To choose an AI consultant, look for one who starts from one task and a money number, tests more than one model on your own messy cases, builds the first working piece inside the tools your team already uses, and gives you a fixed price for a defined first step. Walk away from anyone who leads with a strategy deck, a favourite model or an open-ended discovery phase.
I run Skcript Applied Labs, so I am one of the people you might be choosing between. That is an obvious bias. I'll try to write the guide I would want if I were the one hiring, and I'll point out where a different kind of firm is the better fit.
The short checklist
If you only have five minutes before a call, check these.
- They ask about the task before the technology. The first questions should be about volume, time per item, error rate and cost, not which model you want.
- They name the number they plan to move. Hours per week, cost per invoice, error rate on claims. Something you can measure before and after.
- They test on your work. A test set built from your real cases, messy ones included, with the right answers attached. Not a public leaderboard.
- They are not tied to one model. They can tell you which models they will compare and why.
- They build the real thing, small. The first deliverable runs inside your inbox, ERP or ticketing tool, not a chat window on the side.
- The people on the call build the work. You meet the engineers, not only a partner and a slide.
- You own what they build. Code, test cases and numbers, from day one.
- The first step has a fixed price and an end. You know what you pay, what you get and when you decide.
- They will tell you to stop. If the numbers are poor, the pilot ends with a plain recommendation, not a bigger proposal.
None of this is exotic. It is the same discipline that closes the five gaps where AI pilots stall. A consultant who can't show you how they close each one will probably leave you in one of them.
Red flags
These are the patterns I would walk away from, or at least slow down on.
A strategy deck before anyone has seen your work
A roadmap written before anyone has sat with the people doing the task is a guess with nice fonts. Strategy has its place, but for a first AI project the useful output is a measured task and a decision, not a vision.
One model for everything
"We use the best model" is not an answer. Ask which one, tested on what. Models change every few months, and the right one for reading supplier invoices may not be the right one for sorting customer emails. A consultant who always lands on the same vendor may have a reason that has nothing to do with your work.
Demos on clean sample data
If every demo uses neat examples, ask to see it on yours. Real work has crooked scans, missing fields and emails that switch languages mid-sentence. A system that has only seen clean cases will look strong in the room and struggle with real volume.
Open-ended discovery
"We'll start with a discovery phase and scope from there" can be fine. It can also be where budgets disappear. If discovery has no fixed price, no end date and no defined output, you are buying time, not answers.
Nobody can say what you keep
Ask what you are left with if the project stops after the first step. If the answer is "a report", you are paying for a document. If it is "the starting numbers, the test set and the working first piece", you are paying for something you can use again.
The sales team and the build team never meet
Large programmes often separate the people who sell from the people who build. That is not always a problem, but you should know it before you sign. Ask who on the first call will write the code.
No answer on data
Every consultant should ask early where your data is allowed to go and how sensitive it is. If they don't, they will find out later, usually after the demo. I won't repeat the UAE specifics here; if that is your market, read AI consulting in the UAE, which covers data residency in detail.
Questions to ask before you sign
Here are the questions I would put to any firm, including us. Good answers are short and specific.
Which number will you move, and how will you measure it? Look for a baseline measured on day one and the same measurement repeated at the end. I explain how we count this in cost per task, not cost per token.
Which models will you test, and on what? You want a comparison on your own cases, with accuracy, the kinds of mistakes each model makes, cost per task and where the data goes. Here is how we test AI models on your work, if you want a reference point.
Do you earn more if we pick one vendor over another? Partnerships, resale margins and referral fees are all normal in this industry. You just want to know about them.
Who on this call builds the work? On our side, the engineers on your first call are the ones who build, and I am on that call too.
What do we need to give you? A good consultant is specific: one owner who knows the task, a set of real examples, access to the systems involved and a few hours a week. Vague answers here usually mean the scope is vague too.
What do we keep if it stops? The starting numbers and the test set should stay with you. The test set alone lets you check any AI tool you try later.
What is the price, and which part is fixed? Ask for a fixed price for the first step, and ask how the next step will be priced before you commit to it.
When would you tell us not to do this? The best answer is a list. Mine is in when not to use AI.
Fixed price vs retainer, fairly
I am biased toward fixed prices, because that is how we sell our first step. So here is the fair version.
| Fixed price for a defined step | Monthly retainer | Time and materials | |
|---|---|---|---|
| Best for | A first project with a clear task and end | Ongoing work: watching a live system, testing new models | Work whose scope truly can't be known upfront |
| What you know upfront | Price, output and end date | Monthly cost | Day rate |
| Main risk | Scope set too narrow, so change means a new agreement | Hard to tell when you are done, or what moved | Spend grows with every unknown |
| What to ask for | A written definition of done and a stop point | A monthly report on what changed and what it was worth | A cap, and a check-in before you reach it |
A retainer is not a bad thing. Once a system is live, someone has to watch it, fix issues and test new models as they ship, and that is ongoing work. We offer that too, as monthly care with a three month minimum. Where retainers go wrong is as the starting point for a first project, because nobody can say when it is finished.
Time and materials is honest when the work genuinely can't be scoped. It is less honest when it replaces the effort of scoping.
The pattern I trust most is a fixed price for a short first step, a fixed price for the rollout agreed before it starts, and a retainer only once there is something live to look after.
What this looks like in practice
Our first step is a 10-day AI pilot on one task. It has one fixed, published price: USD 7,500 in the US, with fixed local prices for other markets, such as AED 27,500 in the UAE and EUR 7,000 in Europe. If you already tried AI and it stalled, there is a pilot check-up from USD 4,000 that tests what you have on your real work.
At the end you keep the starting and new numbers, the test set, a model scorecard, the working first piece in your tools, a runbook, and a fixed price and 4 to 6 week plan for the rollout. Or a plain recommendation to stop. The day-by-day version is in what a 10-day pilot actually delivers, and what happens after is in the AI pilot to production playbook.
We are not the right fit for everyone. If you want a multi-year transformation programme, or one person on your team long term, a different kind of partner suits you better. I compare the options in AI consultant vs AI agency.
The one question that matters most
If you only ask one thing, ask this: "What will we know at the end of the first step that we don't know now, and what will it cost to find out?" A good consultant answers with a number, a date and a price. Everyone else answers with a process.
You can see how we work on our AI consulting page. If one task is eating your team's week, start a chat with us on the site, or book a call through the Applied Labs page. Bring the task and a few real examples, and I'll tell you honestly whether we are the right people to help.
