The one-page plan
Kavita's CEO asked what the AI plan was. Here is the answer she gave, five months later, on one page — and the structure of it is the deliverable of this whole course.
The page
Where we looked. We audited 41 tasks across the team on volume, text-shape, and cost of error. Four scored as strong candidates. One task that scored high — fraud detection — was excluded on error cost and regulatory exposure, and three vendors proposed it anyway.
What we did. Piloted AI-assisted claim file summarisation for eight weeks with twelve reviewers, against a measured baseline. Deployed it. Configured two smaller tools internally for acknowledgement letters and complaint classification. Redesigned the intake form, which removed most of the document-chasing work and needed no technology at all.
What it achieved. Review time 18 → 11.5 minutes. QA error rate flat at 2.1%. Throughput up 22%, absorbed without additional headcount. 15% of files flagged for source verification, by design.
What it costs. Licences ₹X per year. Internal time roughly 75 hours to date, and an estimated day a month ongoing. Review time is counted inside the 11.5 minutes.
What is allowed. A one-page data policy, in force, with a named owner. Every customer-facing output is reviewed by the person accountable for sending it. No AI system makes a decision affecting a customer's claim.
What would make us stop or change. QA errors above 2.5%. Checker modification rate falling below 10%, which would indicate the checkpoint has stopped functioning. Any regulatory guidance requiring explicability we cannot provide.
What is next. Evaluate claim extraction into the core system, which needs integration and is therefore a build decision rather than a configuration. Not before Q3.
Why that page works
It answers a business question with business evidence. It names what was excluded and why, which is what makes the inclusions credible. It states what would make the answer change. And it recommends one thing for one task rather than a transformation.
Compare it to the version most leaders receive — a strategy deck with a maturity model and a roadmap of capabilities. Kavita's page is smaller and worth considerably more, because every line of it is a thing that happened.
The four decisions, revisited
Where does it fit? The audit answers this, and only the audit does. Three hours with your own people beats any consultant's framework, because the answer is specific to your operation.
Is it working? Baseline first, three numbers, honest about the checking time, re-measured at three months. A number without a baseline is an opinion.
What is allowed? One page, recitable, with a named person to ask. Written before deployment rather than after an incident.
Who does what? Everyone competent, a few who build, one who goes deep, and a prompt library so that what one person learns does not stay with them.
The short version of everything
Audit before you shop. The order determines whether you are solving your problem or the vendor's.
Pilot against a baseline, with written kill criteria. A pilot that cannot fail is a purchase.
Count the checking. It is the most under-counted cost in this field, by a wide margin.
Put a person before anything irreversible, and make sure they can actually see the error.
Accountability does not transfer. Design as though you will have to explain one specific case, because eventually you will.
Watch the five early signals. Usage in week three. Modification rate. Second maintainer. Pilot end date. What the team is really using.
Where to go from here
If you want the mechanism under all of this, How AI Actually Works is an afternoon and will make vendor conversations easier for the rest of your career.
If your instinct after the audit was "we could just build most of this ourselves", that instinct is usually correct at your scale, and Build AI Tools Without Code is where to send the two people who will do it.
And put AI for Everyday Work in front of the whole team, because competent individual use is the part with the best return and the one that gets skipped in favour of projects.
Do this today: write your own version of the page, with the sections empty. The gaps are your plan — and the first one to fill is almost always the audit, because everything else depends on it.