AI industry observation: what changed after a year of LLMs in enterprises

The short version: the biggest change of the past year is not that the models got smarter. It is that enterprises finally evaluate AI by whether it embeds into a process rather than by how impressive the demo is. We keep watching the same turning point at client sites: everyone applauds the demo, then it stalls the moment it enters a real workflow. The cause is almost never the model — it is data, permissions and process, three unglamorous layers.

Set the frame first. An enterprise AI capability is not a standalone product; it is a set of functions embedded in business systems: it reads the company's knowledge base and business data, works under the company's rules, and produces output that a human reviews, that is traceable, and whose errors can be located. There is exactly one success metric: this step used to take 30 minutes, now it takes 5, and accuracy did not drop.

1 · What changed: from ten pilots to the ones that pay for themselves

The 2025 pattern was to launch ten AI pilots at once and pick the best screenshots for the year-end report. In 2026 the practice is converging: define an evaluation baseline first, then talk about go-live. The pilot funnel is brutal, and that is healthy — the pilots that die fail on three things: no evaluation baseline (quality judged by feel), a cost case that does not close (tokens and ops outweigh the saved hours), and an unchanged process (AI as a skin over the old workflow).

Out of ten pilots, the number that end up embedded in a process and still running a year later is often less than one. What filters them out is not model capability — it is ownership, evaluation and process.

2 · Four layers — and the one most projects are missing

Enterprise AI stacks in four layers: a data layer (knowledge base governance: document structuring, de-identification, versioning, who may see what), a model layer (cloud API, self-hosted open weights, or retrieval-augmented generation), an application layer (embedding into CRM/OA/MES, workflow orchestration, and evaluation, permissions and audit), and a scenario layer that answers only for hours saved and errors avoided.

The layer most failing projects skip is evaluation, permissions and audit. Models are bought as a service and data layers can be outsourced, but three questions must be designed into the system: what happens when the answer is wrong, who is allowed to see which data, and which document does each answer cite.

  • Evaluation. Before go-live, build a regression set of 50–200 real business questions. Re-run it every time you change the model, the prompt or the knowledge base. Without it, every change silently breaks something.
  • Permissions. What the AI may answer depends on who is asking. Knowledge-base permissions must follow the org structure in the business system — not one shared bucket for everyone.
  • Audit. Every answer carries citations (which document, which passage) and important actions leave logs. When something breaks you can tell whether the data was wrong or the interpretation was wrong — that is what makes the system improvable.
In one sentence: the model makes it sound right; evaluation and audit make it right and compliant. What enterprise buyers actually pay for is the second half.

3 · Three integration paths and how to choose

PathFits whomStrengthCost & constraint
Cloud APIPublic data, general writing, coding assistanceStrongest capability, zero hardware, rides model upgradesData must be de-identified before it leaves; usage-based billing needs a long-run cost model
Self-hosted open weightsDrawings, process parameters, finance and HR dataData never leaves the domain; auditable and controllableGPU capex plus dedicated ops; weaker than frontier models, so RAG and process design carry complex tasks
Hybrid (the usual end state)Most mid-size enterprisesSensitive steps local, general steps in the cloud — a workable cost/compliance balanceTwo routing and evaluation tracks; highest engineering complexity

The practical rule: split scenarios by data sensitivity first, then pick the path. Whether data may leave the domain decides 80% of the architecture; the remaining 20% — which model, whether to fine-tune — is the easiest part to change later, provided the application layer treats model calls as a replaceable interface.

4 · Five scenarios where the math actually works

  • Enterprise knowledge Q&A. Policies, manuals and historical tickets become a searchable knowledge base; new hires stop asking the veteran how to file an invoice. Precondition: documents must be structured first — a pile of scans retrieves poorly no matter the model.
  • First-draft documents. Tenders, ticket replies and acceptance reports generated from templates, humans revise. The zero-to-draft stretch is usually over 60% of the effort, and that is what gets saved.
  • Data Q&A (ChatBI). Business users query data in natural language instead of filing a report request and waiting. Precondition: metric definitions agreed first, or the same number will get two answers.
  • Inspection and visual judgement. Surface defects and document checks are small-model territory: enough labelled samples, running at the line edge, at a fraction of frontier-model cost.
  • Engineering productivity. Code completion, unit-test generation and legacy-code walkthrough. The most certain ROI today — and the most dependent on engineering discipline: a messy codebase feeds no good assistant.

5 · Agents: what they can do, what they cannot yet

  1. Already dependable: single-step tasks with verifiable output. "Read this batch of contracts and fill the key terms into a structured table" — every field can be checked, errors are visible.
  2. Conditionally usable: multi-step flows with a human confirmation point between steps. "Draft the delivery note from the order and submit for approval" — approval is a natural checkpoint while the agent assembles the material.
  3. Not yet autonomous: open-ended decisions in an unbounded environment. Letting an agent decide which customer gets which discount fails not on capability but on accountability: if the system cannot say who is responsible for an error, it should not be allowed to act alone.
  4. The rule of thumb: the more a task is verifiable, reversible and checkpointed, the safer it is to agentify. For everything else, build "assist the human" before "replace the human".

6 · Five recommendations before you invest

  1. Choose scenarios by choosing an owner. Every AI project needs a named business-side owner who benefits and signs acceptance. Pilots without owners end at the demo.
  2. Build the evaluation set before the feature. Fifty real questions, reference answers, scoring rules — one week of work and the most valuable asset of the whole project.
  3. Budget data governance as half the work. Scans to structured documents, permission mapping, de-identification rules. Unsexy, decisive, and routinely underfunded.
  4. Embed into existing systems, do not open a new webpage. Support uses it inside the ticketing system; sales inside the CRM. Every extra entry point halves adoption.
  5. Review quarterly; kill what stalls. If the evaluation numbers do not move for two quarters, stop the pilot. Ten mediocre pilots are worth less than one scenario genuinely embedded in a process.

The end state we observe is not spectacular: the winners are not the companies using AI most aggressively, but the ones that embed it into one or two core processes and watch quality with an evaluation set. Models improve every year. What does not change is data governance, evaluation baselines and process design — which is, precisely, classical software engineering.

If you are evaluating enterprise AI adoption, bring two things and the conversation gets short: a description of the process you want to improve, and the current state of your systems' data and permissions. We will first say whether the bottleneck is data, evaluation or process, then propose scope and a phased quote.

Planning a similar production line or enterprise system?

Start with our pricing logic, then let's talk requirements — cost structure, technology comparison and acceptance criteria, all laid out in the open.

0.406340s