Subscribe
Share
edge ai
Trending → →

How much does AI cost? Well, it depends

By Timofey Kruk, Senior AI Scientist at ITRex, who develops and fine-tunes AI systems for complex business tasks.
Published on

TL;DR

  • A focused AI MVP built with proven tools typically costs between $20,000 and $60,000. Production-grade systems that combine multiple models with real-time data cost $80,000–150,000 for the AI part alone. The web, mobile, cloud, and embedded components are billed separately. Source: ITRex project estimates.
  • Five ITRex projects demonstrate the AI cost range: a generative art MVP at $19,000-34,000, a recommendation engine at $20,000-35,000, a speech analysis pilot at $36,000-56,000, an edge AI theft detection prototype at $65,000-125,000, and a rooftop analysis and lead scoring platform at $80,000-145,000.
  • Seven factors drive the cost of artificial intelligence in 2027: the type of AI you’re looking to build, its level of autonomy, your company’s data readiness, target accuracy, integration depth, compliance, and running costs. The last one is the hardest to forecast: one in five organizations already limits AI use because of operating costs (McKinsey, 2026).
  • Running costs can rival AI development costs: AI spend nearly quadruples, on average, when a tool moves from pilot to production (McKinsey). Agentic workflows consume roughly 15× more tokens than human-led interactions (OpenRouter), and Uber spent its entire 2026 AI coding budget in four months.
  • Compliance has a date: the EU AI Act’s obligations for standalone high-risk systems take effect on December 2, 2027. If you plan to deploy high-risk AI in Europe (think systems used in recruitment or credit scoring), you should start budgeting for documentation, oversight, and audits now.
  • You can cut AI implementation costs by starting with one workflow, using pre-trained models, routing tasks to the cheapest model that does the job, and measuring cost per accepted task instead of cost per token.

How much does AI cost? In the software world, it’s the same as asking how much a vehicle costs. A cargo bike and a freight truck both move goods, and nobody expects them to share a price tag.

The market numbers show how much money is in motion. Gartner expects worldwide AI spending to reach $2.67 trillion in 2026 and $3.64 trillion in 2027, with infrastructure taking about 56% of the total investments. Returns lag behind. In McKinsey’s 2026 State of AI survey, only 37% of respondents attribute any EBIT impact to AI, the same share as a year earlier.

That gap is both a budgeting and a technological problem. According to the same McKinsey survey, one in every five organizations limits AI use due to operating costs, which include token fees. So, the useful question is how much a specific AI system will cost to build, run, and govern until it pays for itself.

That’s the question ITRex’s AI consultants answer, focusing on customized and fully custom AI solutions. In this guide, you’ll find the factors that move the price, ballpark estimates from five projects in the ITRex portfolio, and the steps we use to keep AI pilots small enough to fund and useful enough to scale.

What are the key factors shaping AI cost?

In 2027, seven factors will determine where your project falls between a five-figure pilot and a seven-figure program. The first five shape the build, while the last two determine how much you pay each month after launch.

Factor Stage What it covers How it moves the budget
Type of AI Build Rule-based automation, classical ML, computer vision, or Gen AI on foundation models Sets the cost floor; Gen AI adds token fees and GPU-heavy inference
Level of autonomy Build How many decisions the system makes on its own, from answering questions to acting in your systems Chatbot: $5,000–40,000; autonomous agent: from $100,000
Data readiness Build Data quality, structure, labeling, availability, and privacy constraints Data preparation often takes ~30% of the budget
Target accuracy Build The error rate the business can tolerate and the testing needed to prove it Each extra point of accuracy costs more than the last
Integration depth Build Back end, data pipelines, APIs, user interface, and connections to legacy systems About 20% of the build budget, more with legacy ERPs
Compliance and regulation Build and run HIPAA, GDPR, AML, and EU AI Act duties for high-risk systems from December 2, 2027 Roughly 25% premium in regulated industries
Running costs Run Inference, token fees, cloud hosting, monitoring, and retraining after launch 15–20% of the build cost per year, plus usage fees

Factor #1: The type of AI you need

The technology you pick sets the AI cost floor.

Rule-based automation and classical machine learning work with structured data and require minimal hardware; see our ML cost guide for ballpark estimates from the ITRex portfolio. Computer vision needs labeled images or video and often specialized hardware. Generative AI runs on foundation models, the large pre-trained models behind tools like ChatGPT, which bring token fees and GPU-heavy inference.

Here’s the honest part: plenty of business problems don’t need Gen AI. If an if-then rule or a regression model solves the task, it will be cheaper to build, cheaper to run, and easier to audit. Our Gen AI cost guide explains when the premium is worth paying for.

