Hiring guide
Which AI projects are actually worth doing? A scoring method
By Samuel Marndi · Updated
The short answer
Score every candidate on four things: how much the current process costs, whether it can be built with today's models, whether the data exists in a usable form, and what a wrong answer costs you. The fourth is the one nobody scores, and it is the one that kills projects after launch rather than before.
Run honestly, this rubric tends to produce an uncomfortable result: the website chatbot — almost always the first idea — scores poorly, and document extraction somewhere in the back office scores highest. The chatbot is visible, which is why it gets proposed; extraction is invisible, which is why it is still being done by a person at a desk.
The four scores, and why the fourth one matters most
Score each candidate 1–5 on all four, multiply nothing, and look at the shape rather than a total. A project that scores 5-5-5-1 is not a good project with one weak spot — it is a project that will be switched off within a quarter.
| Score | The question | A 5 looks like | A 1 looks like |
|---|---|---|---|
| Value | What does the current process cost, in hours or errors per month? | Someone spends 30+ hours a month on it and can tell you so | Nobody can name the cost, because nobody has measured it |
| Feasibility | Can today's models do this reliably, not impressively? | The task is extraction, classification, routing or summarising | It requires reasoning over a long chain where one error ruins the output |
| Data readiness | Does the input exist, in a usable form, without a project to create it? | Files in a folder, rows in a database, emails in a mailbox | It is in people's heads, or in scans nobody has OCR'd |
| Cost of being wrong | What happens when it gets one in twenty wrong? | A human notices and corrects it in seconds | A customer is told something false, or money moves |
Eight projects businesses actually propose, scored
These are the eight that come up most often in scoping conversations. The scores are my judgement, not a measurement, and the point is the ordering rather than the numbers — if you disagree with a row, the rubric is still doing its job, because now you are arguing about the right thing.
| Project | Value | Feasible | Data | Wrong-cost | Verdict |
|---|---|---|---|---|---|
| Invoice / document extraction | 5 | 5 | 4 | 4 | Do this first. Boring, invisible, and the arithmetic works |
| Enquiry routing and triage | 4 | 5 | 5 | 4 | Strong. The data is already in your inbox |
| Internal search over your own documents | 4 | 4 | 3 | 5 | Strong, if the documents are findable. Often they are not |
| Meeting notes and follow-up extraction | 3 | 5 | 5 | 5 | Easy win, small prize. Good first project to build confidence |
| Quote and proposal drafting | 4 | 4 | 3 | 3 | Worth scoping. Needs a human signature step, always |
| Website chatbot for sales | 2 | 4 | 2 | 2 | Usually not first. Visible, popular, and it answers questions few people were asking |
| Outbound voice calling | 3 | 3 | 3 | 1 | Regulated. Read the rules before the architecture — see the voice agent page |
| Fully autonomous decision-making | 4 | 2 | 2 | 1 | No. Not a model problem; an accountability problem |
Why the chatbot keeps winning the argument and losing the rubric
It scores 2 on value because the questions a website chatbot answers are mostly questions a good FAQ page answers for free — and the enquiries that actually convert tend to arrive by phone or WhatsApp from someone who has already decided. It scores 2 on data because "our website content" is usually thinner than people think once you try to answer real questions from it. And it scores 2 on cost-of-being-wrong because a confident wrong answer to a prospect is worse than no answer.
None of that makes chatbots useless. A support bot on a real documentation set, answering genuinely repetitive questions, scores well — and what an AI chatbot costs works through the economics with an estimator. The claim here is narrower: it is rarely the right first project, and it is almost always the first one proposed.
The four questions that kill a project after launch
A project can pass the rubric and still fail, and it usually fails on one of these. Ask them during scoping, when the answers are cheap.
- Who checks the output, and is that in their job? Every useful AI system needs a human in the loop somewhere. If nobody owns that step, the system is unmonitored by design.
- What does it cost per month at ten times the volume? Model usage scales with use. A pilot that costs ₹4,000 a month at fifty documents a day costs rather more at five hundred, and that arithmetic should be done before the build, not after.
