Frequently Asked Questions

"How much does it cost?"

Three tiers designed for different scales:

• L1 Foundry: $250/mo — 1 agent, POC/validation
• L2 DLPFC Runtime: $2,500/mo — 5 agents, team scale
• L3 Forge Enterprise: $25,000+/mo — unlimited, on-prem option

Commissioning fee: $50k-250k+ for industrial deployments (custom knowledge graph setup, legacy doc ingestion, 90-day shadow mode).

"Why so expensive vs ChatGPT?"

ChatGPT is a goldfish. It forgets everything after 8k tokens.

AMS is a cognitive operating system. It:

• Remembers forever (H-MEM architecture)
• Learns from successes (Bayesian skill validation)
• Never hallucinates unsupported claims (7-layer security)
• Provides full audit trail (Graph View)

"What's the ROI?"

Typical enterprise deployment:

• 2 junior dev salaries saved (~$150k/year)
• 4+ hours/week per knowledge worker recovered
• Reduced onboarding time (new hires productive 3x faster)
• Zero "graybeard leaves" knowledge loss

Payback period: 3-6 months for L3 deployments.

"How is this different from RAG?"

Standard RAG = retrieval + generation. You stuff context into a prompt.

AMS adds:

• Procedural Memory — Learns executable skills from patterns
• Bayesian Validation — Only promotes skills that prove reliable
• Graph Persistence — Context isn't in the prompt, it's in the database
• Security Pipeline — 7 layers before anything touches the LLM

"What LLM do you use?"

Flexible. Currently:

• Default: Gemini 1.5 Flash (cost-effective)
• Reasoning: Kimi K2 Thinking (complex tasks)
• Local: LM Studio / Ollama (air-gapped deployments)

We're model-agnostic. The memory layer is the product.

"Can this run on-premise?"

Yes. Three deployment options:

• Managed Cloud — We host everything
• BYOC — Our software in your VPC
• Air-Gapped — Full on-prem, data diodes, your H100s

Defense/finance clients typically go BYOC or air-gapped.

"How do you handle PII?"

7-Layer Security Pipeline includes:

• Input sanitization
• PII detection & redaction
• RBAC enforcement
• Audit logging

"What about hallucinations?"

Output validation layer cross-references claims against retrieved chunks. Unsupported claims are blocked before response.

Plus: Every answer includes source citations. Users can verify.

"What industries is this for?"

Primary targets:

• Industrial/Energy — HVAC, oil & gas, utilities (Boilerman proves this)
• Defense/Aerospace — Communications-denied, air-gapped requirements
• Finance/Insurance — Audit trail, compliance, explainability

Common thread: High-stakes environments where hallucinations = liability.

"Can you give me a case study?"

Boilerman (HVAC vertical):

• 61,000+ document chunks indexed
• 290+ manufacturer coverage
• Query response: <10 seconds (vs 15-30 min manual)
• 72% token reduction vs vanilla GPT-4
• 89% retrieval precision (vs 71% standard vector search)

"How long does implementation take?"

Depends on scope:

• POC (L1): 1-2 weeks — Connect docs, basic memory, single agent
• Team Scale (L2): 4-6 weeks — Multi-agent, skill learning, integrations
• Enterprise (L3): 3-6 months — Full commissioning, legacy systems, custom security

"We already use [Competitor]"

Let's compare:

• ChatGPT/Claude direct → No memory, no audit trail, hallucination risk
• Glean/Guru → Search, not cognition. Can't learn or execute.
• Custom RAG → You're maintaining it. How's that going?

AMS is the only system with Bayesian skill validation and procedural memory.

"This sounds too good to be true"

Fair. Let's do a POC. $25k, 3 months, specific use case. You'll see the Graph View. You'll watch it learn. Then decide.

"Our IT won't approve external AI"

That's why we offer BYOC and air-gapped. Your data never leaves your walls. We just license the software.