Too many AI programs get approved on the strength of a demo. I’m Denis Salatin, founder of Lumitech in Dubai, and after watching that happen more than once, we changed the order: a working slice of the workflow comes before the client signs the main project contract.
That slice covers one scenario, runs on the client’s own data, and gets reviewed by the people who would use it. For us, that’s where AI prototyping services start, and the reviewers’ feedback decides the next step: what’s wrong with the output, which systems it has to reach into, and whether a limited pilot deserves a budget.
What an AI Pilot Project Has to Prove
An AI pilot project earns a production budget when it proves one workflow end to end, people included.
On one recent project, a legal knowledge assistant for a financial institution in the Middle East and North Africa (MENA) region, thousands of employees needed legal answers just to get through the day. Each answer meant the same small hunt: what’s the context, where’s the document, is this the approved version, and what is this person allowed to do?
At first, nobody touched a model. The team read the questions that reached the legal department and sorted them by business situation, because people ask about situations and rarely know which document applies. That sorting drew a rough line between what an assistant could answer and what needed a lawyer. I say rough on purpose: the legal team moved that line twice during the pilot.
One more decision came early. Explaining a policy and telling someone what to do next sound like the same job and still need different handling, so they went into separate layers. Mix them, and the assistant starts to sound like a lawyer while knowing nothing about the case.
The pilot stayed small, a handful of legal domains, so the legal team could check answer quality long before anyone said the word “rollout.” Business units that used to wait days for a basic answer started getting one in minutes.
AI Proof of Concept vs Pilot
An AI proof of concept tells you whether the model can do the job on sample data. A pilot puts the same model into a working week and measures what happens to the people who work with it.
Three counts describe that second question: answers produced, answers escalated, and minutes of review per answer. They show the flow of work. Whether the answers are usable takes a separate check against an agreed standard on a sample of real questions, including whether escalations went where they should, and that check only means something once three checks are in place.
Three Checks That Come Before Answer Quality
AI use case validation in a regulated workflow starts with three build decisions, and answer accuracy can only be measured once all three hold.
A superseded policy can produce an answer that reads as entirely correct, so document currency comes first. The index carries approved versions only.
Access is the second. Employees hold different clearances, so role-based access control works at the retrieval step, before anything reaches the model.
Retrieval-augmented generation
helps here because the model answers from the documents passed to it as context, so the application decides what the model sees.
Traceability is the third. Each answer links back to its source policy, so a reviewer can check both the source and whether the answer reads it correctly. In a regulated setting, that counts for more than a few points of model accuracy, which engineering teams don’t expect to hear.
All three had to work inside one rule from the client’s compliance team: no internal or legal data through a public AI service, and a cloud-agnostic design for whatever their regulators ask for later.
How We Scope the First Slice of AI Workflow Automation
AI workflow automation becomes measurable when the first slice covers a single question type from end to end. The narrower the slice, the easier those three checks are to hold.
Our scope rule reads the same on every project: one scenario, one user group, one system of record, and a written description of what the output has to contain. When a question falls outside that line, the assistant declines it and names the person who can help. Narrow, admittedly, and that is the point of a first slice.
That line holds even when wider coverage looks tempting. A declined question still takes someone’s time, because a lawyer has to answer it, but that cost is known and planned for. A wrong answer the model should have escalated costs more: someone has to find the mistake, and a few of those can cost the tool the team’s trust.
We Follow an Enterprise AI Pilot Past the Chat Window
A customer types “yes” in the chat, and that’s usually where a demo ends. For an enterprise AI pilot, it’s where the harder half begins, because the order still has to reach whoever prepares it.
We ran into this on a WhatsApp commerce project for a retail network in the Gulf Cooperation Council (GCC). Customers could pick products, pay, and track delivery without leaving the conversation, and the chat had to work with stock, customer, order, and logistics systems so store teams received orders they could act on.
On a smaller project, that’s where I’d look first. Keep the prototype to one order type and ask the people who pack and ship it: did the order arrive complete, and what did you have to chase? What they tell you is a good guide to the integration work ahead, and it’s the reason stock and fulfillment connections belong in the enterprise AI implementation estimate from day one. The same connections make up much of
AI in logistics work, and they’re easier to price once one order has been followed through the whole path.
What to Price Before Moving an AI Pilot to Production
Before moving an AI pilot to production, I ask the client how many runs they expect in their heaviest month. Nobody knows the exact number, but a rough peak figure is enough for our engineers to price the work.
Volume changes the review effort more than people expect, and escalated questions count too. Our AI pilot evaluation times the review on real output, including the questions that go to people, and estimates it for the peak month, since an average hides the queue.
Engineering costs don’t all grow the same way. Some, like model usage, tend to rise with every run, while others, such as hosting, monitoring, and support, often rise in steps once volume passes certain thresholds. That’s why we estimate each component at the expected load, and why multiplying the trial bill by a hundred tells you little.
At the end, the client has an agreed estimate for a month of finished work, including their own team’s time, and the first rollout is sized to match.
Write the AI Pilot Success Criteria Into the Proposal
AI pilot success criteria belong in the proposal, beside scope and price, with a date.
Before AI prototype development starts, we agree with the client on what they’ll see and when we’ll review it. Our AI pilot checklist fits on one page:
- The workflow, and where its approved output has to go
- The current time, cost, and answer quality, with the agreed target beside them
- The expected volume and the time allowed for human review
- The running cost and the person responsible for the process
- The review date, and what would lead us to expand, revise, or stop That last line is the one worth arguing over while nobody has anything invested in the answer. The baseline stays on the page as well, because a good prototype breeds ideas. Someone who came for faster contract review starts eyeing document search, and the review still has to report what happened to contract review time.
Enterprise AI Adoption Starts With Work You Can Judge
Enterprise AI adoption should begin with one recent request, the result that was approved, and the person who handled it.
That person can tell you how long the work took and where they had to make a judgment call, which gives the requirements a solid starting point. Our own delivery work started the same way. We had estimates, time records, and task updates, yet managers still couldn’t see when effort and progress drifted apart mid-project, which is the gap AI delivery intelligence tools are built to close.
For a first AI workflow automation project, pick a task your team can evaluate just as easily. Before anyone talks about the main contract, ask the people who did the work to review a result from your own records. That gives the next decision a firmer basis, and the contract ends up describing something everyone has already seen.
Denis Salatin is the founder of Lumitech, a Dubai-based software development company and the company behind Tolmete.
Top comments (0)