Process Diagnostic on invoice to cash before anything was bought.

One process, one week. From Monday 20 to Friday 24 October 2025, GRABS measured invoice to cash from the company's own records: the Sage aged listing, the accounts@ mailbox timestamps, Dawn's call log, and Credits & Claims.xlsx. Nothing was recalled in a workshop. In the sample week the company raised about 1,100 invoices, opened 27 claims and overrode the stop list six times, all in line with its running averages.

Below is the grab map: the five points in this one process where the work waits. Tap each row for where it waits, how many hands touch it, and what the wait costs at the company's own numbers.

The surprise finding. David expected the diagnostic to confirm that his customers pay slowly. What it showed instead is that the largest single identifiable blocker that week was his own company: £294k of overdue money, 31% of the overdue book, was attached to invoices where the customer was waiting for Calder & Marsh, for a copy invoice, a POD, or an answer to a claim. The slowest payer in the invoice-to-cash process was the process.

The diagnostic closed with two Role Cards and a readiness list. Everything it produced was the company's to keep, hire or not.

Signed by David Calder and Margaret Whitaker on Friday 31 October 2025, before either Worker started. GRABS counts events; the client prices them, in advance, in writing. Every multiplier on every later report appears on this sheet, fixed at signature. The 9.25% overdraft rate, for instance, is the facility letter's rate on the day of signing, and it does not float with the base rate afterwards: a signed sheet is a signed sheet.

Each hire is proposed on one page, the Role Card: six parts, in the language of a job spec, because that is the language a manager can already assess, argue with and hold to account. Both terms ran 90 days from Monday 3 November 2025.

Onboarding: the vault, built in days

Days 1 to 4. Into the vault went: the Sage exports (customer master, terms, aged listing, open orders, price books, SKU file), the Pricing Master's 41 tabs, the rebates and claims spreadsheets, the quote template and sixty sent quotes, Dawn's best chase letters, the facility letter, the FM contract terms and portal dates, and five capture sessions: three half-days with Gary on pricing logic, one with Dawn on accounts and rebates, one with Margaret on the month-end. The connector was built the same week: a nightly Sage export into the vault and a draft-order import back, through Sage's standard interfaces, under one named user licence, owned by Pete Nowak. It failed twice during the term; both failures were caught by the Monday Report's data-freshness line, and neither reached a customer.

Day 5 found the problem the glamourless way: 243 of 1,100 account records had no working invoice email address or a dead contact, and 57 were duplicates. Nothing intelligent can chase a wrong address. Weeks 1 and 2 became a master-data sprint, the Worker drafting corrections, Lynn and Dawn verifying, before a single reminder went out. It was not in the plan. It is in every plan now.

30-day review, Monday 1 December 2025. Vault complete and attested by Margaret; both Workers live; every overdue account now inside a reminder sequence for the first time in the company's history. Debtor days 60.1 against 61.0 at baseline: two weeks of sending proves the machinery, not the outcome, and the review said so.

The week-6 exception, resolved. On Wednesday 10 December David chaired a thirty-minute meeting with Gary, the sales office and both Worker approvers. The Pricing Master in the vault was declared the single source of truth with Gary its named owner; the offline tabs were merged in; parallel keying stopped; the two staff who had been double-keying took on the draft-review and release role instead. The fortnight of double work and the eleven duplicates are in the record. The cause was trust, not technology, and the fix was a chaired meeting, not a patch.

The same month produced a quieter correction. A top-40 account told its rep that the stage-2 reminder wording read cold. Margaret softened two sentences and re-signed the sequence wordings on Friday 19 December. The sequences are hers, not the Worker's; that is what a signature means.

60-day review, Monday 5 January 2026. Debtor days 57.4. 55% of email orders auto-captured, with an error rate of 6 per 1,000 captured lines against the 25 per 1,000 keyed baseline. Release authority to £750 on the clean list granted, as the Role Card provided.

The Worker's own error, on the record. On Tuesday 20 January a stage-2 reminder went to Denholm Fabrications, £6.2k, hours after their BACS payment had arrived. The remittance carried no reference, so it sat unallocated while the sequence fired. Denholm's finance manager replied sharply, and was right to. The correction and apology went out the same morning under Dawn's name, and the rule changed that day: no reminder leaves on any account with an unallocated receipt anywhere on the ledger until a human has matched it, and the receipts sweep moved to 8:30 daily. One error in roughly 2,100 sends over the term is a better record than the Friday call list ever had. It is still one more than the report is allowed to hide.

