Why discovery before build: the single biggest predictor of whether AI actually works

The pattern in failed AI projects

Every failed AI project I have ever seen — and I have seen a lot — fails at the same step. Not the build. Not the model. Not the infrastructure. Always the diagnosis.

Someone reads about AI doing something impressive. They decide AI is the answer to their problem. They commission a vendor (or worse, ask their internal team to "investigate AI"). The vendor builds what they were asked to build. Six months later, the build is deployed and nobody uses it — because the problem the build solved was not the real problem.

The remedy is unglamorous: spend 45 minutes diagnosing before you commit to building.

What a real discovery call covers

AIMindShift's discovery call is a 45–60 minute technical conversation structured in five sections. Tight enough to finish on time. Long enough to surface the bottleneck.

Section 1: Context (10 minutes)

I ask you three questions:

1. What is your operation? (Headcount, revenue band, sector, customers.)
2. What changed recently? (Growth, regulation, churn, founding-team transition.)
3. Why are you asking about AI now? (The trigger always tells me something.)

I am not interviewing you. I am building enough of a model of your business that the rest of the conversation has signal. The shape of your answers tells me which framework to apply.

Section 2: Operations mapping (20 minutes)

We walk through your operation from the outside in. The inbound channels (calls, emails, web, referrals). The internal workflow (qualification, assignment, execution, follow-up). The outbound channels (delivery, billing, post-engagement). The data flow underneath all of it.

Throughout, I am tagging each step against three questions:

• Where is senior judgement being spent on clerical work? (Headcount tax.)
• Where is the workflow duct-taped between SaaS tools? (SaaS-sprawl tax.)
• Where are decisions waiting on one person? (Founder bottleneck tax.)

Sometimes you already know where the leak is. Sometimes you don't, and the conversation surfaces it.

Section 3: Ranking bottlenecks (15 minutes)

We pick the three highest-leakage points and size them. Concrete questions:

• How many hours per week per person?
• What is the senior salary band of those people?
• How many missed opportunities or delays per month does this create?
• What does each missed opportunity cost in revenue?

By the end of this section, we have three bottlenecks with rough hour and £ figures next to them.

Section 4: Build / no-build decision (10 minutes)

For each of the three bottlenecks, I tell you which of four options is the right next step:

1. Process change. The bottleneck is the process itself; AI is not needed.
2. Hire. The work needs judgement that cannot be encoded.
3. Off-the-shelf SaaS. The workflow matches a template; bespoke is overkill.
4. Bespoke AI build. The work is repetitive and rule-based, the rules are specific to you, and the volume justifies the build cost.

Some bottlenecks get answer 1, 2, or 3. That is fine. I am not trying to sell you a build. I am trying to give you the right next step.

Section 5: Scope (5 minutes)

For any bottleneck that scored "bespoke AI build" in the previous section, I give you:

• Approximate cost range (e.g. £25k–£45k)
• Approximate timeline (e.g. 6–10 weeks)
• Approximate ongoing cost (hosting + AI model usage)
• What success looks like in measurable terms

And we agree what happens next — either a proposal in writing within 48 hours, or a clean walk-away with the bottleneck map.

What you leave with: the costed bottleneck map

A one-page document covering:

1. Your three biggest operational bottlenecks, each named and sized.
2. The cost (hours and £) of each, per month.
3. For each, the recommended next step: process / hire / SaaS / bespoke.
4. If bespoke: cost range, timeline, scope.
5. If not bespoke: who or what you should look at instead.

This document is yours regardless of whether AIMindShift builds anything for you. It is a tangible deliverable from an hour you would otherwise have spent on internal speculation.

What this is not

It is not a generic operations audit. It is not strategic consulting. It is not a sales meeting wrapped in friendlier language. It is a technical conversation between an engineer and an operator, designed to identify whether bespoke AI is the right tool for any of three named bottlenecks.

Why it works (when most equivalents don't)

Most "AI strategy sessions" or "AI audits" you find advertised do the opposite of what is useful: they give you a 40-page deck with broad recommendations and nothing actionable. The output is impressive-looking and useless.

The discovery call is the opposite. The output fits on one page. Every line is actionable. Most of the time, the action is clear within 48 hours of the call.

Worked example

Anonymised: a 7-person UK professional services firm engaged me for a discovery call after watching their managing director burn his weekends on quote drafting.

Section 2 mapping surfaced three bottlenecks:

1. Quote drafting (MD's 6 hours/week)
2. Inbound enquiry triage (operations lead's 8 hours/week)
3. Monthly client reporting (every fee-earner's 2 hours per client per month)

Section 4 verdict: bottleneck 1 was actually a process problem (the MD had not delegated the criteria); bottleneck 2 was a clear bespoke build candidate; bottleneck 3 was perfect for an off-the-shelf reporting SaaS the firm already owned but had not configured.

The MD walked away with: a sheet of criteria he could hand to his ops lead (process change, free), a costed build proposal for the enquiry triage (£22k bespoke build), and a 30-minute follow-up he could run himself to configure the reporting tool (~£0 cost).

AIMindShift quoted £22k. The firm engaged the build. The other two bottlenecks were fixed without AIMindShift — and the firm did not pay for advice it did not need.

The decision is yours

After a discovery call, you have everything you need to decide whether to commission a build, fix the bottlenecks yourself, or do nothing. There is no follow-up sequence. There is no second call to "pitch you the proposal". If the answer is yes, you ask for a proposal; if it is no, you keep the bottleneck map and move on.