Auto Alpha AdvisoryFor every 33 AI proofs of concept a company launched, only four reached production. The finding comes from research IDC ran with Lenovo for the CIO Playbook 2025, a survey of IT leaders in Asia-Pacific. I'm quoting the numbers via CIO.com's March 2025 write-up, which puts it this way: 88% of the proofs of concept IDC observed never made it to wide-scale deployment. In the same piece, IDC's Ashish Nadkarni gave the reason without much diplomacy: "Most of the time the POC happens not because of a strong business case."
That number sits behind a question owners keep asking before they've hired anyone: what does an AI consultant actually do? I'm a CA(SA) who builds these systems, and I'd judge the job by what you're holding when it ends. An audit ends with a ranked shortlist of use cases, each with a rand cost and a payback period. A sprint ends with something running in production. A retainer has one number, reported monthly, that should be moving. If a proposal can't tell you which of those you'll have at the end, you're paying for a deck.
An audit ends with a ranked, costed shortlist
An AI readiness audit takes a few days. It turns a general sense that the business should be doing something with AI into a short, ordered list of things worth doing. Each line on that list should show:
- the process it touches, and what that process costs today in hours or rand
- the build cost
- the monthly running cost, meaning model API fees, hosting and licences
- what it gives back each month, in hours or rand
- the payback period, in months
- the person in your business who will own it
The order comes from payback. How interesting the technology is doesn't come into it. The top item is usually dull: invoice capture, debtor reminders, the weekly report somebody pastes together by hand every Monday morning. I'd take dull and shipped over clever and stuck.
The cost columns matter because cost is what stops these projects. In S&P Global's 2025 survey of companies in North America and Europe, cost was among the top obstacles respondents cited, alongside data privacy and security risk, according to CIO Dive's write-up. An audit that leaves cost out is answering a question you didn't ask.
A good audit also tells you what not to build. Sometimes the most useful line on the list says to use the ChatGPT seats you already pay for properly and come back in six months. If every audit a consultant runs ends in a large build, they're selling the build.
A sprint ends with something running in production
A sprint is a few weeks spent building one item from the shortlist. Mine run four to eight weeks with a fixed scope, price and deadline, as my services page sets out, and the deadline is the date it goes into production.
Get production defined in writing, because the word stretches. My working definition: it runs on your real data, inside your systems, without the consultant in the room, and someone on your team uses it on an ordinary Tuesday. A demo on sample data doesn't count. A prototype that only works on the consultant's laptop doesn't either.
This is where most AI work stalls. IBM's 2025 CEO Study, fielded between February and April 2025, found that only 16% of AI initiatives had scaled enterprise-wide. Scaling across a large company is a bigger job than getting one system live in a mid-sized firm, but it points the same way: building the thing is the easier half.
A sprint should close with a handover. The code sits in your repository, the accounts are in your company's name, the running costs go on your card, and there's a one-page note on how to switch it off. If the consultant keeps the keys, you've rented the system rather than bought it.
A retainer is judged on one number
A retainer is the ongoing engagement: a monthly fee for someone who keeps the systems running, builds the next item on the shortlist and answers questions before they turn into projects. Nothing forces an end date, so it's also the engagement most likely to drift.
What stops the drift is one number, agreed in the first month, with a baseline measured before any work starts and a report every month after that. It might be hours of admin returned per week, or the days between finishing a job and invoicing it. Choose the one the business cares about most, even if something else would be easier to move.
That kind of measurement is rarer than it should be. IBM's summary of its Q4 2025 Think Circle reports that only about 29% of executives said they could measure AI ROI confidently. A retainer without a number puts you with the majority, paying every month for a feeling.
The number I watch on my own work is small, but it's real. I built a portfolio intelligence pipeline for my own use, and when I built it in March 2026 it replaced one to two hours of daily research. A retainer report should lead with that kind of figure: plain, checkable, and in units the business already thinks in.
If the number hasn't moved after a quarter, you've learned something about the retainer. A quarter is the fair unit to judge one on, because a single month is too short for most changes to show, which is why mine has a three-month minimum before it goes month to month.
What should an AI consultant cost?
It should cost less than the payback the audit identifies, because the fee belongs inside that calculation. The build cost on the shortlist includes what you pay the consultant, and the payback period still has to make sense with it in.
Fees usually come in three structures, and each one decides who carries the risk.
A fixed fee for a fixed scope means you agree the deliverable and the price up front. The consultant carries the risk of the work taking longer; you carry the risk that the scope was badly written. It suits audits and sprints, where the output can be named in advance.
A day rate means you pay for time, so any overrun is yours. It suits genuinely open work, like finding out whether a messy data source is usable at all. Even then, ask for a cap and a stated output for each block of days.
A retainer is a monthly fee for a defined amount of ongoing work and access. It fits the period after a sprint, when there's a system to run and a shortlist to keep working through.
Two costs usually sit outside the fee. One is running cost: model API fees, hosting and licences, billed monthly and usually to you. Ask for a monthly estimate in rand for each system, and ask whether it rises with usage. I've broken those costs down in what AI really costs a South African business. The other is your team's time for testing and sign-off, which you pay in hours rather than invoices. And check whether the quoted fee includes VAT.
I don't put my own prices in posts, because prices change and a stale figure is worse than none. They're on the services page.
Read the proposal like an engagement letter
I read AI proposals the way I'd read any engagement letter, and three sections do most of the work.
Scope comes first: which processes, which systems, which data, and what is explicitly left out. A line about exploring AI opportunities across the business is not a scope.
Then deliverables, meaning the named thing you'll hold at the end. A ranked shortlist of use cases with build cost, running cost and payback is a deliverable. Strategic recommendations, with nothing more specific attached, are not.
Then acceptance criteria, which say how you'll both know it's done. For a sprint, that might be: it processes last month's supplier invoices from the live inbox, and your bookkeeper signs off the exceptions list. For a retainer, it's the number, its baseline and the day each month it gets reported.
After that, look for the terms that bite later: who owns the code and the accounts, who pays the running costs, and what happens to your data. If the consultant will handle your customers' personal information, POPIA's section 21 is short and specific: "A responsible party must, in terms of a written contract between the responsible party and the operator, ensure that the operator which processes personal information for the responsible party establishes and maintains the security measures referred to in section 19." A proposal that says nothing about data handling isn't finished.
The test before you sign
Before you sign, write down the one thing you expect to hold when the engagement ends: the costed shortlist, the running system or the number. Then find it in the proposal, with a date and a test attached. If it's there, you're hiring a consultant. If you can't find it, what you're buying is a deck.
Three questions in the first call tell you most of the same thing. What do you have running in production right now? What would you tell me not to build? What do I own at the end? A consultant who can't answer the first, or has never said the second, is selling the deck.