OPTI Software

AI cost optimization in 2026 is not only choosing between Gemini, Claude and GPT. The following matter as well:
❓ Which requests actually need the strongest model?
❓Can repeated context be cached?
❓Does every request need grounding?
❓How many retrieval steps are triggered?
❓What happens after a failed tool call?

A cheaper model used badly can be more expensive than a premium model used selectively. And an optimized token bill can still hide an inefficient application.

So the KPI I would want for most production AI projects is increasingly:
cost per verified business outcome, not cost per token.

That is the perspective behind our comparison of AI costs in Google Cloud, AWS and Azure in the first half of 2026.

👉 Read it here https://www.opti-software.com/post/ai-costs-2026-google-cloud-aws-azure/

👉️ Romanian version here https://www.opti.ro/post/costuri-ai-2026-google-cloud-aws-azure/

What does a practical cyber resilience programme actually look like?

We focus on the real organizational activities beyond documentary audits and cyber procurement decisions. We publish a six-phase model that connects cybersecurity, business continuity and tested recovery:

💡Phase 0: Stabilisation
Define incident roles, emergency access, backup verification and an out-of-band communication channel.

💡Phase 1: Inventory & Business Impact Analysis
Know your assets, dependencies, RTO, RPO, temporary workarounds and recovery order.

💡Phase 2: Attack surface reduction
Privileged access, MFA, patching, EDR, encryption, segmentation and vulnerability management.

💡Phase 3: Tested recovery
Restore a known-clean copy into an isolated environment, rebuild services and integrations, then validate the actual business process.

💡Phase 4: Incident response
Playbooks, staff training and tabletop exercises involving management, IT, and legal.

💡Phase 5: Continuous operation
Recurring controls, access reviews, restore tests and resilience exercises.

The goal after 12 weeks is to get evidence of reselience:
“We restored the service. The business validated it. RTO was 4h20. These are the issues we found. This is when we test it again.”

Our new article explains the complete framework and how it maps to NIST CSF 2.0, CIS Controls, ISO 27001, ISO 22301, ISO/IEC 27031, GDPR and NIS2.

👉 Read the framework here https://www.opti-software.com/post/cyber-resilience-program-12-weeks/

👉️ Read the romanian version here https://www.opti.ro/post/program-rezilienta-cibernetica-12-saptamani/

De ce AI nu găsește întotdeauna răspunsul potrivit?

Cum transformi o bază de cunoștințe într-un sistem care oferă răspunsuri relevante.

În ediția precedentă am organizat cunoașterea.

Dar simpla existență a informației nu este suficientă: AI trebuie să găsească exact sursa potrivită atunci când are nevoie de ea.

Într-un catalog cu milioane de produse sau într-o bază mare de documente, căutarea trebuie să înțeleagă atât termenii exacți, cât și intenția utilizatorului.

Citește ultimul newsletter săptămânal la 👉️ https://buff.ly/YD4fEti

What is an AI Agent allowed to act on?
An enterprise AI agent needs a limited scope, clear tools, verified data, permissions, logging and security boundaries. That is where agentic architecture starts. You can start with three layers:

Business goal
What task should the agent actually help with?

Tools and data access
Which CRM, ERP, documents, APIs or databases can it use?

Guardrails and audit
What must be blocked, logged, verified or escalated to a human?

This can be illustrated with cloud projects, where Gemini, BigQuery, and security services can be combined into production-ready systems.
The future of agentic AI in business is controlled delegation.

We wrote more about this in our guide on Agentic AI in Google Cloud.

👉 Read the article here: https://www.opti-software.com/post/three-differentiators-ai-google-cloud-2026/

👉️ Romanian version: https://www.opti.ro/post/trei-diferentiatori-ai-google-cloud-2026/

How do you build a cyber resilience program in 12 weeks?

You don’t start by buying a SIEM. Nor with a strict documentary audit. Nor by assuming that “we have backup” means that the activity can be resumed.

A practical program starts by answering three questions:

1. What is critical?
Processes, systems, people, suppliers and dependencies.

2. How long do we need to recover?
Define the factors RTO, RPO, workaround and restore order.

3. Can we demonstrate that it works?
Through real restore, business validation and incident response exercises.

The methodology we describe has six phases:
0. Stabilization and Accountability
1. Inventory and Business Impact Analysis
2. Attack Surface Reduction
3. Tested Restore and Continuity
4. Incident Response and Training
5. Recurring Operation

The framework uses principles and requirements from NIST CSF 2.0, CIS Controls v8.1, ISO 27001, ISO 22301, ISO/IEC 27031, GDPR, and NIS2.

Ultimately, the organization should not only have policies, but evidence that it has a measurable capacity to recover.

👉 Full article: https://www.opti-software.com/post/cyber-resilience-program-12-weeks/

👉Romanian version: https://www.opti.ro/post/program-rezilienta-cibernetica-12-saptamani/

AI is getting cheaper. AI applications are not necessarily getting cheaper.
That is the main tension we looked at in our latest analysis of AI costs across Google Cloud, AWS and Azure in H1 2026.

The price of inference can be very low for compact models. Caching, batching and routing can reduce it further, so you should activate them.