Factor #2: The level of autonomy

The more decisions a system makes on its own, the more it costs. Most business AI is narrow AI, designed for a single task, such as recognizing text in PDF files. An AI agent goes further: it plans, calls tools, and acts inside your systems. An agentic system combines several agents, often alongside fixed workflow steps and human approval points, under one orchestration layer. In insurance claims, for example, one agent can handle document intake while another checks for fraud. Every agent you add to the setup means more integration work and more token consumption, so multi-agent systems land toward the upper end of the AI cost range.

ITRex estimates that a conversational chatbot typically costs $5,000-40,000, while the price for an autonomous agent that orchestrates workflows starts at about $100,000. Our AI agent development cost guide breaks that range down.

One common misconception: most deployed models don’t learn from new data on their own. Some systems, such as smart thermostats that learn a household’s schedule, are built to update themselves continuously. That capability has to be engineered, bounded, and monitored, and it adds to the upfront budget. An LLM-based agent can adapt its answers using conversation history, stored memory, and newly retrieved information, but the model itself changes only through retraining, fine-tuning, or a new version from your provider. That someone can also be your model provider, who rolls out a new version. Meanwhile, the world the model was built for keeps changing: customer behavior shifts, documents go stale, and accuracy slips. Any company planning to build an AI system should budget for monitoring and updates from the beginning.

Factor #3: Data readiness

Data quality and relevance matter more than data volume. A small, well-curated dataset often outperforms a large, noisy one. Gartner predicted that through 2026, organizations would abandon 60% of AI projects unsupported by AI-ready data. In a Gartner survey of 1,203 data management leaders, 63% said their organizations lack the right data practices for AI or aren’t sure they have them.

Structured data stored in relational database management systems (RDBMS) is the cheapest to work with. Unstructured data, such as emails, images, and videos, needs extra labeling and pipelines that keep it flowing between systems. In our experience, data preparation often eats around 30% of the budget before a final model is even selected.

When data is scarce or sensitive, you don’t always need a large new dataset. Few-shot learning adapts a pre-trained model to a new task from a handful of labeled examples. On a Gen AI sales training platform we built, we used few-shot prompting, a specific variation of few-shot learning, to cut the training cycle from three weeks of classic fine-tuning to a few hours. The approach works best for narrow tasks where an occasional error is tolerable. For precision-critical uses such as diagnostics, a larger labeled dataset remains the safer investment.

Where real records can’t be shared, as in medical imaging, engineers can augment small datasets or generate synthetic training data with Gen AI and validate it against real samples before training. A data platform assessment conducted before the AI work starts usually costs less than discovering the gaps mid-project. Our guide to data analytics costs covers that side of the budget.

Factor #4: Target accuracy

The more accurate an AI system must be, the more it costs, and each improvement costs more than the previous one. Getting a model to handle typical cases is relatively cheap. The unusual cases are the most expensive: an invoice with an unfamiliar layout, a blurry scan, or a shoplifter who turns away from the camera. Teaching the model to handle such instances requires additional training examples and testing cycles.

Gen AI adds a second cost: more capable models, and models set to reason longer before answering, charge more for every task they process.

In the Artificial Analysis benchmark data cited by Benjamin Aubron, our Gen AI evangelist, in his token economy analysis, GPT-5.6 Sol scored 59 on a combined intelligence index, and running the full benchmark cost about $2,824. DeepSeek V4 Pro scored 44 for about $176. In other words, 15 more points of capability cost roughly 16 times as much.

However, the cheapest option is rarely the safest choice. A model that is cheap to run but frequently fails generates additional work because someone must retry, review, or fix its output, and human time is rarely free. Here is an illustrative example for 10,000 document-processing tasks, assuming that each failure requires $4 of human labor to fix:

Illustrative example: 10,000 tasks, $4 per human fix Budget model Premium model
Inference cost per task $0.01 $0.05
Tasks completed without rework 70% 92%
Tasks needing a human fix 3,000 800
Cost of human fixes $12,000 $3,200
Total cost $12,100 $3,700
Cost per correctly completed task $1.21 $0.37

The premium model costs five times more per task and about a third as much per correct result. The numbers flip when errors are cheap to catch, so run the same calculation with your own volumes and review costs. This metric, cost per accepted task, is also how you should track AI spending after launch (more on that in the running costs section below).

How accurate should your AI system be?

