Amira van Weegen
Marketing Manager
July 20, 2026
Developing an AI Product: How We Turned the AI Hype into a Real Product at anny
8 min.
Introduction
AI is everywhere, but not every product delivers on the hype. What truly makes an AI product defensible and valuable is shown by the following insights gained from building anny AI:
According to an MIT study, 95% of AI projects yield no measurable impact — the issue is rarely the model itself, but rather the lack of integration into the product.
Simply building an AI wrapper around GPT offers no competitive moat — the decisive factor is which layer of the architecture we genuinely control (data, workflow, business logic).
Existing B2B SaaS products with clear use cases and robust customer data have a structural advantage over pure AI startups — provided they integrate AI deeply into their product development.
Context Engineering (MCP, custom context layer, live data) beats the model hype — that is how we at anny AI built an AI phone assistant that handles actual bookings instead of just sounding "smart".
Why the AI hype rarely leads to real product success
Between 2023 and 2026, billions have poured into Artificial Intelligence. Every other SaaS product launched an "AI feature." And yet, for most, a measurable business impact remains out of reach.
The statistics are sobering. A MIT study shows that around 95% of all generative AI initiatives in companies have no measurable effect on revenue or costs. In its Hype Cycle for Generative AI 2025, Gartner predicts that many generative AI projects will slide into the "Trough of Disillusionment." The reason? A mismatch between soaring expectations and real-world utility.
The issue is rarely the model itself. It lies in how AI is embedded into product development—or rather, how it isn't. AI is a powerful lever for smarter decision-making and streamlined processes, not an end in itself. Especially in complex SaaS workflows like resource booking, it's clear: without deep integration into existing systems, any AI feature remains purely cosmetic.
The AI Wrapper Dilemma: Why Some AI Products Lack Long-Term Defensibility
An "AI wrapper" is a product that, at its core, merely wraps a thin interface plus some business logic around an external language model API. The chatbot answers FAQs, summarizes texts, or generates images – but that's about it.
The problem: any other company can connect to the same API. Feature parity is achieved within weeks. No technical moat, no lock-in, no differentiation. The choice of technology stack is critical for developing an AI product, and "just OpenAI API" is not a stack. Integrating an AI model requires technology layers such as frontend, backend, and databases.
The comparison with Salesforce makes this clear: Salesforce deploys AI as an additional layer over an extremely rich CRM data model. What is defensible is the business logic, the permissions and roles model, and the deep integrations – not the underlying language model itself.
The core product question is: Which layer do we control ourselves?
Proprietary customer data (booking history, usage profiles)
Workflow engine and business rules
Integrations (calendars, HRIS, smart locks)
Permissions and roles model
A sustainable AI product combines its own strengths – existing data, domain expertise, user workflows – with language models, rather than just tacking on an "AI frontend". This insight from AI product development should feed into architecture design and the roadmap before the first AI features go live.
Two Strategies for AI Products in B2B SaaS – and Which One Wins
In the B2B SaaS market, we see two fundamentally different strategies for building AI applications.
Strategy A: New AI startups without a foundation. A focus on generic AI use cases, barely any proprietary customer data, often a pure AI wrapper around large language models. High pressure to find product-market fit quickly. The chances of success are slim if there is no in-house data or domain foundation.
Strategy B: Existing SaaS companies with a native AI layer. An existing user base, clearly defined workflows, and historical data. Established sales channels and market trends are already being served. A solid market analysis is often already done—target groups and the competitive landscape are well known.
At anny, we chose Strategy B. We built our AI layer on top of a stable foundation. For B2B SaaS founders, AI in product development means: embedding AI features along the existing customer journey (onboarding, search, booking, support)—no loose labs projects.
What really makes a great AI feature
Many AI features only deliver answers. An FAQ bot, a summary, a generated text. But a truly great AI function executes actions and changes the system state.
Three core criteria distinguish great AI features from mere gimmicks:
Execution over answers: The AI phone assistant in anny AI books a meeting room, reserves a parking spot, or completes a booking directly on the phone.
