Agent Optimization
An AI agent that knows exactly where it falls short
Every call your agent handles generates a transcript. Agent Optimization reads those transcripts, finds the questions the agent could not answer — and the ones it answered confidently but should not have — then routes them to you for a quick review. Once you confirm the right answer, a compliance gate checks it, the agent learns it, and the next caller gets a clean response.
The result is a live score — the Agent Optimization score — that measures what fraction of your call types the agent now handles to completion. It moves up every time a gap closes.
How we built this
Why we built gap detection into the platform from the start
We deploy AI receptionists and outbound agents for UAE businesses — clinics, real-estate brokerages, home-services firms. In every deployment, the first few weeks tell a consistent story: the agent handles the common questions well, then a caller asks something that sits just outside the knowledge base. What happens next depends on how the agent is built.
A naive agent escalates or apologises. A worse one answers confidently from training data — fabricating a policy, a price, or a product detail that the client never stated. Both outcomes erode trust. The second one can create real problems, particularly in regulated sectors where what the agent says on your behalf carries weight.
Agent Optimization is what we built to close that gap systematically. The key insight was that "I don't know" is not the only failure mode worth catching. A confident-but-unverified answer is harder to detect and harder to defend. The gap-detection engine flags both, and keeps the client — not the AI — in control of what the agent says.
The QA and compliance gate is not cosmetic. Calling into UAE businesses means every confirmed knowledge-base entry gets checked against TDRA telemarketing regulations and data-handling obligations under Federal Decree-Law No. 45 of 2021 (UAE PDPL) before anything goes live. That is a design constraint, not an afterthought.
The mechanism
From call transcript to smarter agent — five steps
The loop runs automatically after every call. Your job is a single review step; everything else is handled by the platform.
Every call is transcribed and analysed
After each call the agent handled, the MAJ Console runs the transcript through the gap-detection engine. It looks for two things: questions the agent could not answer, and moments where the agent stated something confidently but the assertion was not grounded in the knowledge base — a subtler failure mode than a flat 'I don't know'.
The gap surfaces in your Console queue
Detected gaps appear in the Agent Optimization queue inside the MAJ Console. Each item shows the original caller question, the agent's response, and — where applicable — a flag indicating that the agent answered confidently but may have fabricated or extrapolated outside the verified knowledge base.
You confirm or correct the answer
You review the gap and either confirm the agent's phrasing, supply the correct answer, or mark the item as out of scope. The review takes seconds for routine gaps. You're approving the content your agent will say, not writing prompts — that distinction matters for audit trails.
A QA and compliance gate runs before it goes live
Before the confirmed answer is pushed to the live agent, it passes a QA gate: length and tone checks, TDRA-alignment screening (so nothing in the knowledge base inadvertently crosses telemarketing rules under Cabinet Resolutions 56 and 57 of 2024), and personal-data handling review under UAE Federal Decree-Law No. 45 of 2021 (the UAE PDPL). Only clean answers make it through.
The live agent is updated — calls improve
The agent learns the verified answer and applies it on the next matching call. The Agent Optimization score — the percentage of your call types the agent handles to completion without escalating — moves up. You watch a concrete number rise rather than guessing whether tuning is working.
The Agent Optimization Score
A number that tells you how much of your call volume the agent owns
The score is calculated from real call outcomes: the percentage of call types your agent handles to completion — answering the question, booking the appointment, or completing the intended action — without escalating to a human.
A new deployment typically starts with a score reflecting the completeness of the initial knowledge base. As gaps close through the review loop, the score climbs. The Console shows you the trajectory alongside the gap queue, so you can see exactly which review sessions moved the number and by how much.
The score is not an AI confidence rating. Confidence metrics are internal to the model and are not reliable as client-facing proof. The Agent Optimization score is computed from what actually happened on the call — whether the agent resolved it or did not.
Gap types caught
Unanswered + confident-but-unverified
Human review required
Every knowledge-base change, no exceptions
Compliance gate layers
TDRA alignment + UAE PDPL data check
Score basis
Real call outcomes, not AI confidence
Design decisions
Why we built it this way
Agents that know what they do not know
Most AI agents fail silently — they give a confident wrong answer and the caller hangs up unconvinced. Gap detection catches confident-but-unverified responses, not just outright failures. That distinction is what separates an agent that sounds good from one that is actually reliable.
A score you can show your stakeholders
The Agent Optimization score is a single number: the share of your call types the agent resolves without escalation. It is calculated from real call outcomes, not from internal AI confidence metrics. That makes it meaningful to clients who need to justify the deployment to management.
No unchecked learning — every update is human-approved
The agent does not update itself from caller input. Every knowledge-base change passes through your review and the compliance gate. That is a deliberate constraint, not a limitation — it is what keeps the agent accurate and keeps you in control of what it says on your behalf.
Regulatory context
The compliance gate is not optional in the UAE
UAE telemarketing activity is governed by two instruments that came into force in 2024: Cabinet Resolution No. 56 of 2024, which sets the substantive telemarketing rules (who may call, when, with what consent, and with what record-keeping obligations), and Cabinet Resolution No. 57 of 2024, which establishes the violations and administrative-penalties framework. Both are published by the Telecommunications and Digital Government Regulatory Authority (TDRA). Official consolidated texts are available via the UAE Legislation Portal.
The Agent Optimization compliance gate checks confirmed knowledge-base entries against both resolutions before they are pushed to the live agent. Separately, any entry involving a caller's personal data — name, phone number, medical detail, financial information — is checked for handling compliance under Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data (UAE PDPL).
These are not tick-boxes. An agent that states incorrect policy information, makes undisclosed recordings, or mishandles a caller's personal data creates liability for the business running it, not just for MAJ Leads. The compliance gate is where that liability is managed before it reaches a live call.
Common questions
Questions about Agent Optimization
How long does a knowledge-base review typically take?
Most gap items take 30 to 90 seconds to review. The Console presents the caller question, the agent response, and — where flagged — the reason the answer is unverified. You confirm, correct, or mark it out of scope. Batches of five or six gaps from a busy day can usually be cleared in under ten minutes.
Can the agent update itself from caller input without my review?
No. Caller-supplied information never goes directly into the knowledge base. Every entry that reaches the live agent has passed through your review and the compliance gate. This is by design: an agent that learns from callers without a human checkpoint is an agent that can be corrected — or manipulated — by callers. We chose not to build that.
What does "confident-but-unverified" mean in practice?
It means the agent gave an answer that wasn't in the approved knowledge base — not a refusal or an escalation, but a statement that looks credible and probably came from the underlying language model's training data. Common examples: a price the client never specified, a policy the agent inferred from a similar question, an opening-hours assumption. These are harder to catch than failed responses, which is why the engine specifically looks for them.
Is Agent Optimization available on all MAJ Leads plans?
Agent Optimization is part of the standard MAJ Console. The gap queue, review workflow, compliance gate, and score are available to all active deployments. Contact us if you have questions about which plan tier applies to your deployment.
How does this interact with TDRA compliance on outbound calls?
Outbound campaigns already run inside the TDRA window (09:00–18:00) against pre-screened lists, aligned with Cabinet Resolutions 56 and 57 of 2024. The Agent Optimization compliance gate adds a second layer: any knowledge-base entry that would affect outbound call content is checked against the same rules before it goes live. The two systems are complementary.
Book a demo
See Agent Optimization in the Console
We will walk you through a live deployment, show the gap queue, run a review, and you will see the score before the call ends.
Dubai, UAE · +971 58 950 9633 · hello@majleads.com