The customer

A Thailand-based enterprise cybersecurity consulting and solutions provider. The company aligns its practice to the NIST Cybersecurity Framework and ISO/IEC 27001, has delivered more than 500 cybersecurity projects, and partners with global security vendors including Palo Alto Networks, CrowdStrike, Fortinet, Check Point, CyberArk, Splunk, Zscaler, SentinelOne, Imperva, Trellix and Devo. Roughly 90% of its staff are technical practitioners.

AWS Partner: G-Able Public Company Limited

The challenge

The customer’s presale function had become the constraint on revenue.

Producing a proposal for an AWS-based cybersecurity engagement required a senior engineer to work manually through requirement analysis, inventory mapping, man-day effort sizing, pricing, proposal writing and solution design. That sequence consumed roughly five architect man-days per proposal, and the elapsed turnaround from brief to submitted proposal ran to 10.5 days.

Three problems followed from that.

Speed capped pipeline coverage. The number of opportunities the team could pursue in a quarter was limited by how fast proposals could be produced, not by how many opportunities existed.

Sizing and pricing varied between engineers. Standards and rate assumptions lived in individual heads rather than in a governed system, so two engineers could size materially similar scopes differently.

Manual pricing carried real financial risk. AWS list prices change. A figure assembled by hand against a stale price is either a loss absorbed at delivery or an uncomfortable conversation with the customer.

Left unaddressed, the immediate exposure was lost opportunities to faster competitors and margin erosion on underpriced deals. The longer-term exposure was an inability to scale presale in step with demand — the same senior architects who write proposals are the ones who deliver billable work, so presale volume and delivery capacity compete directly for one scarce resource.

Why this needed an agentic solution

Proposal production is not a single generative task. It is a sequence of dependent decisions: understand the requirement, select an architecture, size it, price it against live rates, write it, then route it for approval. Each step needs different knowledge and different tools, and a mistake early propagates silently to the end.

A single large-language-model prompt cannot do this reliably, because the model has no way to look up a current AWS price and no structural reason to refuse to invent one. What the work needs is an agent that plans, calls real tools against real data, and is prevented by design from producing a number it

The solution

G-Able designed and delivered a multi-agent presale automation platform into the customer’s own AWS account in the Asia Pacific (Singapore) Region.

At a high level

An Orchestrator plans each piece of work and routes it to specialist sub-agents — pricing, solution architecture, proposal writing, design documentation, approval guidance and gated repository creation. The agents call real tools against live data rather than generating text from model knowledge alone. Every proposal that reaches a customer has passed a human approval gate, and every figure inside it traces back to a live pricing lookup.

The technical detail

Agent runtime and orchestration. All agents run on Amazon Bedrock AgentCore Runtime, so agent sessions are managed by AWS rather than by self-managed containers. The Orchestrator follows a hub model: sub-agents cannot call each other freely. The single sanctioned lateral edge runs from the solution-architecture agent to the pricing agent, which guarantees that no architecture recommendation reaches a human without a verified price attached.

Inference. Amazon Nova Pro through Amazon Bedrock provides all inference, with Cohere Embed 3 Multilingual generating retrieval embeddings over a standards and pricing corpus held in Amazon Bedrock Knowledge Bases. All inference stays inside AWS-managed services in-Region. Agents are built on the Strands agent framework.

Tool mediation. Amazon Bedrock AgentCore Gateway exposes knowledge bases and external tools over the Model Context Protocol, keeping tool execution isolated from orchestration. AgentCore Identity gives each agent its own identity rather than a shared credential, and AgentCore Memory carries durable customer constraints across sessions, so a constraint stated once is not re-learned each time.

Grounding, and why no figure can be invented. The pricing agent queries the AWS Price List API for unit prices and AWS Cost Explorer for actual consumption. A deterministic grounding validator then rejects any quantitative claim in a generated document that cannot be traced to one of those live lookups. The writer agents hold zero tool access by architectural design — they can restate already-validated facts and nothing else, so they are structurally incapable of originating a number. When the pricing source is unavailable the platform produces no figure rather than an unverified one.

Durable state. Every state transition is written to an append-only statusHistory ledger in Amazon DynamoDB, which the agents, the user interface and the audit trail all read from. An interrupted workflow resumes from its last recorded transition rather than being reconstructed from agent memory, and an operational investigation and a business report resolve against identical data.

Platform and network. The human-facing web application, the API tier and the EKS-native agents run on Amazon EKS in private subnets across three Availability Zones, behind an Application Load Balancer protected by AWS WAF, with Amazon Cognito authenticating users and the resulting token validated at the AgentCore Gateway on every request. AWS Lambda runs Gateway tool executions and the event-driven dispatch path. Amazon S3 holds the standards corpus, Amazon ElastiCache serves the interactive tier, and Amazon SNS with AWS Step Functions coordinates asynchronous approval notification. Encryption at rest uses AWS KMS customer managed keys scoped per data domain; secrets are held in AWS Secrets Manager and retrieved at runtime rather than baked into images. Amazon CloudWatch and Amazon Managed Service for Prometheus provide metrics, logs and alarms.

Delivery. The environment is provisioned by Terraform through a gated CI/CD pipeline with a manual approval stage. No infrastructure change reaches production through the console.

Governance: AI proposes, humans decide

Three controls carry the governance model, and none of them depends on the model behaving well:

1. Mandatory human approval gates on proposal release and on any infrastructure change. A human engineer reviews the generated design document before anything is applied.
2. Grounding enforcement — an unsupported figure is rejected by deterministic code, not by prompt instruction.
3. Least authority by construction — writer agents cannot reach tools at all, so widening their own scope is not available to them.

Retrieved external content is wrapped as data rather than as instructions, so untrusted content in the corpus or in a customer input cannot alter agent behaviour.

Results

Measured across the pilot and confirmed by the customer’s presale team:

At roughly 10 proposals per month, that is approximately four architect man-days released from every proposal produced. The platform runs entirely on consumption-priced and On-Demand AWS services with no upfront cost and no commitment, and its monthly running cost is recovered within the first proposal produced each month.

Beyond the numbers, two qualitative outcomes matter as much:

• Pricing consistency. Sizing and pricing now derive from one governed set of assumptions rather than from individual recollection, so materially similar scopes are priced consistently.
• Capacity, not just speed. Removing roughly four architect man-days from every proposal returns that time to billable delivery — which was the actual business constraint, and the reason the engagement was commissioned.

About G-Able

G-Able Public Company Limited is a leading Thai technology enabler with over 37 years of experience, working with more than 100 global technology partners and serving over 1,000 active customers. G-Able’s AI practice spans cybersecurity, cloud and data centre modernization, data and analytics, digital business and application, and managed technology services, delivered by dedicated AWS-focused solution architecture, FinOps and delivery teams.

Interested in what agentic AI could do for your presale or delivery function? [Contact G-Able].