An AI implementation roadmap that survives contact with a real business

There is no shortage of AI strategy advice and very little about the boring middle — the part where a promising pilot has to become something people use on a Tuesday. Having taken AI from idea to daily production use in our own products, here is the sequence we follow when we run an AI implementation for someone else.

Phase 0: Decide what you are actually optimising

Before any use case, get honest about the goal. “We should use AI” is not a goal. “Our reps take four hours to respond and we lose leads to faster competitors” is. Write the problem in a sentence that includes a consequence. If you cannot, the project has no way to prove itself later.

Phase 1: Choose the first use case on boring criteria

The temptation is to start with the most strategically exciting idea. Resist it — those take longest and fail most visibly. Score candidates on four things:

• Volume. Does it happen many times a day?
• Repeatability. Is the right answer broadly the same each time?
• Data availability. Does the answer exist in material you already have?
• Reversibility. If it gets one wrong, can you recover cheaply?

High on all four is where you start. In most Indian SMBs that means first response to enquiries, ticket triage, or reading documents into a system.

Phase 2: Do the unglamorous data work

This phase is where implementations quietly succeed or fail, and it is usually the majority of the effort. You are answering one question: what is the authoritative version of the truth? Businesses routinely discover three different price lists and two contradictory policy documents. An agent built over contradictions produces confident contradictions, at scale, to customers.

Budget real time here. It is not AI work — it is the tidying you have been postponing, now with a deadline.

Phase 3: Build against the smallest useful scope

Build the agent to do the one job, connected to the systems it needs and nothing more. Define the stopping point before the conversation flow: where it hands to a human, and what it must never claim. Instrument it so you can see, later, exactly what it did and why.

Phase 4: Launch on a slice, with a person watching

Do not launch to everything at once. Give the agent one channel, or one product line, or the after-hours window. Have a person review what it did each day for the first weeks. Two things come out of this: real failure cases you could not have predicted, and — more importantly — your team's trust. A team that has watched the agent be right for three weeks will let it do more. A team that was told to trust it will look for reasons not to.

Phase 5: Widen slowly, and keep watching

Expand scope only after the narrow version has been boring for a while. Then keep monitoring: businesses change, price lists change, and an agent that was accurate in March will drift by September if nobody looks. Ongoing improvement is a line item, not an afterthought.

The failure patterns to avoid

• Starting with a platform. Teams that begin by selecting an “AI platform” usually spend the budget before solving anything.
• No definition of working. Agree the number that means success before you build, or you will argue about it after.
• Hiding it from the team. If staff first learn about the agent when it starts answering their customers, expect resistance you will not overcome.
• Skipping the human loop to save money. The review period is what makes everything after it safe.

If you want a view on where your business should start, the readiness assessment takes two minutes and will tell you plainly if the answer is “not yet”. Or read how we build agents, including what we refuse to promise.