Live data over static prompts: Real-time availabilities, current pricing, company policies. anny AI checks team membership and booking windows in real-time, connecting external data sources to answer inquiries about rates or availability on the fly.
Automatically better with every model upgrade: The architecture separates agent logic from product logic. New models immediately deliver higher efficiency without requiring a complete rebuild of the business logic.
In a B2B context, great AI features always require governance: limits, escalation paths, and monitoring. The handoff between humans and AI must be seamless – with clear status indicators and logging for admins to easily understand AI decisions. Automation through AI is a fantastic way to elevate user experience and boost customer service!
anny AI Case Study: Why Customers Keep Calling Despite an Online Booking System
Online booking is supposed to take the pressure off your phones. Instead, customers often keep calling anyway – out of habit, for convenience, because they want advice first, or to clear up quick questions. Especially in sectors like sports classes or other consultation-heavy services, the phone remains the preferred channel. The problem is: Up to 85% of callers who don't get through on the first try won't call back. Every missed connection costs you potential bookings – revenue that is simply lost without anyone ever noticing.
This is exactly where anny comes in. For over 6 years, as a SaaS booking platform, we have been building the technical foundation to ensure that resources of all kinds – spaces, courts, equipment, appointments – can be reliably managed and booked. Over 60 million bookings and 1,500+ corporate clients across Europe and now also in the USA rely on anny today, including for:
Internal resource booking: Desk sharing, parking spaces, meeting rooms, fleet management
External venue and room rental
Visitor management
Equipment rental and leasing
Appointment scheduling
This infrastructure – real-time availabilities, pricing, rules, payments, all live and accurate – is the very foundation that anny AI builds upon. Instead of a call going to voicemail or your team having to juggle phone calls and in-person customers, anny AI takes over the entire conversation as a smart AI phone and chat assistant: A customer says, "I'd like to book two badminton courts tomorrow from 6 to 8 PM" – anny AI checks actual availability, explains pricing or rules if there are questions, and completes the booking directly, payment included. No callbacks, no holding patterns, and no missed calls outside of business hours!
Why "just use an existing GPT" isn't enough for real products
“Let's just plug GPT into our support hotline” – this is a phrase heard in many strategy meetings. In complex B2B SaaS scenarios, this simply doesn't work when real actions are required.
A single booking in anny entails the following complexity:
Dimension | Example |
|---|---|
Catalog logic | Different resource types: desks, rooms, parking spaces, equipment |
Pricing logic | Time slots, number of people, memberships, discounts |
Shopping cart | Multiple resources in a single transaction, cancellation policies |
Form fields | Required fields, consents, data privacy (GDPR) |
Booking rules | Team permissions, booking windows, maximum duration |
Generic AI models have no knowledge of this customer-specific business logic. Data privacy and legal requirements such as GDPR are equally crucial during data processing. Using existing AI models via API is often the more practical route here than training your own models – provided the context is right.
“Just prompting” doesn't solve the challenges: Without controlled tools and structured responses, hallucinations – such as incorrect pricing or availability details – can cause real damage. Especially with business-critical information, it is unacceptable for the technology to “guess.”
Developing an AI product therefore means: building an orchestrated environment in which language models are just one building block. Alongside them are APIs, permission concepts, validations, and observability.
Key Takeaway 1 – The Model Context Protocol (MCP) as the Foundation for AI Products
MCP is easiest to understand as "USB-C for AI": a standardized protocol for connecting tools and data sources to AI models. Any developer and engineer building an AI product today should be familiar with MCP or similar approaches.
In the context of anny AI, we use MCP for:
Accessing booking data, resources, and user accounts
Triggering actions (creating, canceling, and rescheduling bookings)
Integrating external systems (e.g., calendars or industry-specific APIs) via standardized interfaces
Especially for voice applications, MCP is critical: The agent must not "hallucinate freely," but instead only use defined tools. Tools must be callable with parameters that the model understands – e.g., start time, resource ID, and number of participants.