To answer this question, think about what happens when it makes a mistake. A wrong suggestion in a draft that an employee reviews costs a few seconds of their time. The same mistake made by an agent that issues refunds or updates customer records on its own can lead to a financial loss or a compliance incident. In high-stakes uses, such as drones carrying blood samples or models that flag tumors in CT scans, there’s almost no room for error, and the extra data, testing, and validation are part of the price.

Higher accuracy doesn’t always require a bigger budget. Cleaner data (see Factor #3) often improves accuracy at a lower cost than switching to a more expensive model, and delegating each task to the right model (see the running costs section) keeps premium models for the cases that need them. Beyond that, experienced AI teams typically rely on three techniques:

  • Narrower scope. A model designed to recognize a few well-defined situations gains reliability faster than one designed to handle everything. Our edge AI theft detection prototype (Project 4) began with two types of thefts.

  • Confidence thresholds. An AI system handles the cases it is certain about automatically and forwards the uncertain ones to an employee, so people only review what needs a second look.

  • Model compression. For models a company runs on its own servers, techniques called quantization and pruning shrink the model so it needs less computing power. Accuracy usually drops slightly, so the AI team measures the effect on the company’s own data before going live.

The accuracy target is a business decision, and it should be defined at the start of a project, before development begins. For each task, business leaders and the AI team agree on the acceptable error rate, the cost of a single error, and who reviews the output.

At ITRex, we set these success criteria during an AI readiness assessment, which also checks whether the company’s data can realistically support the target. For the rooftop analysis platform (Project 5), the assessment confirmed technical feasibility and the data strategy before development started. Public benchmarks help shortlist models, but the company’s own data is the only way to make the target meaningful. Few other decisions can have such a large impact on the budget.

Factor #5: Integration depth

AI is usually the brain of a larger system, and the body costs money too. A typical AI-based solution needs a cloud back end, ETL or streaming pipelines, APIs for application integration, and a user interface. In our AI agent projects, integration and middleware account for about 20% of the development budget.

Three things can increase this “collateral” cost of AI development:

  • Legacy systems. Modern cloud applications usually offer ready-made APIs. Older ERPs, on-premises databases, and proprietary software often don’t, so engineers have to build custom connectors and test them carefully. Connecting a model to a 20-year-old ERP system is often more difficult than training or fine-tuning the model.

  • Real-time data. A sales forecast that refreshes overnight can run on scheduled batch jobs. An alert that has to reach a store manager within seconds needs streaming infrastructure, which costs more to build and operate.

  • Write access. An AI system that only reads data and shows insights is relatively simple to connect. One that updates records, issues refunds, or triggers orders needs permission controls, audit logs, rollback options, and far more testing. That’s a big part of why AI agents cost more than chatbots.

Businesses frequently use their existing channels for the cheapest integration. In our edge AI theft detection project (Project 4), for example, store owners receive alerts and video footage via WhatsApp, eliminating the need to create and maintain a separate mobile app.

Factor #6: Compliance & regulation

AI compliance increases the cost of any system that handles personal data or influences human decisions, such as hiring, lending, insurance, or medical care. Based on our project experience, an AI system that must comply with HIPAA, GDPR, or anti-money laundering regulations costs approximately 25% more to develop. The extra money is spent on risk assessments, technical documentation, audit trails, explainable outputs, and testing for bias and data leaks.

In the United States, there is no comprehensive federal AI law, so requirements vary by industry and state. Sector regulations, such as HIPAA for patient data and fair lending laws for credit decisions, apply to AI in the same way that they do to other software. States are adding AI-specific rules on top:

  • California. Under the CCPA regulations on automated decision-making technology, companies that use AI to make significant decisions about consumers—for instance, in employment, lending, or housing—must provide pre-use notices and opt-out options starting January 1, 2027.

  • Colorado. The state replaced its original AI Act with a narrower, notice-based law on automated decision-making, which also takes effect January 1, 2027.

A December 2025 executive order directs federal agencies to challenge state AI laws, but Congress hasn’t passed legislation that preempts them, so state rules remain in force.

For companies serving European customers, the EU AI Act sets the calendar. After the Digital Omnibus amendment, which entered into force on July 27, 2026, obligations for stand-alone high-risk systems apply from December 2, 2027, and for AI embedded in regulated products from August 2, 2028. High-risk uses include recruitment, credit scoring, and critical infrastructure. For these systems, risk management, technical documentation, human oversight, and data governance are all budget items. An EU AI Act compliance audit can help you identify which systems are high-risk and what each needs to meet those regulatory requirements.