But a production AI application can also pay for:
💲search and retrieval
💲index and storage
💲grounding
💲tool calls
💲guardrails
💲observability and logging
💲retries and errors
💲application infrastructure

In some architectures, the LLM is no longer the largest line on the bill.
We compared public pricing for Gemini, Claude and GPT models, but also Google AI Commerce Search, Amazon Personalize, RAG / agentic retrieval and broader AI infrastructure spending. With a special chapter on predictive AI: recommendation engines, OCR, forecasts.

That's why the useful question for clients right now is “How much does one correctly completed task cost?”, not "How much does 1M tokens cost"

👉 Full analysis: https://www.opti-software.com/post/ai-costs-2026-google-cloud-aws-azure/

👉️ Romanian version: https://www.opti.ro/post/costuri-ai-2026-google-cloud-aws-azure/

❓Do you cost AI in tokens or in price per deliverable?

Cât costă, de fapt, o aplicație AI în producție?
În noul newsletter OPTI vorbim despre costurile reale ale AI, agenți B2B, 24 de aplicații în medical & pharma și un nou studiu de caz ERP.
Citește ediția completă 👇

Cât costă de fapt o aplicație AI? + 24 de aplicații în medical AI-ul intră în etapa în care modelul e doar o parte a problemei. În ediția aceasta ne uităm la costurile reale ale aplicațiilor AI în producție, la agenți B2B și la 24 de utilizări concrete în medical & farma.

Backup ≠ Continuitate. Ai backup. Dar ai restaurat vreodată compania din el?

Incidentul Agenției de Cadastru din iulie, care este încă în curs de rezolvare completă, arată diferența dintre integritatea datelor și disponibilitatea serviciilor.

Organizația are continuitate atunci când știe minim:
– Ce aplicații depind unele de altele
– În ce ordine trebuie pornite sistemele
– Ce identități și chei trebuie reconstruite în caz de atac
– Cine ia deciziile în primele 60 de minute
– Dacă restaurarea fiecărei componente operaționale (nu doar IT!) durează 4 ore, 24 de ore sau patru săptămâni ca în cazul de față

Nu este adevărată iluzia „Stăm bine, avem backup.”, nici iluzia: avem servicii în cloud, se ocupă altcineva.

Backup-ul este o componentă, cloud-ul este o componentă, iar continuitatea trebuie demonstrată.

Am structurat într-o implementare concretă un program practic de 12 săptămâni în șase faze pe care îl împărtășim pe blogul OPTI Software:
Stabilizare > Inventar & BIA > Hardening > Restore testat > Incident Response > Operare recurentă.

Pentru cei pasionați de standarde și reglementări, fiecare fază este ilustrată cu specificații din NIST, CIS Controls, ISO 22301, ISO/IEC 27031, GDPR și NIS2.

Întrebarea cu care aș începe însă într-o organizație publică ar fi:
Dacă atacul are loc mâine la 08:00, știm exact ce facem până la 09:00?

👉️ Citește aici - https://www.opti.ro/post/program-rezilienta-cibernetica-12-saptamani/

👉️ English version - https://www.opti-software.com/post/cyber-resilience-program-12-weeks/

P.S. Pentru o instituție de dimensiunea ANCPI ar fi insuficiente 12 săptămâni pentru un program complet de reziliență, dar un baseline de reacție rapidă ar putea fi livrat.

In B2B sales, the “next best action” doesn't always exist. Sometimes the best action is: do nothing yet. How can an AI-based system achieve that?

As strange as it sounds in a world where every system wants to push more alerts, more tasks and more recommendations, *noise is expensive*.

In B2B Sales, noisy systems are another dashboard, not a real assistant. We care about practical actions when applicable:
• call the customer
• send a follow-up
• propose a complementary product
• escalate to a key account manager
• block upsell because of overdue invoices
• or simply wait, because the commercial signal is too weak

Such AI-based system needs ERP and CRM data, timing signals, stock and margin checks, business rules, and feedback from the sales team.

The goal is to reduce irrelevant actions and surface the few that actually matter.

👉 We wrote more here: https://www.opti-software.com/post/step-by-step-seller-next-best-action-b2b/

👉️ Romanian version here - https://www.opti.ro/post/vanzator-pas-cu-pas-next-best-action-b2b/

Agentic AI is often described as “AI that acts”. But in real companies, you need to know what is it allowed to act on?

Enterprise AI agents are not only very smart chatbots with potential access to everything, architecture should define the scope and security boundaries for each of them.

A useful agentic AI system usually has three layers:

🌟Business goal
What task should the agent actually help with?

🌟Tools and data access
Which CRM, ERP, documents, APIs or databases can it use?

🌟Guardrails and audit
What must be blocked, logged, verified or escalated to a human?

Google Cloud projects are revelatory, as enterprises need to adapt to new tools and patterns for production-ready systems which enforce controlled delegation.

We wrote more about this in our guide on Agentic AI in Google Cloud.

👉️ Read the article here: https://www.opti-software.com/post/three-differentiators-ai-google-cloud-2026/

👉️ Read the romanian version here: https://www.opti.ro/post/trei-diferentiatori-ai-google-cloud-2026/