90-day review, Monday 2 February 2026. Debtor days 55.2 against 61.0; conversion 44.0 against 42.0; three catches on the term's log; one sending error, corrected and rule-changed; both Role Cards performed inside their authority grades with no breach. Renewal tabled as a decision, not assumed.

• Claims and credit notes were measured and deliberately left alone. The nine-day claim cycle depends on paper pick notes in eleven filing cabinets, and no Worker can pull paper. Credit notes fell only from 118 to about 109 a month, the share caused by keying errors on captured lines. The claims process, and the warehouse paper that gates it, is a next-term question.
• Phone orders were untouched. 380 a week are still keyed live, so the blended error rate improved only from 25 to about 20 per 1,000 lines. The captured lines run at 6 per 1,000; most lines are still keyed.
• The stop-list control holds only where the Worker is. Captured orders cannot bypass a stop, which is how CM-61447 was held. A phone order keyed live into Sage still can, on a verbal word, exactly as before. Closing that route is a Sage workflow change, a management control rather than a Worker task, and it is on the renewal agenda.
• The head of the debtor book did not move. The two largest payers run payment-run policies that no reminder sequence alters. Their terms are a commercial negotiation: TCEL's territory, not a Worker's.
• The webshop is still 2016 vintage. Its orders arrive as emails; the Worker captures them in a minute instead of five. That is a patch on an old pipe, not a new pipe.
• And one failure, on the record. Weeks 5 and 6 produced a fortnight of double work: parallel keying beside the Worker's capture, eleven duplicate drafts, two reaching the warehouse, one shipped twice and credited. The cause was trust, not technology, and the fix was a chaired meeting, a single source of truth declared, and two jobs redesigned.

Teesside, its pick errors and its shuttle delays, was out of scope throughout and remains a question for the strategy, not for this term.

On Monday 2 February 2026 the handover was attested: Margaret signed the record confirming the vault complete and current, every Work Log entry attributable, and the playbooks run once under supervision by Jack Ryder without the Worker's help.

Calder & Marsh keeps, outright:

• the vault: the account master, cleaned and deduplicated; the pricing file as a single source with a named owner; the rebate calendar with every threshold and window dated; the claims file synced nightly; the tone library; and the captured sessions with Gary, Dawn and Margaret, written down for the first time in the company's history;
• the playbooks: the three-stage dunning sequence with Margaret's signed wordings, the escalation routes, the order-capture exception rules, the quote-assembly rules, the rebate calendar procedure;
• the records: the full Work Log, the catch log, the approval-queue history, and the baseline-to-day-90 measurements;
• the trained process itself, which now runs on written rules rather than four people's memories.

The handover notes carry one open item. One member of the sales office was still keeping a private list of "his" email orders in February, checked each morning against the Worker's capture. Nobody has asked him to stop. The note records only that trust is adopted at different speeds.

"It does the chasing I never got to. Fridays I do the ones that need a human." Dawn Hargreaves, at the handover meeting, as minuted

The renewal decision David faces is a real choice, and it is put to him as one. Renew both Workers at £3,300 a month and extend into the claims process. Renew the Credit Controller Worker alone. Or renew neither: take the vault and the playbooks and run the sequences in-house, which the handover genuinely equips the company to do, at the cost of the hours it was short of in the first place. Renewal is a choice, never a default.

The cash position is no longer the binding constraint it was in September. £168k released from the debtor book, added to £170k of facility headroom, roughly doubles the buffer between the company and its limit, and the next Hartside now surfaces in days through the exception list rather than in a facility review months late. Ray's test, that the business should not borrow to pay wages, is back within reach without a penny of new revenue, and the board pack now carries the six rows that were missing: debtor days, quote turnaround, conversion, error rate, OTIF and credit notes, drawn from the same records the Workers keep as a side effect of working.

The Margaret problem has changed shape. Her month-end and her reading of the debtor book are partly written down for the first time, and the dunning process no longer depends on any individual's Friday. Fourteen months remain before she retires, which is enough, provided the capture continues deliberately: the remaining work is her judgement on the bank, the audit and the year-end, none of it yet in the vault. That, not more automation, is the next knowledge task.