- What happens when the model changes? Providers deprecate and re-tune. Anything in production needs a small evaluation set you can re-run, or you will not know the day it gets worse.
- What are you allowed to send it? Customer data, financial records and anything covered by a client contract need an answer on where it goes and under which terms. This is a procurement question, not a technical one, and it stops more projects than feasibility does.
How to run this in an afternoon
This does not need a consultant. It needs the right four people in a room for two hours, and someone willing to write a 1 down.
- List candidates without filtering. Ten to fifteen, from anyone who has an idea. Filtering early loses the back-office ones, which is where the value usually is.
- Score value with whoever does the work today, not with whoever proposed the project. They are rarely the same person and only one of them knows the hours.
- Score data readiness by looking at the data. Actually open it. "We have all that in the system" survives contact with the system about half the time.
- Rank by the lowest score, not the total. A 1 anywhere is a veto until it is fixed, and a project with no 1s and no 5s beats a project with both.
- Pick two: one boring and one visible. The boring one pays for the programme; the visible one buys you permission to continue.
- Write down what would make you stop. A number, before you start. The most expensive AI projects are the ones nobody was allowed to cancel.
Where this sits next to the rest
This page is the triage step — what to do first, out of a list. What the right first move looks like for a startup against an SME against an enterprise is a different question and it is answered on AI consulting in Bangalore, along with what a discovery sprint or proof of concept costs. If you already know what you want built, AI chatbot and agent development and AI voice agents are the build pages.
And if the honest conclusion is that nothing on your list scores well yet — usually because the data is not ready — that is a real finding and a cheap one. The fix is a data project, not an AI project, and it will make every later AI project cheaper.
FAQ
Related questions
How do I decide which AI project to do first?
Score each candidate 1–5 on four things: what the current process costs in hours or errors, whether today's models can do it reliably rather than impressively, whether the input data already exists in usable form, and what a wrong answer costs. Then rank by the lowest score rather than the total — a 1 anywhere is a veto until it is fixed. A project with no 1s and no 5s beats a project with both.
Is an AI chatbot a good first AI project?
Usually not, despite being the most commonly proposed. It scores poorly on value because most of what it answers a good FAQ page answers free, poorly on data because website content is thinner than people expect when you try to answer real questions from it, and poorly on cost-of-being-wrong because a confident wrong answer to a prospect is worse than no answer. A support bot on real documentation is a different and much better case.
What is the highest-value AI project for a small business?
Document and invoice extraction, almost every time. It is invisible, which is why nobody proposes it, and it scores well on all four criteria: someone is measurably spending hours on it, extraction is what current models are genuinely reliable at, the input is already sitting in a folder or a mailbox, and a human notices and corrects an error in seconds.
What does AI readiness actually mean?
In practice it means one thing: does the input data exist, in a usable form, without a project to create it first. Files in a folder, rows in a database and emails in a mailbox are ready. Knowledge in people's heads and scans nobody has run OCR over are not. If nothing on your shortlist scores well on this, the right next project is a data project, and it will make every later AI project cheaper.
How much should a first AI project cost?
Small enough that cancelling it is not a crisis. In Indian terms that usually means a scoped pilot rather than a platform — and the test is whether you have written down, before starting, the number that would make you stop. The most expensive AI projects are the ones nobody was permitted to cancel. Engagement and build costs are set out on the AI consulting and chatbot pages.
Who should be in the room when prioritising AI projects?
Whoever does the work today, whoever owns the budget, whoever owns the data, and one person willing to write a 1 down. The common failure is scoring value with the person who proposed the project rather than the person doing the task — they are rarely the same, and only one of them knows how many hours it takes.
Do we need a consultant to do this?
No. This is two hours with the right four people, and the article is written so you can run it yourself. Where an outside view earns its fee is when the shortlist is long, the data readiness answers are contested, or somebody is already selling you a platform — because then the question is not which project, it is whether the project you have been shown is the one you need.
What should we do if every idea scores badly?
Treat that as the finding rather than a failure, and look at why. It is almost always data readiness, which means the useful next project is making the data findable and machine-readable rather than buying AI to sit on top of it. That work is unglamorous, it has value on its own, and it changes the score of everything else on the list.