On a brighter note, compliance affects AI development costs to a far lesser extent when it is planned in advance. Adding audit logs, explainability, or human review to a system that’s already in production usually means reworking its architecture. US state rules and the EU AI Act also ask for similar things, such as documentation, risk assessments, notices, and human oversight. A well-designed governance framework will help you launch an AI solution that’s compliant in several markets at once.

Factor #7: Running costs

Creating an AI system is a fixed cost. Once it is live, it continues to generate costs that traditional software does not, and the price increases with usage. We advise clients to set aside 15-20% of the initial budget each year for maintenance and model drift corrections, with additional usage fees.

Those usage fees have become harder to predict. LLM APIs have always charged per token, but many enterprise AI tools, such as coding assistants and office copilots, were sold as flat-rate licenses per user. In 2026, many of these providers moved to consumption-based pricing, so companies now pay for every request the tools handle. McKinsey’s research shows that AI spend nearly quadruples, on average, when a tool moves from pilot to production.

AI operating costs come from more sources than a single provider invoice:

  • Model & API fees. Providers charge for every token a model reads and writes, including instructions, conversation history, and attached documents. Models don’t remember earlier requests, so whatever context they need has to be sent again or retrieved from storage on each call. Systems designed to keep that context lean cost less with every request.

  • Cloud infrastructure. Hosting, computing for each model call, vector databases for retrieval, and data pipelines all contribute to the monthly cloud bill. Self-hosting a model shifts spending from API fees to this line, as the company pays for its servers whether they are active or idle.

  • AI inside other software. Many SaaS tools now bundle AI features with their own usage-based pricing, often outside the central IT budget. That spending is easy to miss, because it shows up under marketing, sales, or HR software rather than under AI.

  • Human review. Someone checks outputs, handles exceptions, and fixes errors, and that time belongs in the total cost. In high-stakes workflows, review time can cost more than the model itself, which is why cost per accepted task includes it.

  • Monitoring & retraining. Accuracy tracking, drift detection, and periodic model updates keep the system useful as data and business conditions change. This is where most of the 15–20% annual maintenance budget goes.

Of these five cost sources, model and API fees are the hardest to forecast, because they rise with every new user, every extra step an AI agent takes, and every retry. The section “How much does AI cost to run after launch?” below explains why falling token prices haven’t made AI cheaper to run and how companies keep these fees under control.

How much do AI projects cost? Five examples from the ITRex portfolio

To demonstrate how these factors play out in practice, we estimated the AI part of five projects from our portfolio. At one end sits a generative art MVP an artist needed in a few weeks for an exhibition, at $19,000-34,000. The other is a multi-model sales platform for a US solar manufacturer, priced between $80,000 and $145,000. The figures only cover the AI components; apps, dashboards, and hardware that work in conjunction with intelligent models must be estimated separately.

Project AI scope Main technologies Stage AI cost estimate
AI-driven art generator Image generation in an artist’s style Python, PyTorch, TensorFlow MVP $19,000–34,000
Intelligent recommendation engine Matching user queries to local service providers Amazon Personalize, Python, Amazon S3 MVP $20,000–35,000
Telemedicine speech analysis Speech-to-text and linguistic analysis of consultations Python, speech recognition SDKs Pilot $36,000–56,000
Edge AI theft detection Behavior detection on CCTV, repeat-offender recognition Amazon SageMaker, Python, edge device Prototype $65,000–125,000
Rooftop analysis and lead scoring Satellite image analysis, lead scoring, sales pitch analysis OpenCV, Python, Java, Google Maps API Production $80,000–145,000

Project 1: How much does a generative AI art MVP cost?

An MVP of a similar generative AI system could cost $19,000-34,000, depending on the training data (abstract vs. figurative art), output resolution, and deployment approach.

A couple of years ago, a visual artist asked ITRex to create a generative AI solution that would produce new paintings in his style and present them at an exhibition within several weeks.

We proposed a neural network based on Python frameworks (PyTorch, TensorFlow) that learned his trademark style and displayed results on his website. To keep the MVP lean, we recommended a 1000 × 1000 image resolution and local deployment, with the option to move to the cloud later.

Today, we’d most likely fine-tune a pre-trained open-source diffusion model instead of training a network from scratch. That would put a similar MVP closer to the lower end of the range.

Project 2: How much does an AI recommendation engine MVP cost?

Five map pins, one highlighted
Developing, testing, and deploying a similar recommendation MVP would cost $20,000-35,000.

A founder wished to replace cumbersome search filters on a B2C platform that connects users with local service providers with AI-powered recommendations. ML algorithms would analyze a user’s request and return a list of providers who matched it.

