AI ROI: What Return to Realistically Expect from Custom AI in the First 90 Days

Photo: Calculator, Pen and Calendar by photosteve101, CC BY 2.0. Cropped.

Why most businesses can't say whether AI paid off

Ask companies that bought AI tools whether they got a return and many give a vague answer. The usual reasons: no baseline was measured before launch, the goal was general ("be more efficient"), the tool was never connected to real processes, or the benefits were real but nobody turned them into money. AI ROI is not mysterious. It needs the same discipline as any other investment.

The AI ROI formula

The basic formula:

ROI (%) = (Value created − Total cost) ÷ Total cost × 100

And the payback period:

Payback (months) = Total upfront cost ÷ Net monthly value

The work is in estimating "value created" honestly. Four sources cover most business AI projects:

What should you include in total cost?

• Setup or build: design, integration, configuration, testing.
• Running costs: AI model usage, WhatsApp Business API messages, hosting, software licences.
• Internal time: staff hours for providing information, testing and training.
• Ongoing improvement: reviews, tuning and new features, whether in-house or a retainer.

Leaving out internal time or running costs makes ROI look better on paper and causes disappointment later.

Worked example 1: WhatsApp AI agent for a clinic

Baseline: a clinic receives about 1,200 WhatsApp enquiries a month. Around 25% arrive after hours and many go unanswered until the next day. The front desk estimates 60 potential bookings a month are lost. Average visit value is KES 3,500.

• After launch, the agent answers all enquiries instantly and books directly. Suppose it recovers 30 of the 60 lost bookings: 30 × KES 3,500 = KES 105,000/month extra revenue.
• Reminders cut no-shows by 15 a month: 15 × KES 3,500 = KES 52,500/month.
• Front-desk time saved: 40 hours × KES 400/hour = KES 16,000/month.
• Running costs: about KES 20,000/month.

Net monthly value ≈ KES 153,500. Divide the setup cost by that number for the payback period. For many clinics it's measured in months, not years. Your numbers will differ. The method is what matters. See AI receptionists for clinics.

Worked example 2: M-Pesa reconciliation for a school

Baseline: a bursar spends 3 hours a day matching fee payments, and about 5% of payments go unidentified for weeks.

• Time saved: about 55 hours a month × KES 500/hour = KES 27,500/month.
• Faster collections: accurate balances and automated reminders bring payments in earlier, which improves cash flow and reduces arrears carried into the next term.
• Fewer disputes and calls, which saves time and improves parent trust (harder to price, but real).

See M-Pesa reconciliation.

What's realistic in 90 days?

That's why we back our AI systems with a 90-day ROI promise: if you follow the plan and the agreed metric doesn't improve by at least 30%, we work free for up to 60 more days or refund 50% of the fee. A promise like that is only possible when the metric is specific and measured from day one.

Which AI projects deliver ROI fastest?

• Revenue leaks: unanswered enquiries, slow follow-up, no-shows, abandoned carts.
• High-volume repetitive work: data entry, reconciliation, standard replies.
• Cash-flow processes: invoicing, reminders, payment matching.
• Visibility: dashboards and alerts that catch costly problems early.

Slower or harder to measure: broad "AI strategy" projects, experimental tools with no process owner, and content generation without distribution.

How to set up measurement before launch

1. Choose one primary metric and two supporting ones (e.g. bookings from WhatsApp, response time, no-show rate).
2. Measure at least two to four weeks of baseline.
3. Agree how each metric is counted and where the data comes from.
4. Put the metrics on a dashboard from day one.
5. Review at 30, 60 and 90 days, and record what changed and why.

Our guide to AI KPIs lists the metrics worth tracking and the ones to ignore.

Common reasons AI projects fail to deliver ROI

• No clear problem: "we should use AI" instead of "we lose 60 bookings a month".
• Not connected: AI that talks but can't act in your systems.
• No owner: nobody reviews results or improves the system.
• Poor data: wrong prices, outdated FAQs, messy customer records.
• Staff work around it: because they weren't involved or trained.
• Measuring activity, not outcomes: messages sent instead of bookings made.

Beyond the first 90 days

The best returns compound. Once the first automation is stable, the infrastructure (WhatsApp API connection, CRM, payment integration, dashboard) is already in place, so the second and third automations cost less and pay back faster. Many businesses move from one agent to an operations command center over a year. See what an AI operations command center is.

Soft benefits that still count

Some benefits are real but harder to price: owners getting evenings back, staff spending less time on repetitive work, customers getting faster and more consistent answers, and better decisions because numbers are visible. Don't inflate ROI with them, but do record them. They often decide whether a team embraces the next project.

Build your own estimate

Use this quick worksheet:

1. Monthly volume of the task (enquiries, invoices, payments): ____
2. Minutes per item today: ____ → hours per month: ____
3. Cost per staff hour (fully loaded): KES ____
4. Revenue lost per month today (missed enquiries × value × conversion): KES ____
5. Expected improvement (be conservative, e.g. 30–50%): ____
6. Monthly running cost: KES ____ and setup cost: KES ____

Then apply the formulas above. If the payback is under 12 months with conservative assumptions, it's usually worth a pilot.

Who should own the ROI number?

The business owner of the process (sales manager, bursar, operations lead), not the developer. They agree the baseline, review results monthly and decide whether to expand, adjust or stop.

Next step

Pick the process where you can most easily count the waste. That's where AI ROI is easiest to prove. We start every project with a free 7-day working prototype so you can see results before committing.

Related articles

The Metrics That Prove Your AI Is Working (and which to ignore)How to Choose a Custom AI Software Agency (Without Getting Burned)From Prompt to Prototype in 7 Days: the rapid build method

Frequently asked questions

Explore Kavenacc services & pricing →