Anand Abhishek.
Monthly report
Advisory · The playbook

Diagnosis before prescription

How an AI transformation should be run, and why it should start the way a good doctor's appointment does.

AvailabilityPersonal advisory is currently limited. This page sets out the playbook I believe every AI transformation should follow, whoever runs it.

When you are unwell and see a good doctor, a few things happen in a fixed order. You describe what hurts. The doctor asks questions, examines you and sends you for tests. The reports come back. Only then is there a prescription, and after that a follow-up to check it is working.

A bad doctor hands you the medicine in the first two minutes.

Most AI programmes are run the second way. The prescription arrives first: a chatbot, a forecasting model, an agent platform. The diagnosis never happens. A year on, the business is no healthier and concludes that the medicine doesn't work.

How it usually goes
  1. Pick a tool
  2. Find a place to use it
  3. Run a pilot
  4. Wonder where the return went
How it should go
  1. Hear the symptoms
  2. Diagnose with the data
  3. Prescribe the whole treatment
  4. Monitor, and stop if it isn't working

What the organisation has to bring

A doctor cannot treat a patient who will not be examined. A serious transformation needs three things from the business, and the quality of everything after depends on them.

The symptomsWhat hurts, in business terms. Margin leaking, stock in the wrong place, slow turnaround, customers leaving.
The process mapHow the work runs today, as it really runs. Who hands what to whom, and where it waits.
The dataAccess to the systems behind that process. This is the blood test.

The business knows itself far better than any adviser can. What an adviser should bring is the ability to read it in the data.

The treatment plan

  1. Consultation

    Listen first

    Sit with the people who run the process, not only the people who sponsor the project. Leave with the symptoms in their words and a first map of the work.

  2. Diagnosis

    Test the symptoms against the data

    Symptoms and causes often sit in different departments. Trace the problem through the data to where it starts, and size it in money. Sometimes the diagnosis is that AI is the wrong treatment, and that should be said plainly.

  3. The report

    Issues, written down and ranked

    A plain document: what is wrong, what it costs, what is known with confidence and what is not. The organisation should own it, whatever happens next.

  4. Prescription

    Redesign the handoffs

    Redraw the process around the decisions that need to be made together. Then choose the lightest tool for each piece: a model where prediction is needed, automation where a rule will do, a person where judgement matters.

  5. Treatment

    Build in-house, or with specialists

    Each piece can be built by the internal team or by a specialist who does that one thing best. Whoever advises should have no product to push, so the choice follows the fit.

  6. Orchestration

    The craft that decides the outcome

    The pieces have to work as one system: shared data, clean handoffs, one owner, one set of numbers. Most programmes fail here, in the stitching. This is the part that should never be handed off.

  7. Follow-up

    Monitor against the agreed return

    Agree the measure of success before building and track it after. If the return is not appearing, close the work down and explain why. Nobody should keep paying for a treatment that is not working.

Why orchestration is the part that matters

A patient seeing five specialists needs one physician who holds the whole picture. Otherwise each prescribes well for their own organ and the drugs interact. Businesses are the same. The evidence on where AI value comes from points at the joins, not the parts.

Independent research agrees. McKinsey finds that redesigning workflows is the factor most associated with profit impact from generative AI.1 BCG finds that most of the difficulty in AI programmes lies in people and process, and the least in algorithms.2 MIT finds that work done with specialist partners reaches deployment more often than work built entirely in-house.3

That is why sending a piece of the work to a specialist is healthy. It is also why someone must own the whole.

Three beliefs behind the playbook

When this playbook fits

  • There is a business problem you can name, even if you cannot yet name the cause.
  • You are willing to show the real process and the real data.
  • You would sooner fix one thing fully than pilot ten.
  • You have tried AI before, and it did not pay.
Personal advisory

Currently limited

I take on very few advisory conversations at the moment. If you have a problem you can name and want to compare notes on it, get in touch.

Get in touchRead the deep dives

References

  1. McKinsey, The state of AI: How organizations are rewiring to capture value, March 2025. Link
  2. BCG, Where's the Value in AI?, October 2024. Link
  3. MIT NANDA, The GenAI Divide: State of AI in Business 2025. PDF
  4. RAND, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, August 2024. Link