Using AI to find margin leaks you already have

5 days ago
3 min read

Updated: 2 days ago

Introduction

Most conversations about AI in a business start with revenue, and most of the available money is on the other side. A firm turning over two million at a nine per cent margin does not need more work; it needs to know which of its existing work is quietly unprofitable, and it almost never does.

The reason is not carelessness. The evidence sits in thousands of invoices, timesheets, purchase orders and job records, and reading all of them at customer level and job level is a week of work nobody has. This is the exact shape of problem where AI earns its cost, because the finding is arithmetic performed at a volume a person will not sustain.

1. Using AI to find margin leaks means analysing what you already sold

Start with history, not forecasts.

Twelve months of invoices, costs and time records contain the answer. You are not predicting anything; you are measuring what each job and each customer actually returned. This is the least speculative AI application in a business and usually the most profitable.

2. Job-level profitability is where the biggest surprises are

Averages conceal it.

A firm with a healthy overall margin routinely has a quarter of its jobs at or below break-even, offset by a handful of very profitable ones. Nobody sees this because the monthly accounts report a single blended figure, and the blend is the problem.

3. Customer-level profitability changes who you chase

Revenue and profit rank differently.

Your largest customer by turnover is frequently not your most profitable, once discounts, payment delays, rework and unbilled extras are counted. Ranking customers by contribution rather than revenue tends to reorder the list substantially, and it changes where the sales effort goes.

4. Look for the leaks that repeat

One-offs matter less.

A single bad job is an event. The same underrun on the same job type every month is a pricing or estimating fault, and it is worth far more to find. Pattern detection across many records is precisely what the tooling is good at and what manual review misses.

5. Unbilled work is the most common single leak

It looks like generosity.

Extras agreed verbally, additional visits, revised drawings, small variations. Each one is minor and none are invoiced, and across a year they are frequently the difference between the margin you have and the margin you assumed. Comparing scope records with what was billed surfaces them quickly.

6. Check discounts against who authorised them

Authority drifts.

Discounts given to win work become permanent, get applied by more people than intended, and are rarely reviewed. A list of every discount by customer, size and approver is uncomfortable reading and usually recovers real money within a quarter.

7. Input cost changes that never reached your prices

The silent one.

Material and subcontract costs move continuously and quoted prices do not. Matching your current purchase costs against the assumptions inside your price list identifies the lines you are now selling below the margin you intended, without anyone having made a decision.

8. Validate every finding before acting on it

The output is a hypothesis.

Take the three largest apparent leaks and check them against the actual records by hand. Some will be coding errors, misallocated costs or a job that was always intended as a loss leader. Acting on an unverified finding damages a customer relationship for nothing.

9. Fix the process, not just the instance

Otherwise it returns.

Each confirmed leak points at a process: an estimating assumption, a variation procedure, a discount authority, a price review interval. Correcting the instance recovers money once; correcting the process recovers it every year, and that is where the return actually comes from.

Do the exercise on a schedule rather than once. Quarterly is enough for most firms, and the second run is far faster than the first because the questions are already defined.

Conclusion

Point the analysis at your own history, because the evidence of where margin is lost already exists.

Measure profitability at job and customer level rather than trusting the blended figure, look for leaks that repeat rather than one-off bad jobs, check unbilled extras and authorised discounts specifically, compare current input costs against the assumptions in your price list, verify the largest findings by hand before acting, fix the process that produced each leak rather than only the instance, and repeat the exercise quarterly.