The AI Consulting Process: Diagnose, Build, Run with Confidence

Your business works, but understanding the AI consulting process can help you see what AI could take off your plate. It just works because you’re working it. So you’ve started looking at AI, and maybe you’ve talked to a consultant or two. Each one gave a different answer about what happens next.

One wanted to run a six-week discovery. One wanted to sell you a chatbot. One sent a 40-page report you never finished reading.

This post sets out the process I use with founder-led businesses, from the first conversation to the system running without you. It has three phases: diagnose, build, run. You’ll see what happens in each one, what you should walk away with, how long it should take, and the questions that separate a partner who ships from one who just sends invoices.

Key Takeaways

• A good AI consulting process has three phases: diagnose the business, build one bounded system, then run and improve it every month.
• Most AI projects stall because they start with a tool. The diagnosis should start with your bottlenecks, your data and the decisions only you make.
• The first build should be one system, scoped tightly and live within weeks, not a six-month transformation programme that never ships.
• Before any build, your AI needs context: how you sell, how you deliver, who does what. Without it, every output is generic and needs rewriting.
• Handover is where weak consultants disappear. The run phase, with training, tuning and support, is where the system starts paying for itself.
• Measure three things: hours you can step away, share of recurring tasks automated, and revenue per person. If none move, the process failed.
• Ask any AI consultant for a written diagnosis, a named first system and a success measure before you sign. Vague scope is the biggest warning sign.

Why Most AI Projects Fail Before the AI Consulting Process Even Starts

Here’s the pattern I see over and over. A founder reads about AI, gets excited, buys a tool, and gives it to someone on the team to “figure out”. Three months later, the tool is another login nobody opens. The founder decides AI is overhyped and goes back to doing everything themselves.

The tool wasn’t the problem. Nobody asked what the tool was for.

Tool first versus problem first

MIT researchers looked at hundreds of enterprise AI rollouts and found that around 95% of generative AI pilots delivered no measurable return. The few that worked had something in common. They started with a specific business process and fitted the technology to it, not the other way round.

Small businesses fall into the same trap, just with smaller budgets. A tradie buys an AI quoting app when the real leak is that nobody follows up on quotes. A clinic adds a website chatbot when the real leak is that 40% of phone calls ring out. The AI works fine. It’s pointed at the wrong problem.

What a consulting process is meant to fix

A proper AI consulting process exists to stop that. It’s a structured way of working out three things before anyone spends money building:

1. Where your time and revenue are actually leaking
2. Which of those leaks AI can plug reliably today
3. What order to fix them in so each one makes the next easier

That’s the diagnose phase. Then you build one thing that matters. Then you run it, tune it and add the next one. Diagnose, build, run. Simple to say. Most of the AI consulting market skips at least one of those steps.

Some skip the diagnosis and sell the same package to everyone. Some do a beautiful diagnosis and never build anything. A lot of them build something, hand it over and vanish, which is how you end up with an automation that breaks the first time a supplier changes their email format.

If you’ve already been burned that way, I wrote a separate piece on rebuilding after a bad AI vendor. The short version: the problem usually wasn’t the AI. It was the process around it.

Phase One of the AI Consulting Process: Diagnose

The diagnosis is where a good partner earns trust. You should come out of it with a clear picture of your business you didn’t have going in, even if you never hire that person to build anything.

Start with the founder, not the tech stack

The first conversation should be about you, not your software. What do you spend your week on? What only you can answer? What happens if you’re off sick for a fortnight?

I ask founders to describe a normal Tuesday. The answers are almost always the same shape. Checking messages before breakfast. Answering five questions the team should be able to handle. Chasing a quote, a payment or a supplier. Real work starting at 10:30, if it starts at all.

That Tuesday is the raw material. Every recurring task in it is a candidate.

Map and score the recurring work

Next comes a task map. Every recurring task across the business, yours and your team’s, listed in plain English. Most businesses with 5 to 30 people turn up somewhere between 50 and 100 of them once you actually write them down.

Each task then gets scored on how far AI can take it:

• Fully automatable: the system does it end to end. Lead follow-up, appointment reminders, data entry between tools.
• Assisted: AI does 80% and a person steers. Drafting proposals, summarising calls, preparing reports.
• Supervised: AI does nearly all of it and a person signs off. Client emails, pricing changes, anything public.
• Human only: judgement calls, relationships, the work you started the business to do.

