Over the last two years we've shipped roughly thirty AI and automation projects for businesses between 20 and 500 employees. Some paid for themselves in six weeks. A few never should have been built. This is the honest ranking — what worked, what didn't, and how to tell the difference before you spend.
The scoreboard, ranked by payback speed
When we lined up all thirty projects and calculated months-to-payback, the pattern was almost embarrassing. The unglamorous projects won, decisively.
1. Document intelligence — median payback: 2.5 months
Every mid-size business has a person (usually three people) whose job is retyping. Supplier invoices into the accounting system. Bank statements into reconciliation sheets. Purchase orders into the ERP. Extraction models now read these documents with better accuracy than a tired human at 4 p.m., and the integration work is straightforward because the destination systems already have import APIs.
One distributor client processed 3,800 supplier invoices a month through a two-person data entry team. The extraction pipeline we built handles 94% of them untouched; the two staff members now manage exceptions and supplier disputes — work that actually needs judgment.
2. Demand forecasting — median payback: 4 months
If you hold stock — retail, food, manufacturing — forecasting is where AI earns its keep. Not because the models are exotic (they're often gradient-boosted trees, not deep learning), but because the baseline is so weak: most businesses forecast by "same as last month, plus a feeling."
The food delivery client in our case studies cut prepared-food waste 31% in one quarter. The model wasn't clever. The discipline of acting on it was.
3. Lead scoring and routing — median payback: 5 months
Sales teams waste their best hours on their worst leads. A scoring model trained on two years of CRM history — which leads closed, which ghosted — reorders the call queue every morning. The reps who trusted it closed measurably more. The reps who didn't provided a nice control group, which is how we know it works.
Where the money went to die
Three project types consistently underperformed, and we now warn clients away from them as a starting point:
- Customer-facing chatbots as project one. They touch your brand, they need constant tuning, and the failure mode is a screenshot on social media. Do them after you've built AI muscle on internal work, not before.
- "AI strategy" without a workflow. If nobody can name the specific task, the specific person, and the specific hour of the week the AI will save, the project has no floor under it.
- Building what you can rent. Custom-training a model for something a commercial API does for cents per call. We've talked clients out of six figures of unnecessary ML engineering.
The test we apply before any AI project
Before we write a line of code, a project has to pass three questions:
- Is the task frequent? A thousand times a month beats ten. Volume is what turns percentage improvements into money.
- Is being 95% right acceptable? AI is probabilistic. If a 5% error rate is a compliance incident, you need a human-in-the-loop design — which changes the economics.
- Does the data already exist? The best projects train on history you already have: invoices, orders, tickets. If step one is "spend a year collecting data," the payback clock starts a year late.
Pass all three and the project almost can't fail. Fail one and there's usually a cheaper, dumber automation that should come first — and "dumber but shipped" beats "smarter but stalled" every time.
Start smaller than feels impressive
The clients who got the most from AI all started the same way: one workflow, one metric, six weeks. The demo was boring. The savings weren't. Once the first project banks money, the second one funds itself — and by the fourth, the organisation has something more valuable than any model: the habit of automating.
If you want help finding your first workflow, that's literally what our AI Solutions discovery week is for.