Your AI programme has a data problem, not a model problem
Organisations rarely fail because the model is inadequate. They fail because no single department agrees on which customer record is authoritative.
Square Kufic module · move your cursor
Taqania is a Saudi technology firm working across three connected disciplines: artificial intelligence and data, IT management, and technology consulting. Most of our work involves all three, because a recommendation that nobody can operate is not a recommendation.
We remain accountable after the strategy document is delivered. That single commitment shapes how we scope, how we price, and which engagements we decline.
Each one stands alone. Most organisations start with an assessment and decide what follows once they can see the picture clearly.
Use-case selection, data foundations, model deployment and the governance around it. We start with data quality and access controls, because that is where most AI programmes stall — not in the model.
See the work→We operate infrastructure, applications and cloud environments against agreed service levels: monitoring, incident response, patching, capacity planning and cost management, reported monthly.
See the work→Technology strategy, target architecture, vendor selection and transformation roadmaps — delivered as decisions with owners, costs and sequencing attached.
See the work→Legacy systems separated in stages rather than replaced at once, and connected to the platforms the organisation already depends on, so operations continue while the architecture changes.
See the work→An AI deployment with its guardrails, a managed service reported against its targets, and the output of a consulting assessment.
# assistant.yaml — retrieval assistant, governed deployment model: claude-sonnet languages: [ar, en] retrieval: source: policy-archive access: inherit-from-user # no privilege escalation citations: required guardrails: pii_redaction: on out_of_scope: "refuse and route to service desk" confidence_floor: 0.72 evaluation: golden_set: 480 questions, human-labelled accuracy_gate: 0.91 # deployment blocks below this review_cycle: monthly audit: log: prompt, sources, response, reviewer retention: 24 months
AI READINESS ASSESSMENT — summary of findings Data foundation ███████░░░ 3 / 5 Governance ████░░░░░░ 2 / 5 Platform ████████░░ 4 / 5 Skills and operating █████░░░░░ 2.5 / 5 RECOMMENDED SEQUENCE 1. Consolidate customer records 8 weeks 2. Establish access and retention policy 4 weeks 3. Pilot: Arabic document processing 10 weeks 4. Reassess before further investment — NOT RECOMMENDED YET Customer-facing generative assistant — insufficient data governance to manage the risk responsibly.
We work alongside our customers' teams rather than around them. These are results measured after handover, not at go-live.
Retrieval with mandatory citations, restricted to each user's existing permissions, so nobody gained access to material they could not already open.
Most alerts were not failures. Rewriting the monitoring thresholds before adding staff removed the noise rather than escalating it.
The assessment found that two of five planned workstreams addressed problems the business had already resolved. Removing them shortened the programme.
Organisations rarely fail because the model is inadequate. They fail because no single department agrees on which customer record is authoritative.
If most alerts require no action, additional operators simply learn to ignore them faster. Repair the thresholds before expanding the team.
Every recommendation should name the person accountable, the cost, and what happens if it is deferred. Anything less is an opinion in presentation form.
We are not tied to a single vendor. Selection follows the assessment, not the other way round.
With an assessment at a fixed fee, lasting between two and four weeks depending on scope. It produces a written set of findings and a recommended sequence of work. If the finding is that the investment should be deferred, we state that clearly and you retain the document.
Assessments are fixed price. Delivery runs in defined stages, each priced separately, so you may stop at any stage boundary. Managed services are priced against agreed service levels on an annual contract.
Primarily enterprises and government entities where a technology failure affects operations beyond the IT department. We are not well suited to early-stage products that are still establishing their market.
The consultants and engineers you meet during the assessment. We do not substitute less experienced staff after a contract is signed, and the people who design a system remain accountable for its operation.
In most cases, no. The majority of organisations obtain better results by applying established models to well-governed internal data than by training their own. We will say plainly when a custom model is justified, which is uncommon.
Arabic is treated as a primary requirement rather than a later addition. This affects document processing, retrieval quality, evaluation sets and interface design, and it is assessed explicitly rather than assumed to work.
Access controls inherited from existing user permissions, mandatory source citations, redaction of personal data, defined behaviour for out-of-scope requests, an accuracy threshold that blocks deployment, and a full audit log. These are configured before launch, not afterwards.
Monitoring, incident response, patching and updates, capacity planning, cost management and monthly reporting against agreed service levels. Scope and exclusions are defined in the agreement so that responsibility is never ambiguous during an incident.
Yes, and this is the more common arrangement. We typically assume responsibility for defined services while internal teams retain the areas closest to the business, with responsibilities documented rather than assumed.
You receive operational documentation, configuration and access, and a transition period during which your team or another provider takes over. We do not retain control of anything required to operate the service.
Send a short description of the problem and the constraint you are working within. A consultant replies within two working days, in Arabic or English.