We chose Amazon Personalize, a managed AWS service that handles training, deployment, and hosting of recommendation models and charges no upfront fees. The back end was written in Python, with user data stored in Amazon S3. A managed service made sense for an MVP because it removed infrastructure work.

If you’re scoping a similar product today, also consider semantic search built on text embeddings, which can match free-text requests even before you have much interaction data.

Project 3: How much does an AI speech analysis pilot cost?

Stethoscope with glowing blue waveform
A basic speech analysis AI pilot may cost $36,000-56,000.

A healthcare technology company approached ITRex to add video recording to a telehealth system used in multiple US hospitals. The long-term idea was to apply facial recognition and natural language processing to consultation videos and help physicians improve doctor-patient interactions.

During discovery, we confirmed the pilot was feasible and chose Python with speech recognition frameworks and SDKs. The client then made a wise decision: the pilot focused on transcribing consultation audio, laying the groundwork for future analysis of communication patterns to reveal shifts in a patient’s well-being, which would require separate clinical validation.

Starting with one capability instead of the full vision kept the pilot affordable and gave the client real data before committing to facial analysis.

Project 4: How much does an edge AI theft detection system cost?

White dome security camera and edge box
The AI part of a similar prototype would cost $65,000-125,000.

A UK security provider wanted to give small retailers a theft detection solution they could afford: no expensive surveillance upgrades and no ongoing cloud analytics fees. ITRex’s edge AI experts designed a platform that runs video analysis on an in-store edge device, a Lenovo ThinkEdge SE70 optimized for AWS Panorama, Amazon’s managed service for computer vision at the edge. One device connects to up to ten existing CCTV cameras. When the system spots suspicious behavior, it sends a WhatsApp alert with the footage. We skipped a dedicated mobile app on purpose to keep costs down.

Here’s how the AI estimate for a similar prototype breaks down:

  • Data preparation & labeling: $10,000-20,000. The client’s proprietary dataset of shoplifting footage had to be curated and annotated.

  • Behavior detection models: $30,000-55,000. We trained models in Amazon SageMaker for two priority scenarios: concealing an item in clothing and grabbing an item before sprinting to the exit.

  • Edge optimization & deployment: $15,000-30,000. Models were compressed and tuned to process multiple camera streams on a single edge device.

  • Repeat-offender recognition: $10,000-20,000. Facial recognition flags known offenders when they enter the store. For more details, see our facial recognition cost guide.

The edge-first design paid off: 50% savings in the first year compared with cloud-based alternatives and up to 8× lower costs in later years.

Managed platforms can change under any project, and this one shows why that matters for edge AI budgets. In May 2025, AWS announced it would end support for Panorama on May 31, 2026, after which Panorama devices no longer function. For a system like this, the trained detection models and most of the application code carry over to new hardware, while the device runtime has to be replaced.

That’s why our estimate for building a similar system today assumes a vendor-neutral edge stack that can move between hardware providers. Facial recognition also processes biometric data, a special category under GDPR, so a data protection impact assessment is typically required before deployment.

Project 5: How much does an AI rooftop analysis & lead scoring platform cost?

The AI components of a similar platform would cost $80,000-145,000.

A US residential solar manufacturer relied on door-to-door sales, and every new state meant offices, licenses, and manual research to find suitable roofs. ITRex built a location intelligence and sales enablement platform that scores rooftops remotely, ranks homes by solar potential, splits them into territories, and coaches sales reps. The client cut lead qualification time by up to 80%.

Here’s how the AI development costs were distributed in the project:

  • Rooftop & area scoring: $45,000-80,000. Computer vision built on OpenCV, Python, and Java analyzes satellite data from the Google Maps API to estimate roof shape, orientation (azimuth), pitch, and shading, then factors in weather patterns.

  • Lead scoring & ranking: $10,000-20,000. The model combines solar viability with open real estate data, such as roofing material and age, to rank addresses.

  • Sales pitch analysis engine: $25,000-45,000. Reps record their door-to-door pitches in the mobile app, and NLP scripts assess emotional tone, compliance, clarity, and structure.

The platform relies on classical computer vision with OpenCV rather than heavier deep learning models, which keeps the processing cost per address modest. Part of that rooftop analysis functionality is now available off the shelf: Google’s Solar API returns roof segments, pitch, azimuth, and shading for the addresses it covers.

On a similar project today, we’d start with the API wherever data is available and build custom computer vision only for coverage gaps and client-specific scoring. That could bring the $45,000-80,000 rooftop scoring estimate down noticeably.

All five projects from the ITRex portfolio follow a clear pattern: if you want to reduce AI development costs, start with one clearly defined business problem and address it with the simplest technology available.

