AI-assisted diagnosis in consulting: a founder's guide

AI-assisted diagnosis in consulting means pointing AI at your own firm to find where the work jams up, and which of your decisions are worth automating. It has nothing to do with medical diagnosis. The name is unlucky, but the job is simple enough.

What you get out of it is a short list of repeatable workflows. They hand you back your time and recover the margin you are losing to manual work you do over and over.

---

A practical 90-day starter plan you can run right now

Phase	Weeks	What to achieve
Audit and mapping	1-4	List your top 10 recurring decisions. Score each by frequency and £ cost. Choose one pilot. Map the inputs it needs.
Prototype	5-8	Build the diagnostic workflow. Test on real (anonymised) data. Produce a sample output for review.
Test and handoff	9-12	Run three live cycles. Measure time saved. Document the process. Train the team. Hand off to operations.

Minimum viable scope for your first pilot:

• One decision type only
• Inputs you already have (no new data collection)
• A human sign-off step built in from day one
• A clear KPI you can measure before and after

Leave out: multi-system integrations, client-facing automation, and anything that touches personal data before your GDPR review is done.

Build it by hand before you automate it. Unscalable first, scalable second. That is the order that works.

Sit with Claude and do the diagnostic job yourself, over and over, until it comes out right every time. Each time it nails a step, save that step as a skill. Then chain the skills together. Then automate the chain with Claude Code.

Skip that and you are building the magic button on day one. It does not work and nobody uses it. Most dead pilots are this same mistake in a different coat.

Selling the diagnosis as a small paid job works well too. It brings in cash, it shows you the follow-on work, and it keeps the scope tight enough to finish in weeks.

---

What to measure and what a 90-day prototype costs

Core KPIs:

• Hours saved per engagement
• Cost saved (hours × rate)
• Delivery velocity (days from brief to draft)
• Conversion rate on proposals
• Client satisfaction score

Example calculation. If your team spends 8 hours per client on discovery synthesis, and AI cuts that to 2 hours, you save 6 hours per client. At a £150 internal rate, that is £900 per client. Across 10 clients a month, that is £9,000 recovered monthly.

A 90-day prototype costs money in three places: people time, tooling, and cleaning up your data. People time is nearly always the biggest line. Tooling for a one-workflow pilot is usually the smallest, which surprises people.

To sell the ROI internally, set the cost of the pilot against the monthly saving over a year. Use your own numbers, not someone else's. That is the whole point of measuring before and after. You end up with a figure you can defend in the room.

The discovery phase of any AI build exists for the same reason. Skip it and you build on guesses.

---

Who owns the diagnosis, and what does UK GDPR require?

Clear ownership stops pilots from stalling. Assign these roles before you start:

• Data owner. Accountable for input quality and access control.
• Model reviewer. Signs off every AI output before it reaches a client or influences a decision.
• Process owner. Responsible for the workflow running correctly and for logging changes.

UK GDPR minimum checklist for a pilot:

• Confirm your lawful basis for processing. Legitimate interests is common for internal operational data
• Apply data minimisation: use only the fields the workflow actually needs
• Anonymise or use synthetic data wherever possible in early testing
• Log every human review decision so you have an audit trail

The ICO publishes its own guidance on AI and data protection. Read it before the pilot, not after. It comes from the regulator, so it beats any blog summary of it, including this one.

On pilot data, three rules. Use anonymised extracts. Never put live client personal data through a new workflow until it has been through a review cycle. Keep the outputs in the access-controlled system you already have, rather than standing up a new tool for them.

Training your team to check AI outputs is part of governance, not an extra.

---

What tends to go wrong, and how to fix it fast

Most pilots fail for one of four reasons:

Red flag	One-line fix
Draft AI output sent to clients as-is	Add a mandatory human review step before any output leaves the firm
Scope creep beyond the pilot decision	Write a one-page scope document and review it weekly
Missing evidence discipline	Require a source reference for every AI-generated claim
Overtrusting model numbers	Cross-check any figure the AI produces against your own records

One early warning sign. If the output keeps using different names for the same client, service or metric, stop the pilot and fix the input data. Messy naming does not stay small. It compounds through every step that follows.

Drift like that is a structure problem, not a model problem. Orchestration is the fix: saved skills and decision-first templates that hold your standard no matter who is running the job.

---

What a real diagnosis turns into

The biggest AI rebuilds I have worked on did not start out as big projects. They started as small jobs that grew.

I once took on what looked like a tiny project for a disability services group. Audit their digital platforms. Find the overlap and the redundancy. Tidy the tech stack, maybe save a bit on subscriptions. One and done. I had no vision for anything bigger.

That audit found a lot of overlap. It saved them several thousand a month. The CEO was pleased enough that he wanted to go after the next problem, which was getting all their data into one place. That is when it clicked. We were not fixing isolated problems. We were building toward something much bigger without ever calling it that.

So we solved one problem a month, for six to twelve months. We automated onboarding. We built a system that found funding opportunities in minutes instead of hours. What began as a platform audit ended up rebuilding how the whole business ran, and it never disrupted the day job.

I worried that going slow would kill momentum. The opposite happened. Each win built the appetite for the next one.

Three things to take from that:

• Start with the decision the founder touches most. That is where the time is hiding.
• Work back from the end goal, then solve one problem at a time. Each win funds the next.
• Measure before and after from week one, even roughly. You need the number to justify the next phase.

---

The part most consultants get backwards

Most founder-led firms come at AI the wrong way round. They pick the tool first, then go hunting for a problem it might solve. That is backwards.

Do it the other way. Find the decision that costs you the most time or money. Put a £ figure on it. Then ask whether AI can do the pattern-spotting part of that decision. Yes, build a tight prototype. No, move to the next decision on the list.

People also underrate the diagnosis itself. A clear read on how a business really runs is a product you can sell. Clients pay for clarity. Show a prospect exactly where their firm leaks margin and what fixing it takes, and you have a paid engagement instead of a free pitch.

There is a bigger prize sitting behind the first pilot. Every workflow you prove and save is another piece of your method written down in a form software can run. Do that across enough decisions and you stop owning a pile of loose automations. You own an AI Operating System: AI employees built with Claude Code that carry your judgement into lead handling, onboarding, delivery and admin.

That is what gets you out of every loop. One automation never does.

AI speeds up the pattern spotting. You still bring the judgement, the accountability and the final call. That split is what makes the whole thing hold.

---

Ready to turn your expertise into a repeatable system?

We run a 90-day program for founder-led consultancies and education businesses doing £1M or more. It maps your IP, finds the decisions worth automating, and builds the workflows your team can run without you in the room.

It suits founders who already have a proven method and want to scale delivery without hiring more senior people or working longer weeks. You get custom diagnostics, hands-on prototyping, and a full handoff to your operations team.

Start here: run the IP monetisation assessment and find out which parts of your expertise are ready to automate now.

---

Useful sources

• Guidance on AI and data protection, Information Commissioner's Office
• A guide to lawful basis, Information Commissioner's Office
• Claude Code documentation, Anthropic
• The AI software discovery phase, explained for founders, Devwiz
• How AI supports consultant decision trees, The AI Orchestrators

---