What companies are actually paying for in AI.
Everyone can write the code now. The money goes to the work around it: choosing the job, redesigning the workflow, training and shrinking the model, serving it cheaply, running it on the device, and keeping it honest. Ten categories, with the evidence.
In 2026 companies pay for agentic systems wired into real workflows, workflow redesign and operation rather than software, custom and distilled models, inference cost optimisation, on-device AI, voice agents, labelled data and evaluation, knowledge systems, integration and industry-specific applications, and sovereign deployment. Zingaro AI offers a service for each, and the figures below are the analysts', linked to their sources.
- 01 · What is paid for
Agentic AI development
What buyers ask for: Agents that finish a job in our systems, not a chatbot. Most of what is called an agent today is a fixed workflow around one model call.
What we do: Agents with tools, permissions, confidence thresholds and a review queue, built for one bounded job first and measured against a number written down before the build.
The evidence for Agentic AI development The evidence Source 40% of enterprise applications will carry task-specific AI agents by the end of 2026, up from under 5% in 2025. Gartner, August 2025 Agentic AI already represents 10–15% of enterprise IT spending in 2026. IDC FutureScape 2026 Only 16% of enterprise deployments qualify as true agents that plan, act and adapt. Menlo Ventures, December 2025 - 02 · What is paid for
Workflow redesign and automation
What buyers ask for: Take the whole process, not a step of it. Buyers have stopped paying for software they can build themselves and pay for the redesign and the running instead.
What we do: We take over the workflow: the queue, the exceptions, the systems it writes to, and the weekly report. The code is the smallest part of it.
The evidence for Workflow redesign and automation The evidence Source Nearly three-quarters of AI high performers fundamentally redesign workflows because of AI, against a quarter of everyone else; redesign has one of the strongest links to profit of any factor tested. McKinsey, State of AI 2026 Nearly a third of organisations decided against buying at least one software product because they can now build it with agentic coding tools. McKinsey, State of AI 2026 Over 40% of agentic AI projects will be cancelled by the end of 2027 for cost, unclear value or risk. Gartner, June 2025 - 03 · What is paid for
Custom LLM training, post-training and distillation
What buyers ask for: A model that does our job in our format at a cost that works at volume, and that we own.
What we do: Supervised fine-tuning, reinforcement post-training with verifiable rewards, distillation into small models, and the evaluation suite that gates all of it. The weights are yours.
The evidence for Custom LLM training, post-training and distillation The evidence Source Spending on generative AI models is forecast to grow 117% in 2026. Gartner, September 2026 The cost of GPT-3.5-level performance fell from $20 to $0.07 per million tokens between November 2022 and October 2024, a 280-fold drop: small models became good enough for narrow jobs. Stanford AI Index, as summarised Enterprise AI budgets are up about 65% in 2026, and the share reserved for pure innovation fell from 25% to 7%: the rest is core operating expense, where cost and reliability decide. a16z CIO survey, February 2026 - 04 · What is paid for
Model optimisation and inference cost
What buyers ask for: Our AI bill is the fastest-growing line we have and nobody forecast it. Make it smaller without making the product worse.
What we do: Quantisation, batching, speculative decoding, caching and routing, on your GPUs, measured with load tests on your real traffic, with the evaluation suite as the gate.
The evidence for Model optimisation and inference cost The evidence Source 2026 is the first year inference spending passed training in the AI cloud: $23.3B against $19B. Gartner, reported August 2026 73% of organisations say their AI costs exceeded the original budget; 98% now manage AI spend, up from 63% a year earlier. FinOps Foundation, State of FinOps 2026 - 05 · What is paid for
On-device and edge AI
What buyers ask for: Run it on the phone, the laptop or the camera. The hardware shipped; the models and the engineering did not come with it.
What we do: Small language, speech and vision models distilled, quantised and shipped inside your app or device, with a hybrid path to a larger model for the hard cases.
The evidence for On-device and edge AI The evidence Source End-user spending on GenAI smartphones is forecast at $393.3B in 2026, up 32%. Gartner, as reported AI PCs will exceed 50% of new PC sales in 2026, and 40% of software vendors will have AI built for PC deployment by the end of 2026, up from 2% in 2024. Gartner, as reported - 06 · What is paid for
Voice agents and custom speech models
What buyers ask for: Answer the phone in our customers' language and dialect, finish the task, hand off when it should.
What we do: Our specialism: custom ASR and TTS on your recordings, voice agents on our own telephony, low-latency streaming, deployment and optimisation of speech models.
The evidence for Voice agents and custom speech models The evidence Source Conversational AI will reduce contact-centre agent labour costs by $80 billion in 2026, with one in ten agent interactions automated. Gartner, August 2022 prediction The voice agents market is projected to grow at a 42% compound annual rate from 2026 to 2032. MarketsandMarkets - 07 · What is paid for
Labelled data, evaluation and guardrails
What buyers ask for: Data that makes the model right, and proof before every release that it still is.
What we do: Labelling by people who know the domain, synthetic data where it is honest, an evaluation suite that gates releases, and guardrails as tested code.
The evidence for Labelled data, evaluation and guardrails The evidence Source Demand for AI data annotation and labelling on Upwork grew 154% year on year. Upwork, In-Demand Skills 2026 Surge AI reached $1.2B in annualised revenue in 2024, ahead of Scale AI's $870M: human data is a billion-dollar line. Reuters, July 2025 The EU AI Act's high-risk obligations, now pointed at December 2027, require conformity assessment, technical documentation and monitoring: evaluation becomes a compliance artefact. Travers Smith on the EU timeline - 08 · What is paid for
Knowledge systems and enterprise search
What buyers ask for: Answers from our own documents, with the source shown, for staff and customers.
What we do: Retrieval over your documents with cited answers, a fallback to a person, and evaluation on the questions your people actually ask.
The evidence for Knowledge systems and enterprise search The evidence Source Retrieval-augmented generation is the common pattern in production systems, and general-purpose copilots took $8.4B of enterprise spend in 2025. Menlo Ventures, December 2025 - 09 · What is paid for
Integration and industry-specific applications
What buyers ask for: Put AI inside the product and the systems we already run, for our industry, not a generic tool.
What we do: AI products and features built into what you run, by industry, with the evaluation and deployment that let them pass review.
The evidence for Integration and industry-specific applications The evidence Source Applications took $19B, more than half of the roughly $37B enterprises spent on AI in 2025; industry-specific solutions $3.5B, led by healthcare. Menlo Ventures, December 2025 Demand for AI integration grew 178% year on year, and AI chatbot development 71%. Upwork, In-Demand Skills 2026 - 10 · What is paid for
Sovereign and on-premise deployment
What buyers ask for: Inside our boundary, in our country, with keys and updates we control.
What we do: On-premise and air-gapped installation with signed updates and rollback, private cloud in your account, and the five sovereignty questions answered in writing.
The evidence for Sovereign and on-premise deployment The evidence Source AI infrastructure absorbs $1.48 trillion of the $2.7 trillion forecast for 2026; AI services $576 billion, growing 13.9%. Gartner, September 2026
The figures are the analysts' and the surveys', quoted with their dates and linked to their sources. They describe the market, not our results. Where a figure is reported second-hand, the link goes to the report we read.
Bring the line item, not the idea.
Twenty minutes is enough to say which of these applies to your job, and whether we can take it.
A pilot starts within 5 working days of agreed scope · Nothing upfront · No seat licences