What the term has really bought is the standing to address the strategy questions from evidence. The FM contract can be renegotiated with measured OTIF and a known cost to serve. The Pricing Master, now a single controlled source, makes the 1.1 points of drifted margin recoverable account by account. Teesside can be decided, kept, shrunk or closed, against numbers rather than pride. None of these is an AI question. They are owner's questions, which is where they always belonged; the engine just had to stop consuming the owners first.

Everything you just read is one method applied once. It has three steps, and you have now seen each of them working.

And underneath all three, the counting discipline from Part seven: events counted by the supplier, prices signed by the client in advance, and three columns of different evidence that are never blended into a number you would have to take on faith.

Run it on your own numbers

Three questions, straight from this file to your business:

1. What is your debtor-days number today, and what was it two years ago? If you do not know either, that is the first finding.
2. Who owns your shared inboxes, and how long does a copy-invoice request wait for an answer? The Calder & Marsh diagnostic found 31% of the overdue book waiting on the company's own replies.
3. If your biggest account went silent at 90 days, what in your month-end pack would catch it before your bank review did?

The board pack is a ready-to-edit presentation for taking this case to your own board, in your voice, with every figure left as a slot for your numbers.

Or write to [email protected].

Calder & Marsh is fictional; the world it trades in is not. The company's shape, its pains and the engagement's results were calibrated against the published sources below, graded by how much weight each will bear. Where the simulation needed a number the public record does not supply, it stayed inside the nearest published range, and the ranges are shown.

Official and statutory

• Department for Business and Trade, large businesses' payment practices and performance statistics, 2025 edition: large payers average 32 days to pay with 15% of invoices late; manufacturing payers, the customer base of an MRO distributor, average 45 days with 21% late.
• GOV.UK, employer National Insurance from April 2025 (15% above £5,000) and auto-enrolment minimums (3%): the basis of every loaded rate in the file.
• Bank of England base rate and standard UK SME overdraft pricing of base plus 3 to 8 points: the 9.25% facility sits inside the band.

Large-panel and sector research

• FSB and GoCardless late-payment research: 52% of UK SME invoices paid late; around £22,000 of overdue invoices owed to the average small firm at any time.
• Xero Small Business Insights: UK small-business invoices paid an average of 8 days late against agreed terms.
• Atradius Payment Practices Barometer, UK: roughly two thirds of UK firms report late B2B payment; 30-day terms with a book running at 55 to 70 days is a recognised problem state.
• UK SME benchmark aggregations of filed accounts: wholesale and distribution at a median 24% gross and 5% net margin; Calder & Marsh at 28.5% and 2.3% is a technical-range distributor with a squeezed bottom line.
• Consero and APQC mid-market benchmarks: a finance function at 1.2 to 2% of revenue; the simulated £178k on £11.2m sits at the top of the range, as stated in the profile.
• Leeds salary data (Indeed, Totaljobs, CV-Library): credit controllers at £23k to £30k; Dawn at £28k.
• APQC order-management benchmarks: 1 to 3% of manually keyed orders contain errors; the simulated 2.5% of lines is at the loaded-engine end of that band.

What AI delivers, and how often it fails

• Brynjolfsson, Li and Raymond, NBER Working Paper 31161, the strongest independent study in this field: AI assistance lifted support-agent productivity about 14% on average.
• Forrester Total Economic Impact study of an AR-automation platform (vendor-commissioned): around 25% DSO improvement as a multi-year outcome, not a first quarter's.
• AR-automation industry data, discounted for vendor interest: a 10 to 15% relative DSO improvement is the defensible first-quarter band; headline claims of 75% reductions were discarded.
• Order-capture case studies (Conexiom, Emporix, large mature deployments): error rates falling from 2 to 5% toward under 1% on automated lines, over months; straight-through rates of 50 to 70% in early deployment, 80%+ only at multi-year maturity.
• RAND Corporation, 2024: over 80% of AI projects fail, roughly twice the rate of ordinary IT projects, most often on misstated problems and poor data quality. The day-5 master-data sprint in Part six is that finding, lived.
• S&P Global Market Intelligence, 2025: 42% of companies abandoned most of their AI initiatives that year; shadow processes and change management, not models, are the usual cause. The week-6 double-keying failure is that finding, lived.
• The Klarna reversal, 2024 to 2025, as reported by Bloomberg and others: aggressive AI substitution of service staff partially reversed after quality complaints. The reason both Role Cards keep humans on approvals, escalations and everything ambiguous.