Ting Guides · For the CTO and CIO
What should an AI solution design document contain?
The short answer
Enough for an engineer to build the system and for a tester to reject it. That means the task and its boundaries, the users and their workflow, the data and who may access it, every system it reads from or writes to, what it does when unsure, the permission behind each action, and acceptance criteria written as tests.
Reviewed by Abdullah Al-RashedUpdated
What it depends on
Who will build from it
If your own team builds, the document can lean on shared context. If a vendor builds, or you are tendering, it must stand alone: every assumption written down, every term defined.
Whether it acts or only advises
A system that suggests needs output formats and review screens. A system that acts needs, for each action, the permission, the limit, the log and the way to undo it. OWASP traces "excessive agency" to too much functionality, permission or autonomy; the document is where you cap all three.
How many systems it touches
Each integration adds a section: the API or export used, the fields mapped, what happens when the other side is down, and who owns that system.
The risk of the use
Higher-risk uses spend more of the document on failure handling, review and monitoring, and less on features.
Sections a Saudi design document should not skip
General information, not legal advice. Each regulatory point summarises the official source cited beside it, as checked on the date shown under Sources; confirm how it applies to your company with your legal adviser.
Data location and personal data
Name where each component processes data, the model provider included. If personal data may leave the Kingdom, state the purpose and the conditions relied on under Article 29 of the Personal Data Protection Law, or the exemption and safeguards relied on under the Transfer Regulation, in the document itself, not in a later email.[2][3]
Arabic acceptance cases
Write acceptance tests in the Arabic your users actually produce, dialect messages and angled scans included, with their own pass marks, separate from English.
Lifecycle and risk category
SDAIA's AI Ethics Principles describe an AI system lifecycle of plan and design, prepare input data, build and validate, and deploy and monitor, and sort uses by risk. The design document is the first stage's output: it sets the tests and monitoring for the other three and records the risk category it assumes.[1]
Inside one design document: referral intake at a group of clinics
ExhibitWorked example
Rasikh Care, the six outpatient clinics of the fictional Rasikh Holding, specifies a referral-intake assistant. One line from each section, and the document opened at its acceptance tests.
AI Solution BlueprintBP-RC-03 · v1.2
- 01Task and boundaryRead referrals that arrive by email, scanned letter and WhatsApp at six clinics, extract the agreed fields and route each one to the clinic and specialty it names. It never books a slot, gives medical advice or decides urgency.
- 02Users and workflowReferral coordinators clear exceptions in a review queue; the duty nurse sees every referral marked urgent before anyone else.
- 03Data and accessReads the referrals inbox and the clinic and specialty directory. No access to clinical notes, results or accounts.
- 04IntegrationsCreates a referral record in the booking system through its API. If the API fails, the referral stays in the queue and the booking-system owner is alerted.
- 05UncertaintyAny field below its confidence threshold, or a specialty no clinic offers, goes to a coordinator. Thresholds are set per field during testing.
- 06Permissions and logOnly a named coordinator can release a referral. Every extraction, edit and routing decision is logged with user and time, and can be reversed.
- 07Acceptance tests
07Acceptance tests
AT-07Pass mark
- Given
- A handwritten Arabic referral, scanned at an angle
- When
- The assistant reads it
- Then
- The patient's file number and the requested specialty are read correctly, or the referral goes to a coordinator. It is never routed to the wrong clinic.
Pass mark: 42 angled scans · every one read correctly or sent to review · 0 routed wrong
AT-08Pass mark
- Given
- A referral marked urgent, by email, letter or WhatsApp
- When
- The assistant reads it
- Then
- It reaches the duty nurse first, whatever else it says, and never waits behind routine referrals.
Pass mark: every urgent referral reaches the duty nurse first · 0 held in the routine queue
Referral intake and routing · BP-RC-03Page 14 of 22
The decisionReady for approval: every section holds a line a tester can check, and the hardest case has a written pass mark.
What Ting does and doesn't do here
What we do
- We write this document as the AI Solution Blueprint: workflow, architecture, data sources, integrations, permissions and acceptance criteria written as tests, ending in a defined implementation quote.
- We turn each phrase of the business request into a line an engineer can build and a test can check.
- We hand it over so you can build from it with Ting, with your own team or with another vendor.
Where Ting helps: AI Solution Blueprint
What we don't do
- We don't build without one. A paid, approved Blueprint comes before any Ting-built implementation.
- We don't require a Transformational Audit first. A Blueprint can be commissioned directly.
- We don't quote an implementation before the specification exists. The Blueprint is what fixes scope and quote.
Sources
Primary sources only, each read in full on the date shown.
- AI Ethics Principles, version 1 (May 2025)SDAIA · checked
- Personal Data Protection Law (Royal Decree M/19 of 9/2/1443H, amended by Royal Decree M/148 of 5/9/1444H)SDAIA · checked
- Regulation on Personal Data Transfer outside the Kingdom, version 2.0 (August 2024)SDAIA · checked
- LLM06:2025 Excessive AgencyOWASP GenAI Security Project · checked
