Glean

Artificial intelligence software covers the platforms, models, and infrastructure used to build AI-powered applications. The category sits below the AI agent layer and above the raw cloud GPU layer — it is the toolkit that turns foundation models into production capabilities. The category spans foundation model providers, fine-tuning and training platforms, model gateways and routers, vector databases, evaluation and observability tools, and the MLOps systems that operationalize models in production. The unifying theme: software that helps teams ship AI features without rebuilding the infrastructure each time.

Four forces push organizations toward dedicated AI software: • Model choice matters. No single foundation model is best at everything. Routing requests across providers — cost-optimized for one task, quality-optimized for another — requires platform support. • Speed beats sophistication. The team that ships an AI feature in weeks beats the one that ships a perfect feature in months. AI platforms collapse the integration work. • Cost discipline is real. AI inference costs scale with usage. Without observability, caching, and routing, costs grow unpredictably. AI platforms turn that variable into something measurable. • Governance is no longer optional. Auditing what model was used, what prompt was sent, what response came back, and who saw it has become a baseline requirement for regulated industries — and a wise practice for everyone.

The capabilities that define modern AI platforms group into seven areas: Model access and routing • Access to multiple foundation models through one API • Dynamic routing by cost, latency, or quality • Provider failover and load balancing • Model versioning and pinning Fine-tuning and customization • LoRA and adapter-based fine-tuning • Full fine-tuning with managed compute • Distillation from larger to smaller models • Continuous training pipelines Prompt and context management • Prompt versioning and A/B testing • Template libraries with variable substitution • Context window optimization and chunking • Retrieval-augmented generation (RAG) tooling Vector and retrieval • Embedding generation and storage • Hybrid search (vector + keyword) • Metadata filtering and namespace isolation • Index management at scale Evaluation and observability • LLM-as-judge and human evaluation workflows • Regression detection across model versions • Trace logging of every inference • Token and cost tracking Safety and guardrails • Content filtering on inputs and outputs • PII detection and redaction • Jailbreak and prompt injection defense • Policy enforcement engines MLOps and deployment • Model registry and lineage • Deployment to managed and self-hosted infrastructure • A/B traffic splitting and shadow deployments • Drift monitoring on input and output distributions

Teams that invest in AI platforms report three durable benefits: • Faster feature delivery. What used to require custom inference infrastructure, evaluation scripts, and monitoring code is collapsed into platform features. • Cost predictability. Routing, caching, and per-feature budgets turn AI spend from an unbounded variable into a managed line item. • Safer rollout. Evaluation, guardrails, and audit logs let teams ship AI features with the same confidence as any other production system.

• AI engineers and ML practitioners — building and fine-tuning models • Application developers — shipping AI-powered features into existing products • Data science teams — running experiments and putting models in production • Platform engineers — operating the inference and serving infrastructure • Security and compliance teams — auditing AI usage and enforcing policy • Product managers — measuring AI feature adoption and quality

The market is fragmented and the pace of change is relentless. Evaluate against these criteria: 1. Multi-provider support A platform that locks you into one foundation model provider is a strategic risk. Look for true multi-provider abstraction — switching models should not require rewriting application code. 2. Inference economics Hosted, self-hosted, and routed inference have very different cost curves. Confirm pricing transparency, the ability to bring your own provider keys, and caching strategies. 3. Evaluation infrastructure Without systematic evaluation, model upgrades become Russian roulette. Confirm the platform supports your evaluation methodology — golden datasets, LLM-as-judge, human review — and integrates with CI. 4. Latency and throughput First-token latency, total-token latency, and concurrency limits vary dramatically across providers and configurations. Test against your real workload, not synthetic benchmarks. 5. Data residency and isolation Where models are hosted, where prompts are logged, and where embeddings are stored matters for regulated data. Confirm the platform supports your residency requirements. 6. Open vs proprietary models Some platforms only serve proprietary models; others support open-weight models you can self-host. The choice affects cost ceiling, customization depth, and exit cost. 7. Ecosystem integrations The AI stack rarely lives in isolation — it integrates with feature stores, data warehouses, observability tools, and application frameworks. Confirm the integrations your team already depends on.

• Start with one foundation model and one use case. The temptation to evaluate every model on every workload is paralyzing. Pick a default, ship something, then expand. • Build evaluation before scale. A test set with expected outputs catches regressions long before production users do. The cost of building it is repaid many times over. • Cache aggressively. Many AI workloads have repeated inputs. Semantic caching can cut inference cost by 30–60% with minimal quality loss. • Plan for the model upgrade treadmill. New models ship monthly. Your platform should make upgrades a configuration change, not a rewrite. • Instrument from day one. Trace every inference, log every prompt, attribute every dollar. Retrofitting observability is harder than building it in.

AI software pricing typically combines: • Per-token / per-request inference fees — pass-through of model costs with platform margin • Platform subscription — base fee for the AI platform itself • Compute hours — for self-hosted or fine-tuning workloads • Storage — for vector indexes, fine-tuned models, and logs Hidden costs surface in egress fees, fine-tuning storage, and the cost of evaluation runs at scale.

• Specialized models. The trend from one large generalist to many smaller specialists is accelerating. Routing to the right specialist beats forcing one model to do everything. • Edge and on-device inference. Mid-sized models running on edge hardware or end-user devices reduce latency and cost for many workloads. • AI gateways. A dedicated middleware layer that handles routing, caching, observability, and policy across providers is emerging as a standard architecture pattern. • Open-weight competitiveness. Open models have closed much of the quality gap with frontier proprietary models, making self-hosting viable for far more use cases. • Evaluation as a discipline. Systematic evaluation, golden datasets, and continuous benchmarking are moving from research practice to engineering hygiene.

What is the difference between AI software and AI agents? AI software is the platform layer — models, infrastructure, tooling. AI agents are applications built on top of that layer to execute autonomous tasks. Most agent platforms use AI software underneath. What is a foundation model? A foundation model is a large model trained on broad data that can be adapted to many downstream tasks. Large language models are the best-known example; the category also includes image, audio, and multimodal foundation models. Should I use a proprietary or open-source model? Proprietary models often lead on capability and ease of use. Open models offer lower per-inference cost at scale, full data control, and no vendor lock-in. Many production systems use both — proprietary for hard tasks, open for high-volume routine ones. What is retrieval-augmented generation (RAG)? RAG combines a retrieval step (finding relevant documents) with a generation step (producing an answer grounded in those documents). It is the dominant pattern for grounding LLMs in private or up-to-date data. What is an AI gateway? An AI gateway is middleware between an application and one or more foundation model providers. It handles routing, authentication, caching, observability, rate limiting, and policy enforcement — much like a traditional API gateway, but tuned for LLM workloads. How much does AI infrastructure cost? Costs vary by model size, request volume, context length, and inference pattern. A useful planning rule: assume the inference cost of a feature scales linearly with adoption, and instrument accordingly. What is model drift? Model drift is the gradual degradation of model output quality over time as the world changes around it. For deployed models, drift comes from changing user behavior, vocabulary, or upstream data. Monitoring input and output distributions catches drift early. ---