Small and mid-sized companies run on spreadsheets, and that is not a failure of discipline. Before replacing anything, it is worth being honest about why the spreadsheet won in the first place.
Why spreadsheets are still useful
A spreadsheet is the fastest modelling tool most businesses will ever have. It requires no project, no vendor, and no specification. The logic is visible: anyone can click a cell and see the formula, which is more transparency than most business software offers. It is universally understood, easy to hand over, and it adapts the same afternoon the business changes. For scenario work, one-off analysis, and pricing questions, it is still the right answer.
Where spreadsheets become fragile
Problems begin when a file stops being a model and becomes an unowned system of record that several people depend on.
- Manual data entry re-keyed from another system every week, with no reconciliation
- Version drift: final_v3_new_JK.xlsx circulating alongside two other truths
- Formulas nobody can explain because the author left eighteen months ago
- Broken ranges after an insert, silently producing a wrong total that nobody notices
- No history, so you cannot see what a number looked like when the decision was made
- One person who is the process, and whose holiday stops the reporting cycle
The important observation: none of these are calculation problems. They are problems in the manual work that grew around the calculation — getting data in, checking it, distributing it, and reacting to it.
How dashboards, automation, and AI fit together
A realistic modernisation path rarely begins by replacing the spreadsheet. It begins by automating the surrounding steps while the calculation logic stays where the business already trusts it. Think in layers, each with a clear owner.
- Source layer: ERP exports, bank statements, invoices, forms, emails
- Structure layer: a spreadsheet or a small database with defined fields and types
- Automation layer: scheduled jobs that move, join, and validate data without a human
- AI layer: extraction, classification, summarisation, and anomaly flags
- Presentation layer: one dashboard or one short recurring report, not eleven
- Decision layer: a named human who owns the action
Automation handles the deterministic movement. AI handles the messy, text-shaped parts that rules cannot express — reading a supplier email, deciding which category a free-text request belongs to, writing a first draft of the commentary. Dashboards handle attention. Keeping these responsibilities separate is what makes the system debuggable when something goes wrong at 07:40 on a Monday.
Example: email request to tracker to summary
A distribution company receives around sixty customer emails a day into a shared inbox: delivery questions, complaints, price requests, document requests. Today someone reads each one and forwards it, and nobody can say how many complaints arrived last month.
The assembled version: a new email triggers the workflow, an AI step classifies it into one of five categories and extracts customer, order reference, and requested action, a rule rejects anything without a recognisable customer and routes it to manual handling, a row is written into a tracker with a timestamp and an owner, a draft reply is prepared for the routine categories, and a person reviews before sending. On Friday a second workflow summarises the week: volumes by category, average response time, unresolved items older than three days.
Nothing here is technically impressive. The value is that a previously invisible process now has numbers attached to it, and the summary lands without anyone building it.
Example: monthly numbers to red flags
The second example runs on the finance side. Once monthly figures are closed, a scheduled job pulls revenue, gross margin, operating costs, cash balance, receivables, and payables into a small structured table with twenty-four months of history. Deterministic rules then evaluate a fixed set of patterns: margin declining for three consecutive months while revenue grows, receivable days extending, cash falling while profit is positive, fixed costs growing faster than revenue.
The rules produce the flags — deliberately, because a threshold is auditable and a model is not. The AI step only turns triggered flags into a readable paragraph explaining what the pattern usually means and which three questions to ask. The owner receives one page: what changed, what it might mean, what to check. That is a decision support system, and it fits inside a spreadsheet, an automation tool, and one AI call.
A practical conclusion for smaller businesses
The most common way these projects fail is scope. A company tries to build the full stack at once, with a tool nobody owns after the consultant leaves. The alternative is to ship one narrow layer that removes one visible irritation, keep it maintainable by the person who suffers from the problem, and only then extend.
A healthy small-business system is boring: predictable inputs, a clear owner per step, an obvious place to look when something is wrong, and a documented way to switch back to manual. AI belongs inside such a system as one component — not as its foundation.
Worked examples of the finance calculations behind this kind of monitoring are available at SME Finance Helper