Most wasted AI spend I've seen starts the same way: a team pays for an API, then finds its own data can't feed it. Two weeks later they're cleaning spreadsheets instead of shipping.
Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data, and found 63% of organisations lack the right data practices for AI or aren't sure (Gartner, 2025). Here's the one-afternoon audit I run before anyone enters a card number.
1. Data: can you reach it, and trust it?
Can a script read your data, or does someone export a spreadsheet? Does each record have an owner and one consistent ID? Which fields are personal data (anything that identifies a person, like a name or phone number)? If half your support tickets say "see email", a ticket summariser has nothing to work with.
2. Infrastructure: can you call an API safely?
Calling a hosted model needs little compute: keep the API key outside your code, add a timeout with a fallback, and log errors without logging what users typed.
3. Security and compliance
India's Digital Personal Data Protection (DPDP) Rules were notified in November 2025 with an 18-month phased compliance timeline, and failing to keep reasonable security safeguards can cost up to ₹250 crore (PIB, 2025). Before customer data reaches an AI provider, you need a stated purpose, a notice and a way to delete it. My rule: if you can't say in one sentence where a customer's data goes, don't send it.
4. Team skills
You don't need a machine-learning team, just one engineer who owns the prompt (the instructions sent to the model), tests answers on real examples and watches cost.
5. Workflow integration
Pick one repetitive reading or writing task, like drafting replies or tagging tickets, and time it today. Without that baseline you can't tell whether AI helped.
Estimate API spend before committing
Providers charge per token, a small chunk of text, often a word or part of one. Say you draft 1,000 support replies a day, each sending 1,500 tokens and receiving 300: that's 45 million input and 9 million output tokens a month.
| Model | Input /1M | Output /1M | Monthly | ≈ ₹ |
|---|---|---|---|---|
| gpt-5.4-nano | $0.20 | $1.25 | $20.25 | ₹1,944 |
| gpt-5.4-mini | $0.75 | $4.50 | $74.25 | ₹7,128 |
| gpt-5.4 | $2.50 | $15.00 | $247.50 | ₹23,760 |
At ₹96 to the dollar, rounded. Check today's rate and prices before budgeting.
The largest model costs about 12 times the smallest. Test the cheapest first and move up only when it fails.
Monthly cost, same workload
Score your stack
Give each layer 0 (not in place), 1 (partly) or 2 (ready):
- Data: readable by a script, owned, personal fields identified.
- Infrastructure: key stored outside code, timeouts, safe logging.
- Compliance: purpose, notice and deletion for personal data.
- Skills: one person owns prompts and testing.
- Workflow: one task chosen, baseline timed.
| Score | Next move |
|---|---|
| 8–10 | Pilot one task for two weeks with the cheapest model that passes your tests. |
| 5–7 | Fix your lowest layer first; in my experience, data or compliance. |
| 0–4 | Hold off on APIs. Spend a month on data ownership and access. |
Your next step
Book one afternoon this week with your lead engineer and whoever owns your customer data. Score the five layers together and name the one task you'd pilot. Below 8, spend the API money on raising your lowest score first.
Sources
- Gartner (2025). Lack of AI-Ready Data Puts AI Projects at Risk. Press release, 26 February 2025.
- Press Information Bureau (2025). Government notifies DPDP Rules to empower citizens and protect privacy. 14 November 2025.
- Press Information Bureau (2025). DPDP Rules, 2025 Notified. November 2025.
- OpenAI (2026). API pricing. Standard rates, checked 10 October 2026.