Most businesses are past the question of whether to use AI. The harder question — the one that actually stalls people — is where to start. You’ve seen the demos, your competitors are making announcements, and there’s a quiet pressure to “do something with AI.” But pressure to act is not a plan, and starting in the wrong place is expensive: you can burn six months and a healthy budget on a project that impresses no one and quietly gets shelved.
This is a roadmap for starting deliberately — how to choose a first project that genuinely matters, prove it works, and earn the momentum to do more. It’s the path we walk clients through before a single line of code gets written.
Start with a problem, not the technology
The most common way to begin badly is to start from “we need AI” and go hunting for somewhere to put it. That’s backwards, and it shows up later as a slick tool nobody uses.
Start instead from a problem that’s already costing you: slow response times, a bottleneck task that eats your best people’s afternoons, a decision made on gut feel that could be made on data. AI is a means; the problem is the point. If you can’t name the problem in one sentence and say roughly what it costs you, you’re not ready to build — you’re ready to look harder.
A good first problem tends to be:
- Painful — someone feels it every week, not once a quarter.
- Measurable — you can attach a number to it (hours, dollars, response time, error rate), which is what lets you prove the project worked.
- Bounded — narrow enough to finish in weeks, not a moonshot that reorganises the company.
- Yours to fix — you own the process and, crucially, the data behind it.
Are you ready? An honest readiness check
Before you fall in love with a use case, look squarely at three things. Most stalled projects were doomed here, not in the modelling.
- Data. Do you actually have the information the AI would need — and is it accessible, reasonably clean, and allowed to be used? “We have tons of data” and “we have usable data” are different sentences.
- Process. Is there a defined workflow the AI plugs into, and a human who owns it? AI improves a process; it can’t invent one that doesn’t exist.
- Appetite. Is the team ready to change how they work, and is leadership ready to fund a second step after the first? A pilot with no path to production is a science experiment.
You don’t need all three to be perfect. You need to know which one is weakest, because that’s what you’ll spend your time fixing.
Choose your first use case: score the candidates
When you have a few candidate problems, resist picking the most exciting one. Pick the one most likely to succeed visibly — an early win buys you the credibility to tackle the ambitious stuff later. Score each candidate on four axes:
| Candidate first project | Business impact | Data ready? | Risk if it’s wrong | Good starter? |
|---|---|---|---|---|
| Internal document search / Q&A over your own knowledge | Medium–High | Usually yes | Low | ✅ Excellent |
| Drafting customer-support replies (human reviews) | High | Often | Low — a person approves | ✅ Strong |
| Summarising / triaging inbound requests | Medium | Usually | Low | ✅ Strong |
| Sales or demand forecasting | High | Needs history | Medium | ⚠️ Later |
| Fully autonomous customer-facing agent | High | Demanding | High | ❌ Not first |
Notice the pattern: the best starters are high-value, low-risk, human-in-the-loop tasks where a mistake is caught before it reaches a customer. The worst first project is anything autonomous and customer-facing — the failure mode is public, and you haven’t earned the reliability yet. (If an agent is on your roadmap, read AI agents explained for where they shine and where they bite.)
Crawl, walk, run
Adoption is a sequence, not a leap. Each phase earns the right to the next.
- Crawl — one narrow use case, internal, with a human in the loop. The goal is a real result you can measure, not a platform.
- Walk — integrate it into the actual workflow, expand to more users, and start measuring against the baseline you captured at the start.
- Run — scale what works, add adjacent use cases, and only now start thinking in terms of a shared platform rather than one-off tools. (When you get here, platform vs website is worth a read.)
The teams that succeed treat their first project as the start of a capability, not a one-off. But they still start small.
Your first 90 days
Concretely, a sensible first project fits in a quarter:
- Weeks 1–2 — Define. Lock the problem and a single success metric. Capture the current baseline number so you can prove improvement later.
- Weeks 3–4 — Assess and prototype. Check the data is real and usable. Stand up a rough prototype to test feasibility — is this even tractable?
- Weeks 5–8 — Build narrow. Build the smallest useful version and put it in front of a handful of real users. Watch what they actually do with it.
- Weeks 9–12 — Measure and decide. Compare against the baseline. Then make an honest call: scale it, iterate on it, or kill it. “Kill it” is a valid, cheap outcome here — that’s the whole point of starting small.
If that arc feels familiar, it’s the compressed version of the full custom AI project lifecycle.
Don’t build what you can buy (yet)
For a first step especially, reach for off-the-shelf tools before commissioning custom work. If a mature product solves 90% of your problem today, use it, learn from it, and find its ceiling. Build custom only where the tools genuinely can’t fit your data, workflow, or the experience you need to own — the full framework is in custom AI vs off-the-shelf. Proving the value cheaply first is not a detour; it’s how you de-risk the bigger investment. (And if you’re wondering about that investment, what custom AI costs lays out the numbers.)
A quick gut check
Before you greenlight a first project, you should be able to answer yes to all of these:
- Can you state the problem — and its cost — in one sentence?
- Is there a single number that tells you whether it worked?
- Do you have the data it needs, and permission to use it?
- Is there a named person who owns the workflow it plugs into?
- If it fails, is the failure cheap and contained?
- Is there an appetite (and budget) for a step two if it works?
Any “no” isn’t a reason to stop — it’s simply the first thing to fix.
The bottom line
Starting with AI isn’t about picking the cleverest model or the flashiest use case. It’s about picking a real, painful, measurable problem, proving you can move the number on it, and using that win to earn the next one. The businesses that get value from AI aren’t the ones that started biggest — they’re the ones that started narrow, measurable, and human-supervised, then scaled what worked and quietly killed what didn’t.
Trying to figure out where your best first project actually is? That’s exactly the conversation we like having. Tell us what’s slowing you down, or see how we approach AI strategy consulting.