The Scoutz Independent technology studio · Mumbai WebsitesCustom softwareAI automation
AI

What an AI pilot should cost, and how to tell if it worked

Pick one repeated task, put a number against it before you start, and run it for a month. Everything else is a demo.

Most AI projects in small and mid-sized businesses die the same way. Someone sees a demo, a budget is approved, a tool is bought, and three months later nobody can say whether it helped. The fix is not a better model. It is picking one repeated task, writing down a number before you start, and running it for a month.

Choose the task, not the technology

A good first candidate has four properties: it happens often, it follows a pattern, a human can check the output quickly, and being wrong occasionally is survivable. Drafting replies to the twenty questions you answer every week qualifies. Deciding who gets credit does not.

Tasks that almost always work as a first pilot:

  • Drafting first replies to routine enquiries, from your own templates, for a person to send.
  • Pulling structured details out of documents that arrive as PDFs or photographs, with a human confirming before anything is saved.
  • Summarising long threads or call notes into the three lines somebody actually needs.
  • Checking submissions against rules you can write down, and flagging the exceptions.

Notice what these have in common: the AI prepares, a person decides. That is where the current technology is reliable, and where the savings are real.

Write down the number before you start

This is the step that gets skipped, and the reason most pilots end in an argument about vibes. Spend a week measuring, by hand if necessary:

  1. Volume. How many of these happen per week?
  2. Time each. Timed, not estimated. Estimates are generous in both directions.
  3. Error rate. How often does it currently need redoing, and what does a mistake cost?
  4. Who does it. One person, or everyone a little? Work spread thinly across a team rarely converts into recovered hours; work concentrated in one person usually does.

Those four numbers are your baseline. They also tell you the ceiling: if the task consumes three hours a week in total, no amount of AI will return a large number, and you should pick a different task before spending anything.

What the money actually buys

Our AI work sits in a ₹50,000 to ₹15,00,000 band, which is wide because the work genuinely is. Where a project lands depends on how much of the organisation it touches, not on how clever the model is. A single workflow in one team sits at the bottom. A programme across departments, with data handling to satisfy and a dozen integrations, sits at the top. What is in it either way:

  • An audit. Where the hours and the errors actually go, which is frequently not where management believes.
  • A costed roadmap. What to automate, in what order, with the expected saving against each — written so a finance team can read it.
  • Guardrails. What data may be sent where, what is never pasted anywhere, who can see outputs, and what happens when the model is wrong.
  • A pilot with a number against it, running on your own API keys and your own accounts, so the usage bill is yours and visible and the thing keeps working if we disappear.

Those bands are starting points rather than quotes. We agree what the fix is worth to you first and shape the scope around that — and if the budget is fixed, saying so early gets you an honest answer about what fits inside it.

One cost to plan for separately: model and API usage is billed by the provider, not by us. It is usually modest for a single workflow, but it is a running cost, not a one-off, and it should appear in the business case from day one.

Four ways pilots fail

  • No baseline. Without the before numbers, "it feels faster" is the only available verdict, and it convinces nobody holding the budget.
  • Automating a broken process. If the underlying steps are wrong, AI produces wrong results faster. Fix the process first; it is cheaper.
  • No owner. A pilot without one named person responsible for it drifts back to the old way within a fortnight.
  • Measuring the model instead of the work. Accuracy benchmarks are interesting. Hours recovered, errors avoided and enquiries answered faster are what you are buying.

Judging it after a month

Re-measure the same four numbers, then ask three questions. Did the time per item fall enough to matter? Did quality hold — same errors or fewer, checked on real output, not samples? Would the team complain if you switched it off tomorrow? That last one is the most honest signal available.

Then make a decision on paper: extend, adjust, or stop. Stopping a pilot that did not pay is a successful outcome, not a failure — it cost one month instead of a year, and you now know something specific about your own operation.

Where lessons fit instead

Sometimes the right answer is not a build at all. If two or three people would get most of the benefit by using these tools well themselves, private lessons on your real work cost a fraction of a pilot and the knowledge stays in the business. We quote those separately, and we will suggest them when they are the better buy.

What we would do

In a free 20-minute review we map your week and mark the three places AI would pay for itself first, in order. You get that list whether or not you hire us. If a pilot follows, it starts with the baseline measurement — before any tooling is chosen — because without it nobody can tell you whether the thing worked.