How much does AI cost to run after launch?

Operating costs for generative AI and agentic systems can quickly exceed the initial investment. Token prices continue to fall, but total bills rise, because cheaper tokens make more workloads viable, and agents consume far more of them.

Benjamin Aubron, ITRex’s Gen AI Evangelist, unpacks the mechanics in his analysis of the token economy’s billing problem. Every AI cost discussion should include three numbers from this analysis:

  • Cheaper tokens lead to more usage. When token prices fall, teams use the savings to run more AI tasks. According to Exponential View, every 10% drop in token prices comes with a 12–18% increase in the number of tokens consumed, so the total bill still grows.

  • AI agents use far more tokens. A chatbot answers a question once. An agent plans the task, pulls data, uses tools, checks its own work, and tries again when something fails. Each of those steps consumes tokens. OpenRouter estimates that an agent request uses about 15 times more tokens than a regular human interaction.

  • A lower price per token doesn’t guarantee a lower bill. Kimi K3 charges about half as much as GPT-5.6 Sol for every token it produces. On the same set of benchmark tests, though, Kimi K3 produced about 130 million output tokens to Sol’s roughly 70 million. Its total cost came out at $2,710 against Sol’s $2,824, only about 4% lower.

Some of the world’s largest companies have already learned these lessons the hard way. In late 2025, Uber gave roughly 5,000 engineers access to agentic coding tools and ranked teams on an internal leaderboard by how much they used them. By April 2026, the company had spent its entire AI coding budget for the year. Uber now caps spending at $1,500 per employee per month for each tool. Benjamin’s article describes two more cases: an Amazon project that matched author details to product listings and ran 860% over budget, and an unnamed company that spent roughly $500 million on AI in a single month because nobody had capped employee licenses.

The relevant metric is cost per completed, accepted task. Token price is just the meter.

Benjamin Aubron, Gen AI Evangelist, ITRex

How do you keep AI running costs under control?

The cost of running artificial intelligence in production is easiest to manage when the architecture and spending rules are established from the start. On our projects, four practices make the greatest difference:

  1. A tiered model setup. Not every task needs the most powerful model. Small models running on the company’s own servers handle routine work, such as sorting documents, extracting data, and processing sensitive information. Open models built for high throughput take on large volumes of work that isn’t urgent. Premium frontier models are kept for the hardest, highest-stakes tasks, which usually make up a small share of total usage.

  2. Spending limits for each workflow. Caps on tokens and dollars, daily spending alerts, and automatic stop rules catch an agent that’s running out of control within hours. In the Amazon case, the overspending went unnoticed for about five months.

  3. Cost per completed task as the main metric. Tracking the full cost of each correctly finished task, as in the example under Factor #4, shows whether AI is paying for itself. Token counts alone only show how busy the system was.

  4. Models that can be swapped. In Benjamin’s comparison, the same workload cost anywhere from $2.00 on Claude Fable 5 to $0.0352 on DeepSeek V4 Flash, a difference of about 57×. When a system can switch models without rebuilding the workflow on top, the company can move to a cheaper or better option as soon as one appears, whichever vendor offers it.

Why do so many AI projects fail or overrun their budgets?

Most AI projects fail for business reasons rather than technical ones. A Gartner survey found that only 48% of AI projects reach production, and those that do take an average of eight months. For agentic AI, Gartner predicts that over 40% of projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls.

The same Gartner analysis notes that many use cases positioned as agentic don’t require agentic implementations. We see the same thing in discovery workshops we run for our clients. The failure patterns repeat:

  1. No link to a business outcome. The project measures usage, accuracy, or tokens instead of revenue, cost, or time saved.

  2. Data that isn’t ready. Missing, siloed, or poorly labeled data stalls the model and inflates the budget.

  3. Moonshot scope. Overly ambitious programs promise to “completely change the way the company has operated for decades,” take years, and lose executive support before delivering value.

  4. Disconnected teams. Data scientists build models that software engineers can’t integrate, or that business users don’t trust.

  5. No cost controls. Usage-based pricing without budgets, alerts, or owners turns small mistakes into large invoices.

None of these problems requires more advanced AI to fix. Instead, they call for a clear AI strategy, a company-wide data strategy, and a delivery approach that proves value early. Learn more about common pitfalls in our guide to AI implementation challenges.

How can you reduce AI costs & see value sooner?