Then you weigh each task by how much time or money it costs. The overlap of “fully automatable” and “expensive” is where your first build lives.

Capture the brain

Here’s the part most consultants skip, and it’s the reason so many AI outputs read as if they came from a stranger. Your AI needs context. Who you sell to. What you sell and why. How you price, how you talk, who on the team does what, what you’ve tried before.

I build this as a set of structured context files during the diagnosis. Think of it as the briefing you’d give a senior hire on day one, except the AI never forgets it. Once it exists, every draft, summary and recommendation comes back sounding like your business instead of a template.

Check the data

Last, a quick audit of where your numbers live. CRM, accounting, booking system, spreadsheets. Not a migration. Just enough to know what the system will be able to see and what it won’t.

What you should walk away with

A written diagnosis. Not a slide deck full of AI trends. A document that says: here’s where the time goes, here’s where the money leaks, here’s the first system to build, here’s what it will change and how you’ll know. If the diagnosis doesn’t name one specific first build, it isn’t finished.

If you want a feel for what this looks like in practice for smaller firms, my post on AI adoption for SMEs goes further into how owners decide what to tackle first.

Phase Two: Build One System, Not Ten

This is where the AI consulting process either earns its fee or falls over. The temptation, on both sides, is to build everything at once. Resist it.

Why one bounded build beats a transformation programme

Big multi-month programmes have a habit of never quite finishing. Requirements drift, the team loses interest, and six months later you’ve paid for a lot of meetings and very little that runs.

I build one system first. One clear job, a defined start and finish, live in about two weeks. It might be client onboarding, lead follow-up, call answering or a morning brief. The point is that within a fortnight, something in your business happens without you that used to need you.

That changes how founders think about AI. It stops being a concept and becomes “what else can it do?”

What happens during build weeks

A decent build has a rhythm you can see:

• Days 1 to 3: confirm the exact workflow, step by step, with the person who currently does it. Not you, if you can help it. The person on the front line knows where the exceptions hide.
• Days 4 to 8: build the system against your real tools and your context files. Wire in the CRM, the inbox, the calendar, whatever the job touches.
• Days 9 to 12: test with real examples. Old leads, past client files, last month’s enquiries. Break it on purpose.
• Days 13 to 14: go live with a person watching, then hand over.

You should see working pieces during the build, not just at the end. If a consultant goes quiet for three weeks and then “reveals” the finished product, that’s a red flag.

Human in the loop by default

Nothing that talks to your clients should go fully autonomous on day one. Early on, the AI drafts and a person approves. As trust builds and the error rate drops, you loosen the reins. This is how you avoid the horror story of an AI promising a client something you don’t offer.

It’s also how your team comes to trust the system instead of quietly working around it.

What a good first system looks like

The best first builds hit a leak you already know is there. Take Dr Claire, a dentist I worked with. Her practice was missing close to half its inbound calls despite having two receptionists. The fix wasn’t another hire. It was an AI receptionist answering overflow and after-hours calls, qualifying patients and booking them in. Missed calls dropped to zero and booked appointments rose 44%.

That’s what a first system should feel like. Specific problem, specific fix, a number that moved.

What the build is not

It’s not a custom software project. Most of what a small business needs can be assembled from proven pieces and wired to your context. Borrowing before building keeps the cost down and the timeline short. If someone proposes writing everything from scratch for a 12-person business, ask why.

For a closer look at how these systems hand work between steps without you in the middle, see my piece on agentic workflows.

Phase Three: Run, Train and Improve

Most consultants treat go-live as the finish line. It’s actually the start of the phase where the value builds.

The first 30 days

AI systems don’t arrive perfect. They arrive good, and then they get better as they meet the odd cases your business throws at them. A lead who replies in te reo. A client who books, cancels and rebooks three times. A supplier invoice in a format nobody planned for.

In the first month, I watch the system closely, review what it got wrong, and tune the prompts, rules and context. Your team learns how to use it and, just as importantly, how to tell me when it’s annoying them. By day 30, the error rate should be low enough that people stop double-checking everything.

I wrote a whole post on what that early checkpoint should look like: the 90-day AI automation checkpoint. Worth a read if you’ve got something live and you’re not sure whether it’s working.

Training, optimisation and support

After the first month, the work shifts into a steady rhythm of training, optimisation, implementation and support. That looks like:

• Training: keeping the context current as your business changes. New services, new prices, new staff. An AI that knows last year’s business gives last year’s answers.
• Optimisation: reviewing results and tightening what’s underperforming. Response rates, booking rates, time saved.
• Implementation: building the next system from the diagnosis list, one at a time.
• Support: someone to call when something breaks or a tool changes underneath it.

