A payback period is the only AI metric a finance director needs, and almost no vendor will give you one. Below are seven workflows costed end to end. The arithmetic is shown so you can replace every assumption with your own — which is the point, because the numbers that matter are yours, not mine.
The method — four inputs, one number
Payback in months = build cost ÷ (monthly benefit − monthly run cost). That is the whole model. What makes it honest is discipline about the four inputs:
- Build cost — integration, configuration, data cleanup, training and the internal time your own staff spend. Internal time is the line most people omit, and it is typically 30% of the total.
- Monthly run cost — model API usage, hosting, licences, and the human who handles the exception queue. Budget for the exception queue; it never reaches zero.
- Monthly benefit — only what you can see in the accounts or in a timesheet. If you cannot name the account line it lands in, it is not a benefit, it is a hope.
- A confidence haircut — halve the benefit before you approve. If it still clears, you have a real project.
Assumptions used throughout: a 40–200 person firm; loaded cost of an accounts or admin clerk KES 400/hour; sales executive KES 560/hour; senior manager KES 1,560/hour; overdraft cost 16% per annum; ~KES 130 = USD 1. Figures are rounded.
The seven, ranked by payback
| Workflow | Build (KES) | Run / mo | Net benefit / mo | Payback | Clears 90 days? |
|---|---|---|---|---|---|
| Sales-lead follow-up | 200,000 | 10,000 | 86,000 | 2.3 mo | Yes |
| Meeting actions & task tracking | 90,000 | 4,000 | 34,000 | 2.6 mo | Yes |
| Management reporting | 160,000 | 5,000 | 61,000 | 2.6 mo | Yes |
| Accounts-receivable follow-up | 180,000 | 8,000 | 63,000 | 2.9 mo | Yes |
| Customer-complaint management | 150,000 | 6,000 | 52,000 | 2.9 mo | Yes |
| Tender-document analysis | 240,000 | 14,000 | 80,000 | 3.0 mo | Borderline |
| RFQ & quotation preparation | 220,000 | 12,000 | 63,000 | 3.5 mo | No — phase it |
Five of seven clear 90 days on these assumptions. That is a realistic hit rate, and a list where all seven cleared would be a list you should not trust.
1. Sales-lead follow-up — 2.3 months
The mechanism. Quotes sit in a sheet with a status column updated when someone remembers. Sales staff work the deals they recall — the recent and the loud ones. Everything else expires quietly. The automation schedules a ladder against every open quote and drafts each message for approval.
The arithmetic. 60 quotes a month, average value KES 180,000, gross margin 28%. Follow-up coverage today is about 40% of quotes; a ladder takes it above 90%. Conservatively, the recovered half of the pipeline converts at 3 percentage points below the attended half — call it a net uplift of 1.8 points on the whole book. 60 × 180,000 × 1.8% = KES 194,400 of revenue, at 28% margin = KES 54,400. Add 14 hours a month of manual chasing at KES 560 = KES 7,800, and roughly KES 34,000 of previously lost repeat business from customers who simply were not contacted. Total gross ≈ KES 96,000; net of KES 10,000 run cost = KES 86,000.
Where it goes wrong: unreviewed messages. Route every draft through the owning salesperson for one click. The moment a customer receives something that reads as automated, the channel is worth less than it was before you started.
2. Meeting actions and task tracking — 2.6 months
The mechanism. Weekly management and production meetings generate decisions that are remembered differently by each attendee. Transcription plus extraction produces a decision log and an owner-dated action list.
The arithmetic. Six recurring meetings a week; on average 25 minutes of each is spent re-establishing what was agreed last time. 6 × 25 min × 4.3 weeks ≈ 10.75 hours a month, weighted heavily to senior attendees — call it KES 1,200/hour blended = KES 12,900. Add roughly 15 hours a month of chasing actions that were never written down, at KES 560 = KES 8,400. Add the value of decisions that no longer quietly lapse — measured conservatively at one recovered action a month worth KES 17,000. Gross ≈ KES 38,000; net KES 34,000.
Where it goes wrong: actions land in a tool nobody opens. Push them into whatever people already use daily, even if that is email.
3. Management reporting — 2.6 months
The mechanism. Month-end reporting is assembled by hand from several systems, arrives around day 15, and is then argued about. Automation reads the source tables, produces the pack on day 3, and drafts commentary on what moved.
The arithmetic. 3.5 days of finance and admin effort per month at 8 hours × KES 400 = KES 11,200, plus 6 hours of a senior manager reviewing and correcting at KES 1,560 = KES 9,400. The larger number is decision latency: management currently acts on a 6-week-old picture. Costing one mistimed correction a quarter at KES 135,000 gives KES 45,000 a month. Gross ≈ KES 66,000; net KES 61,000.