  1. 1
    Assess readiness
    Check data, infrastructure, and skills, and spot where simpler technology does the job
  2. 2
    Prioritize use cases
    Rank ideas by value and pick one or two that can prove it within a quarter
  3. 3
    Choose the simplest stack
    Combine custom, open-source, and off-the-shelf parts that are easy to swap
  4. 4
    Prepare only the data you need
    Clean and unify the data for the chosen use case, not the whole data estate
  5. 5
    Validate, then build
    A PoC confirms feasibility; an MVP puts the essentials in front of real users
  6. 6
    Govern and improve
    Track accuracy, drift, and cost per accepted task; retrain on fresh data

  7. Each new use case starts the cycle again at step 1

The cheapest AI project is the one scoped to a single, measurable workflow and expanded only after it pays off. It’s hard to pin down every requirement for a custom AI solution, or a realistic estimate, at the start. That’s why ITRex normally runs AI pilots in short, agile steps that show ROI early and make the case for further funding.

Below are several practical tips for keeping the cost of AI under control:

  • Assess readiness before you build. By conducting an AI readiness assessment, you can validate whether your data, IT infrastructure, and employee skills support the use cases you have in mind. Such assessments often reveal that a simpler technology, such as rule-based automation or standard BI, solves part of the problem for a fraction of the cost.

  • Prioritize use cases by value. Interview internal and external stakeholders to find the processes and decisions that AI can enhance. Next, rank candidates with a prioritization framework like MoSCoW, RICE, or Kano, and pick one or two that can prove value within a quarter.

  • Choose the simplest stack that works. Most cost-efficient AI projects combine custom, open-source, and off-the-shelf components. Pre-trained models available through APIs are a quick, low-cost way to validate an idea. Once traffic is stable, high-volume tasks can move to smaller or self-hosted models. Tools that make it easy to switch providers protect the project if a service is discontinued or prices rise, as the AWS Panorama example in Project 4 shows.

  • Prepare only the data you need. Data preparation is one of the largest items in an AI budget, so it’s worth limiting it to the chosen use case. By gathering, cleaning, and unifying only the data the model will actually use, you avoid paying to process your entire data estate. Our guide to data preparation for machine learning explains the techniques.

  • Validate with a PoC, then build an MVP. A proof of concept tests whether the idea works technically and is meant for internal use only. An MVP puts the essential features in front of real users, so you can see where the model needs improvement before investing in the full system.

  • Govern & improve after launch. Once the system is live, tracking accuracy, model drift, and cost per accepted task shows whether it still delivers value. Regular retraining on fresh data and human approval for high-stakes actions keep it reliable. Each new use case then goes through the same cycle, starting with step 1.

Our AI cost reduction playbook shows how companies recover these investments through automation, fewer errors, and better resource use.

In the end, how much does AI cost?

Most custom AI initiatives land between $20,000 for a narrow MVP and $150,000 for the AI part of a production system, plus 15–20% a year to keep it accurate. Autonomous agents and enterprise programs often sit above that range.

The table below summarizes ballpark estimates from the ITRex team.

Solution type Typical scope Estimated cost
Conversational chatbot Answers questions, hands complex cases to people $5,000–40,000
Narrow AI MVP One model, one workflow, proven tools $20,000–60,000
Production AI system (AI part only) Several models, real-time data, integrations $80,000–150,000
Autonomous AI agent Plans and executes multi-step tasks across systems $100,000–500,000+
Maintenance and retraining Monitoring, drift correction, updates 15–20% of build cost per year

Is the investment worth it? McKinsey’s 2026 survey shows that the 6% of companies it classifies as AI high performers attribute at least 5% of EBIT to AI. Nearly three-quarters of them have fundamentally redesigned workflows around AI. The money pays off when AI changes how work gets done and when someone owns the result long after launch.

AI cost: FAQs

  • Artificial intelligence cost comparison by project complexity level

    AI costs typically rise with solution complexity. A rule-based or simple ML chatbot typically costs $5,000-40,000 to build. A narrow AI MVP with one model and one workflow, such as our recommendation engine for a services marketplace, costs $20,000-60,000. A production-grade system that combines several models with real-time data, like our rooftop analysis and lead scoring platform, needs $80,000-150,000 for the AI part alone, and autonomous agents’ price starts around $100,000 and can exceed $500,000.

  • Typical AI project cost for mid-sized companies

    Based on ITRex project estimates, a first AI project at a mid-sized company typically costs $35,000-150,000 for the artificial intelligence components alone, excluding apps, dashboards, and hardware. Where a project lands in that range depends mainly on how ready the company’s data is and how many systems the AI has to connect to. A single-capability pilot, such as the speech analysis system ITRex built for a US healthcare technology company, sits at the lower end ($36,000-56,000), while a multi-model platform that draws on mapping services and real estate data, such as ITRex’s rooftop analysis and lead scoring solution, reaches $80,000-145,000. Mid-sized companies can optimize AI costs by validating one use case with a proof of concept before committing to a full-blown project.

