Included in a typical proposal
- Assessment and a prototype on the sample you provide.
- Integration labour if you proceed, as a second stage in the proposal.
- An oversight design.
SOFTWARE DEVELOPMENT
AI-assisted business tools
in Dubai & the UAE.
Apply AI to a specific task with clear boundaries and human responsibility. Assess the data, accuracy needs and operational fit before integrating a model into everyday work.
WHAT THIS SERVICE IS
A plain explanation of the work.
AI-assisted business tools means applying a model to a defined task with a human still responsible for the result: drafting a first reply that someone sends, classifying a ticket for a person to handle, extracting fields from a document for a clerk to confirm. It is not an autonomous agent running the company, and it is not a claim of accuracy we have not tested on your examples.
The brief must include data rules. Customer content may not be allowed in a consumer chatbot. We will not paste your files into a public tool to “try it”. Provider choice, retention, and whether you have an organisational account are part of discovery.
Failure is normal: wrong totals, confident nonsense, leaked prompt instructions. Oversight, a fallback, and a way to turn the feature off are in the first release, not in a later ethics workshop.
A CLEAR PICTURE
A labelled diagram, not a decoration.
-
Write the task as a marking scheme
If you cannot mark it, you cannot operate it.
-
Choose a provider you can contract
On an organisational account.
-
Test on held-out examples
Including nasty ones.
-
Ship with a human confirmation step
Until you have evidence to loosen it — and maybe never.
WHO IT IS FOR
The situations this page is written for.
Teams with a high volume of similar documents or messages, and managers who can accept a reviewed draft rather than a magic button. It is a poor fit if you cannot describe a correct answer for a sample set.
- Operations that can supply real (approved) examples and counter-examples.
- Businesses with a policy on where data may go.
- Products that need an assistive feature, not a gimmick on a homepage.
- Owners who will keep a person in the loop.
TYPICAL SCOPE
What a proposal usually names.
01
Use-case assessment
The task, the input, what “good” means, and whether a rules engine would suffice. If it would, we will say so and point at automation.
02
Data and access planning
Approved sources, masking, provider, and logging. Training on your data is a separate, explicit decision. We do not quietly opt you into a vendor’s training corpus.
03
Prototype and evaluation
Run the sample set, show misses, and agree a threshold for going further. No public accuracy percentage is invented for marketing.
04
Integration and oversight
Put the model behind the existing application or a small tool, with review UI, monitoring you can read, and a kill switch. Cost per call is estimated from the sample, not promised as a bill.
HOW THE WORK RUNS
The sequence we follow once the brief is clear.
-
Write the task as a marking scheme
If you cannot mark it, you cannot operate it.
-
Choose a provider you can contract
On an organisational account.
-
Test on held-out examples
Including nasty ones.
-
Ship with a human confirmation step
Until you have evidence to loosen it — and maybe never.
INCLUDED AND QUOTED SEPARATELY
Labour versus things that sit on their own line.
Quoted separately
- Foundation-model research or training your own large model.
- Unattended decisions over money, safety or HR.
- Any accuracy, saving or headcount figure.
- Using AI to generate this website’s marketing claims.
AFTER HANDOVER
What you should hold when the work is done.
Prompts, evaluation notes and provider accounts are yours. Keys are rotatable. If the feature is off by default, that is a valid launch.
Models and prices change. A quarterly review of the sample set is wiser than never looking again. That review can be maintenance.
PRACTICAL NOTES
Details that usually affect the proposal.
Prompt text is an artefact you should keep in version control like any other configuration. If only we hold the prompt, you do not own the behaviour. Evaluation examples belong beside it so a later change can be scored. That sounds fussy; it is how you notice a silent quality drop after a vendor model update.
Cost control is a cap and a log, not a hope. We can set limits on the provider account you own. We cannot freeze their unit price. If the task can be done with a cheaper model, the evaluation will say so. Starting with the most expensive model is rarely a requirement.
PLANNING THE WORK
Questions worth asking.
Can it replace our support team?
Not as a responsible default. It can draft or route. People still own the customer.
Will you use our data to train a public model?
Not unless a contract you sign says so. We treat that as a no unless you instruct otherwise in writing.
Do you build chatbots for the website?
Only with a defined knowledge source, a handoff to a person, and no invented policy answers. Many businesses need a better FAQ page first.
CONNECTED SERVICES
Explore the next piece.
LET’S BUILD WHAT’S NEXT
Your next step starts
with a conversation.
Tell us what you want to improve, build or simplify. We’ll help you define a practical way forward.
Discuss your project