Where it goes wrong: two departments define a metric differently and the pack is disbelieved. Agree definitions in writing before building anything.
4. Accounts-receivable follow-up — 2.9 months
The mechanism. A reminder ladder driven by the ageing report, escalating in tone, with disputes routed to a person rather than chased.
The arithmetic. This one is mostly working capital, not labour. On KES 9,000,000 of monthly revenue with DSO of 68 days, cutting DSO to 56 releases 12/30 × 9,000,000 = KES 3,600,000 of cash, once. Valued at a 16% overdraft rate, that is 3,600,000 × 0.16 ÷ 12 = KES 48,000 a month of avoided finance cost. Add 22 hours of clerk chasing at KES 400 = KES 8,800, and reduced write-offs of roughly KES 14,000 a month from invoices that were previously chased too late. Gross ≈ KES 71,000; net KES 63,000.
Where it goes wrong: chasing a disputed invoice. Suppress any account with an open dispute flag, and make raising that flag trivially easy for your team.
5. Customer-complaint management — 2.9 months
The mechanism. Complaints arrive across email, phone and WhatsApp, are acknowledged inconsistently, and are resolved by whoever notices. Automation classifies, assigns severity, routes to an owner and acknowledges immediately.
The arithmetic. 90 complaints a month; 35 minutes each of coordination overhead saved = 52 hours × KES 400 = KES 20,800. Retention is the real number: if 2 of those 90 customers were previously lost to slow handling, and average annual gross profit per customer is KES 168,000, retaining them is 2 × 168,000 ÷ 12 = KES 28,000 a month. Add roughly KES 9,000 of avoided goodwill credits. Gross ≈ KES 58,000; net KES 52,000.
Where it goes wrong: ambiguous categories. If your own staff argue about whether something is "quality" or "delivery", fix that first — on paper, in an afternoon.
6. Tender-document analysis — 3.0 months (borderline)
The mechanism. Tenders arrive as long PDFs. Extraction produces scope, deadlines, mandatory documents, evaluation criteria and disqualifying clauses, plus a go/no-go brief.
The arithmetic. 12 tenders a month; 3.5 hours of senior review each reduced to 1 hour = 30 hours × KES 1,560 = KES 46,800. Add avoided wasted bids: if 3 a month were previously pursued and were never winnable, at 6 hours of preparation each, that is 18 hours × KES 1,560 = KES 28,000. Add roughly KES 19,000 for compliance misses that previously caused disqualification. Gross ≈ KES 94,000; net KES 80,000.
Why it is borderline: the build carries genuine document-handling complexity and the benefit is concentrated in senior time, which is real but harder to bank. Approve it if tender volume is rising; defer it if it is flat.
7. RFQ and quotation preparation — 3.5 months (does not clear 90 days)
The mechanism. An enquiry becomes a draft quote against the price book, ready for review.
The arithmetic. The estimator saves 30 hours a month at KES 500 = KES 15,000. Quote turnaround falls from a day to under two hours, which on 60 quotes at KES 180,000 and 28% margin is worth roughly KES 60,000 a month at a conservative 2-point win-rate uplift on contested quotes. Gross ≈ KES 75,000; net KES 63,000 against a KES 220,000 build.
What to do about it. Do not abandon it — phase it. Build it for your three highest-volume product families only. The build drops to roughly KES 120,000 and captures around 65% of the benefit, giving payback near 2.9 months. Extend coverage from the savings. This phasing trick works on most workflows that miss the threshold.
Four ways these numbers go wrong
- Counting saved hours as saved money. Freeing 22 clerk-hours reduces cost only if the hours go to something you would otherwise have hired for, or if headcount actually changes. Otherwise the benefit is capacity, which is real but does not appear in the P&L. Say which one you are claiming.
- Omitting the exception queue. Straight-through processing of 90% means 10% needs a person. Staff it in the run cost or the workflow silently fails.
- Attributing everything to the automation. If sales rose in the quarter you launched a follow-up ladder, some of that was the market. Hold a control group where you can — one region, one product line.
- Ignoring maintenance. Price books change, document formats change, people leave. Budget 15–20% of build cost annually or your working automation becomes next year's write-off.
How to use this
Copy the table, replace every figure with your own, and apply the 50% haircut. Anything still under four months should be approved this quarter. Anything over eight months is not a bad idea — it is a later idea, and it usually becomes a good one once an earlier workflow has cleaned up the data it depends on.