  • Comparing AI cost versus traditional software development

    AI development projects usually cost more than traditional software projects of a similar scope because they involve data preparation, R&D, and ongoing model costs. Based on ITRex project experience, data preparation consumes around 30% of an AI budget, and keeping a model accurate adds 15–20% of the development cost each year, plus usage fees for paid model APIs. Traditional software also needs maintenance, but its behavior and costs stay largely predictable after release. The gap shrinks when ready-made AI components fit the task: in ITRex’s guide to software development costs, a cancer-information chatbot built on a no-code platform with a ready-made large language model was scoped at $5,000-10,000 plus about $100 a month to run, against an estimated $50,000 for traditional development. And where a rule-based filter or a standard search feature solves the problem, it’s usually cheaper than any AI model.

  • Artificial intelligence cost range for customer service chatbots

    Based on ITRex estimates, a customer service chatbot typically costs $5,000-40,000 to build, depending on how many systems it interacts with and how many languages and channels it supports. An FAQ bot that hands complex questions to human agents sits at the low end. An AI agent that checks orders in your ERP, processes refunds, and updates the CRM starts closer to $100,000. If you’re creating a customer service bot, it’s usually cheapest to start with the most common questions and a smooth handoff to human agents, then add actions like refunds once the bot has proven reliable. Budget for usage fees, which increase with each conversation the bot handles: Gartner predicts that by 2029, AI agents will resolve 80% of common customer service issues without human intervention, so the volume is likely to rise further.

  • AI development cost for computer vision applications

    Based on ITRex estimates, the AI part of a computer vision project typically costs $45,000-125,000, depending on how much data needs labeling, how accurate the model must be, and whether it runs in the cloud or on edge devices. The computer vision module of ITRex’s rooftop analysis platform, which reads satellite data to estimate roof pitch, orientation, and shading, falls in the $45,000-80,000 range, while the edge AI theft detection prototype, which recognizes two theft behaviors across up to ten cameras, costs $65,000-125,000. Computer vision costs drop when you can use pre-trained models instead of training them from scratch. ITRex uses MotionRex AI, our proprietary platform of pre-trained, edge-optimized computer vision models that run on hardware ranging from NVIDIA Jetson boards to low-power microcontrollers, for tasks like human pose estimation, motion tracking, object detection, and shrinkage detection in retail.

  • Full artificial intelligence cost for predictive maintenance platforms

    A predictive maintenance platform has three cost blocks: sensor data collection, the AI models, and integration with maintenance and ERP systems. If your equipment already streams usable sensor data, the AI part of a pilot on one production line typically costs $50,000-100,000 by our estimates. Retrofitting sensors, building IoT pipelines, and connecting to a computerized maintenance management system (CMMS) are billed separately and can exceed the AI budget. Our guide to IoT costs covers the hardware and connectivity side.

Ready to put a number on your AI project?

At ITRex, every AI estimate begins with three questions: what data do you have, which workflow will the AI support, and what outcome would justify the investment. We design the project around that outcome, so if rule-based automation or an intuitive BI dashboard can do the job for less, we’ll recommend it instead. We select the platform that best fits your needs, even if it comes from a vendor with whom we do not collaborate, and take responsibility for the system’s accuracy and operational costs after launch.

TABLE OF CONTENTS
What are the key factors shaping AI cost?Factor #1: The type of AI you needFactor #2: The level of autonomyFactor #3: Data readinessFactor #4: Target accuracyFactor #5: Integration depthFactor #6: Compliance & regulationFactor #7: Running costsHow much do AI projects cost? Five examples from the ITRex portfolioProject 1: How much does a generative AI art MVP cost?Project 2: How much does an AI recommendation engine MVP cost?Project 3: How much does an AI speech analysis pilot cost?Project 4: How much does an edge AI theft detection system cost?Project 5: How much does an AI rooftop analysis & lead scoring platform cost?How much does AI cost to run after launch?How do you keep AI running costs under control?Why do so many AI projects fail or overrun their budgets?How can you reduce AI costs & see value sooner?In the end, how much does AI cost?AI cost: FAQsReady to put a number on your AI project?
Ready to kick off your AI project?
Contact us
background banner
edge ai

Talk to our AI consultants to estimate your project, or start with an AI readiness assessment or a proof of concept to test the idea before you commit a full budget.