This is the difference between an AI partner and a project vendor. A vendor ships and leaves. A partner stays while the system grows with the business.

Adding the second, third and tenth system

Each new system is cheaper and faster than the last, because the foundations are already in place. The context exists. The data connections exist. Your team already knows how to work alongside the AI. The first build might take two weeks. The fifth might take three days.

That’s the compounding effect. Over time, you end up with an AI brain that knows the business and an AI workforce that does the repeatable work, all pointed at the leaks the diagnosis found in the first place.

How to know it’s working

I track three numbers with every client:

1. Away-from-desk hours: how long you can step away before something needs you. This should climb steadily.
2. Task automation share: the percentage of mapped recurring tasks the system now handles or heavily assists. Feeling the difference usually starts around 20 to 30%.
3. Revenue per person: total revenue divided by headcount, contractors included. If the system works, output grows without matching growth in staff.

If none of these move after a few months, something in the process is wrong, and a good partner should be the first to say so.

How to Judge an AI Consulting Process Before You Sign

You don’t need to understand the technology to tell a good AI consulting process from a bad one. You just need to ask the right questions and listen for vague answers.

Questions to ask on the first call

• What will I have in my hands after the diagnosis? You want a written document naming your first system, not “insights” or “a roadmap workshop”.
• What’s the first thing you’ll build, and how long will it take? A clear answer in weeks is good. “It depends” with no follow-up is not.
• How will we know it worked? Look for a named number: hours saved, response time, bookings, revenue recovered.
• What happens after go-live? If the answer is “we hand it over, and you’re good to go”, expect to be on your own the first time it breaks.
• Who owns the system and the data? You should. Your context files, your data and your workflows should stay yours if you ever part ways.

Your accountant will have their own list. I’ve covered the commercial side in the questions your accountant should ask before you sign an AI contract.

Red flags

• The same solution for everyone. If the proposal reads identically for a plumber and a law firm, there was no diagnosis.
• A tool pitch dressed up as strategy. If the conversation keeps coming back to one platform before anyone has looked at your business, you’re being sold a product.
• A long timeline with nothing shipping. Months of discovery with no live system is a sign of either over-scoping or under-capability.
• No mention of your team. If the plan ignores the people who’ll use the system day to day, adoption will fail.
• Promises of full autonomy from day one. Anyone who says the AI will run unsupervised immediately hasn’t run one in a real business.

Green flags

• They ask more about your business than they talk about AI.
• They can show you systems they run in their own business, not just demos.
• They push back on your ideas when something isn’t worth automating.
• They name one first build and are happy to stop there if it doesn’t deliver.

Consultant, agency or partner?

The labels blur, but the difference matters. A consultant typically diagnoses and advises. An agency typically builds what you ask for. A partner does both and then stays involved while the system runs. For most founder-led businesses, the third option is the one that changes how the business actually operates, because the value comes from the system improving over months, not from a one-off build. I wrote more on that distinction in what an AI automation agency actually does.

Cost without the sticker shock

Pricing varies a lot across the market, so I won’t quote figures here. The useful comparison is this: a well-scoped first system should cost less than a part-time hire, and it should show you a measurable return within the first couple of months. If the numbers in a proposal don’t make sense against what the leak is costing you, ask the consultant to walk you through the maths. If they can’t, walk away.

Conclusion

The AI consulting process isn’t complicated. Diagnose the business properly, build one system that fixes a real leak, then run it, tune it and add the next. Diagnose, build, run.

Most AI projects fail because they skip a step. They buy the tool before the diagnosis, or they build without capturing the context, or they hand over and vanish. Each of those mistakes is avoidable if you know what to look for.

The bigger point is what you end up with. Not a pile of disconnected tools, but an AI brain that knows how your business thinks and an AI workforce that handles the repeatable work. That’s what gets you out of the Operator Trap. The business stops running on your head and starts running on a system, and you get your bandwidth back for the work that actually grows it.

Ready to See What Your Diagnosis Would Show?

If you’d like to know where your business is leaking time and money, and what the first system would be, book a 30-minute Discovery Call. No pitch deck. I’ll ask about your week, tell you honestly whether AI is the right fix, and if it is, show you where I’d start.

Frequently Asked Questions

About Octavius