The limits are real: Having too many tools plus many parameters can overwhelm a model. The larger the tool catalog, the more important clear tool designs, naming conventions, and guardrails become.
For fast, small models in real-time telephony – where a 500–1,000 ms response time is the maximum – the context must remain lean. MCP helps to provide only the truly relevant tools and data instead of loading everything at once. For your own AI architecture, this means clarifying early on how tooling and data access will be standardized, ensuring you avoid vendor lock-in and chaos in your AI integration.
Key Takeaway 2 – Signed Conversation Context: Control hallucinations instead of just hoping for the best
Large language models are prone to hallucinations – especially when context windows are full or instructions are vague. In a B2B SaaS with a booking guarantee, this is a real risk: An invented resource or an incorrect price can cause real damage.
What "Signed Conversation Context" means from a product team's perspective: It is a system that actively manages the conversation context instead of just collecting it – securing certain parts and keeping them accessible across the entire conversation so that crucial information is never lost.
The results at anny AI:
Significantly fewer incorrect tool calls
Protection against "Context Overflow" – old but important information is never lost
More consistent results across numerous conversations
The golden rule is: "A prompt is a sentence, context is a system." It’s not about writing better prompts, but rather designing a robust architecture for context management.
Regarding quality and risks: anny offers a booking guarantee – the AI must never confirm a booking that doesn't comply with the customer's settings, technically secured via MCP tools. A residual risk of incorrect statements (e.g., about staff or prices) still remains, even if it is drastically minimized. That is why: Legal protection and clear disclaimers for customers are not just nice-to-haves, they are essential.
For founders, this context layer is the real value driver – this is where true differentiation from generic AI wrappers is built.
Summary: Building an AI product is all about context engineering – not model hopping
95% of AI projects yield no measurable impact. AI wrappers lack a competitive moat. In contrast: anny AI stands out as a prime example of an AI product that handles real tasks – phone calls, chat, booking, and payment. The difference lies not in the model, but in the architecture behind it.
2026 marks a turning point: It's bound to be the year of AI agents. However, successful agents rely on stable product and context systems, not just on ever-newer models alone. AI technologies like LLMs, Machine Learning, and Voice AI are evolving at lightning speed – innovation cycles are getting shorter than ever. Investing in Context Engineering today means you can swap out models tomorrow without having to rebuild the product throughout its entire life cycle.
For B2B SaaS founders, this means:
Start with the problem: Not "we need AI", but "which workflow can we make more autonomous?"
Design the architecture: Data, MCP/tools, context system – before you select the model.
Measure quality and impact: Not just "sounds good", but booking rate, error rate, and support costs.
Your real competitive edge lies in your existing customer data, workflows, and integrations. The opportunities are massive – provided you stop chasing generic AI features and start building a native AI layer directly into your product.
Try anny AI as a real-world example of AI integration in workspace and resource management. Start for free or book a demo and see for yourself how an AI phone assistant handles actual bookings.
