Comment to add before publication:
Revise before publication. Remove unsupported LeenAI claims about delivery timelines, billing policies, on-premises availability, ownership, and penalty-free exit. Replace claims that build is usually the smallest cost, fixed pricing guarantees cost certainty, or on-premises deployment lowers risk with explicit assumptions and contractual checks. Define a consistent TCO period, distinguish incremental costs from existing costs, and prevent double-counting review, maintenance, and governance. Include usage drivers, failed attempts, third-party licences, support, migration, and decommissioning. Clarify that a pilot measures selected costs while longer-term costs remain estimates. Use authorised representative evaluation data, document tested versions and evidence dates, and distinguish capacity released from cash savings. Tie acceptance, remediation, expansion, and exit to written purchasing decisions. The Real Cost of Owning an AI Agent in Saudi Arabia: A Buyer’s Guide to TCO, Acceptance Criteria and Evals Why is the build price only part of the decision? An AI agent proposal may include a clear implementation fee while leaving its operating costs less visible. Model usage, integrations, human review, monitoring, and future changes can materially affect the business case.
Total cost of ownership, or TCO, brings those costs into one model over a defined period. It helps answer a practical question: what will it cost to implement, operate, control, and eventually replace this workflow?
For Saudi enterprises, the assessment should include the organisation’s data-protection, security, procurement, and operational requirements. Their cost depends on the workflow, data, architecture, and responsibilities agreed with the supplier.
Start with one workflow and one measurement period. A twelve-month estimate and a three-year estimate answer different questions. Neither should be presented without its scope and assumptions. What belongs in the four layers of AI agent TCO? Build, run, govern, and exit provide a useful structure. They are planning categories; individual activities may span more than one layer. Layer Costs to consider Build Workflow design, data preparation, integrations, configuration, evaluation, security review, and user acceptance testing Run Model usage, hosting, search infrastructure, licences, support, human review, monitoring, and maintenance Govern Access reviews, privacy assessments, audit-record management, change approval, incident preparation, and recurring testing Exit Export, migration, replacement testing, handover, access revocation, deletion, and decommissioning
Assign each cost once, even where several teams contribute. For example, a maintenance agreement may already include monitoring and routine evaluations. Adding separate estimates for the same work would inflate TCO.
Also distinguish additional expenditure from existing costs. An existing employee’s review time is an economic resource, but it is not automatically a new cash payment. Finance should agree how both are represented. Which assumptions drive operating cost? Transaction volume alone is insufficient. A short document lookup and a long, multi-step investigation may have very different resource requirements.
Ask what drives the supplier’s charges: input and output tokens, audio minutes, document pages, storage, tool calls, users, or another unit. Include retries, failed processing, testing, and reprocessing after changes.
Human effort needs the same attention. Count preparation, review, correction, escalation, and follow-up. An output that takes seconds to generate may still require substantial checking.
A useful operational measure is:
Cost per accepted output = allocated workflow cost ÷ outputs meeting the agreed acceptance standard
Define “accepted” consistently. Report outputs accepted immediately separately from those accepted after correction. Failed and abandoned tasks still incur costs and should remain visible. How can buyers build a transparent TCO estimate? Separate initial expenditure, recurring costs, usage-dependent costs, and expected transition costs.
For an agreed period:
TCO = initial implementation + recurring operation + internal effort + governance and change costs + exit costs
Use this structure consistently. If human review is already included in operation, do not add it again under internal effort.
An illustrative twelve-month estimate might look like this: Cost item Illustrative amount Implementation and integration SAR 70,000 Hosting, model usage, and software SAR 24,000 Internal review and administration SAR 36,000 Maintenance and recurring evaluation SAR 18,000 Additional governance activities SAR 12,000 Estimated exit and decommissioning SAR 10,000 Total SAR 170,000
These are hypothetical figures, not LeenAI prices or market benchmarks. Assume the categories do not overlap. Some amounts may represent allocated staff time rather than external payments, and exit expenditure may occur beyond the first year.
Show the timing of costs separately when assessing budget needs or cash flow. What do acceptance criteria protect? Acceptance criteria connect the proposed system to the business need. They define the conditions under which the buyer will accept delivery for a specified scope.
For a quotation workflow, criteria might cover field accuracy, pricing-rule compliance, review time, correct escalation, and enforcement of sending permissions.
Specify the test population, measurement method, thresholds, permitted exceptions, and decision owner. Include critical failures that cannot be offset by a strong average score.
User acceptance testing should involve the people who will operate and review the workflow. Record the tested configuration, dates, results, and unresolved findings.
Acceptance establishes evidence for the assessed scope and version. It does not guarantee future performance after data, tools, models, or operating conditions change. How should evaluations continue after launch? Use a representative evaluation set with reviewed reference outcomes. Include authorised historical examples alongside constructed cases for rare failures, missing information, conflicting sources, and prohibited actions.
Keep evaluation cases separate from examples used to tune the system where practical. Otherwise, apparent improvement may reflect familiarity with the test material.
Track both quality and cost. A configuration change may improve accuracy while increasing processing expense or review time.
Retest after material changes to models, instructions, tools, permissions, or knowledge sources. Assign responsibility for investigating failures and approving corrective action.
An acceptance pack should bring together baseline measures, test results, known limitations, operating instructions, and the release decision. Its contents and contractual significance should be agreed explicitly. Does fixed pricing make TCO predictable? Fixed pricing can make a defined portion of delivery more predictable. It does not establish the total cost of operating the system.
Check what the price includes, its assumptions, usage allowances, exclusions, dependencies, and change-control process. A fixed build fee may coexist with variable hosting, support, and consumption charges.
Time-based billing can also be bounded through spending limits, milestones, reporting, and approval requirements. The important question is how expenditure is controlled and reviewed.
Request conservative, expected, and higher-demand estimates. Include scenarios where adoption is slower, review takes longer, or integrations require additional work. How do architecture and ownership affect cost? Compare deployment options against actual requirements. On-premises operation may introduce hardware, staffing, patching, availability, and recovery costs. Managed services may introduce consumption charges, contractual dependencies, and provider-change risks.
Read-only access can limit modification risk while leaving disclosure and access-control risks to address. Hosting location alone does not establish the security or suitability of an architecture.
Ownership also needs precise terms. Distinguish customer data, project deliverables, pre-existing supplier components, third-party software, and hosted services.
Confirm the rights needed to operate, modify, export, and migrate the solution. Ownership of selected deliverables does not by itself ensure portability or an inexpensive exit. What should the purchasing decision contain? Present the measured baseline, the proposed scope, the TCO range, and the evidence supporting expected benefits.
Separate cash savings from capacity released. Faster drafting may create time for additional work without reducing payroll. Revenue benefits should have their own assumptions and should not duplicate the value already attributed to that time.
A pilot can measure selected operating costs and quality outcomes. Longer-term maintenance, renewal, migration, and exit costs will usually remain estimates that need updating.
Possible outcomes include limited deployment, remediation and retesting, further evidence gathering, or stopping. Expansion should follow a separate review of scope, cost, and risk.
To discuss a defined workflow, contact LeenAI or explore pilot scoping. Bring the systems involved, expected volumes, review responsibilities, and operating constraints so the proposal can state what is included, what remains variable, and what evidence will support the next decision.