![](data:image/svg+xml,<svg display="block" role="presentation" viewBox="0 0 120 40" xmlns="http://www.w3.org/2000/svg"><path d="M 55.501 13.158 L 59.216 13.158 L 59.216 30.294 L 55.501 30.294 L 55.501 27.83 C 54.088 29.78 52.094 30.741 49.482 30.741 C 47.113 30.741 45.112 29.883 43.427 28.138 C 41.743 26.393 40.919 24.23 40.919 21.729 C 40.919 19.192 41.743 17.073 43.427 15.321 C 45.112 13.576 47.106 12.718 49.482 12.718 C 52.094 12.718 54.088 13.678 55.501 15.599 Z M 50.071 27.207 C 51.616 27.207 52.925 26.693 53.955 25.667 C 54.986 24.604 55.501 23.306 55.501 21.729 C 55.501 20.153 54.986 18.855 53.955 17.821 C 52.925 16.758 51.616 16.245 50.071 16.245 C 48.526 16.245 47.216 16.758 46.186 17.821 C 45.156 18.848 44.641 20.153 44.641 21.729 C 44.641 23.306 45.156 24.611 46.186 25.667 C 47.216 26.693 48.526 27.207 50.071 27.207 Z M 79.191 19.771 L 79.191 30.294 L 75.475 30.294 L 75.475 20.153 C 75.475 17.586 73.997 16.179 71.628 16.179 C 69.016 16.179 67.265 17.792 67.265 21.113 L 67.265 30.294 L 63.549 30.294 L 63.549 13.158 L 67.265 13.158 L 67.265 15.357 C 68.398 13.612 70.149 12.718 72.591 12.718 C 76.542 12.718 79.191 15.387 79.191 19.771 Z M 98.82 19.771 L 98.82 30.294 L 95.104 30.294 L 95.104 20.153 C 95.104 17.586 93.626 16.179 91.257 16.179 C 88.645 16.179 86.894 17.792 86.894 21.113 L 86.894 30.294 L 83.178 30.294 L 83.178 13.158 L 86.894 13.158 L 86.894 15.357 C 88.027 13.612 89.778 12.718 92.22 12.718 C 96.171 12.718 98.82 15.387 98.82 19.771 Z M 114.049 13.158 L 118 13.158 L 111.401 30.836 C 109.753 35.258 106.964 37.384 103.013 37.142 L 103.013 33.689 C 105.353 33.828 106.655 32.691 107.59 30.33 L 107.759 29.986 L 100.468 13.158 L 104.522 13.158 L 109.642 25.593 Z M 36.232 8.186 C 36.188 5.033 35.879 2.203 35.827 1.77 C 35.798 1.066 35.371 0.458 34.694 0.164 C 33.959 -0.151 33.113 -0.004 32.539 0.531 C 19.487 12.828 19.413 12.916 19.377 12.96 C 19.229 13.128 19.156 13.224 19.082 13.334 L 19.075 13.348 C 19.053 13.385 19.031 13.414 19.009 13.451 C 18.84 13.737 18.744 14.147 18.766 14.477 C 18.788 14.807 18.862 15.152 19.001 15.511 C 19.023 15.592 19.046 15.665 19.075 15.739 C 19.2 16.091 19.568 17.161 19.641 21.003 C 19.722 25.102 19.207 26.965 18.935 27.962 C 18.884 28.138 18.847 28.292 18.81 28.438 C 18.766 28.578 18.656 28.952 18.759 29.465 C 18.81 29.707 18.906 29.912 18.965 30.022 C 18.994 30.074 19.023 30.125 19.06 30.176 C 19.119 30.264 19.215 30.404 19.369 30.536 L 26.712 37.487 C 27.065 37.824 27.536 38 28.014 38 C 28.271 38 28.536 37.949 28.779 37.839 C 29.456 37.545 29.875 36.929 29.875 36.218 L 29.875 24.42 C 29.875 23.504 29.419 22.924 29.073 22.565 L 28.978 22.462 C 29.706 21.803 31.943 19.764 34.452 17.227 C 35.717 15.944 36.298 12.989 36.232 8.186 Z M 27.624 34.605 C 27.624 34.847 27.33 34.972 27.153 34.803 L 21.363 29.326 L 21.378 29.311 L 21.407 29.282 L 21.495 29.194 L 21.532 29.157 C 21.569 29.12 21.606 29.091 21.643 29.054 C 21.65 29.047 21.665 29.04 21.672 29.025 C 21.724 28.981 21.775 28.93 21.827 28.878 L 21.841 28.864 C 21.893 28.82 21.937 28.776 21.988 28.724 C 22.003 28.71 22.018 28.702 22.033 28.688 C 22.084 28.636 22.136 28.592 22.194 28.541 C 22.261 28.482 22.319 28.424 22.386 28.365 L 22.665 28.116 L 22.754 28.035 L 27.308 23.958 L 27.44 24.098 C 27.447 24.105 27.455 24.112 27.455 24.12 C 27.631 24.303 27.631 24.34 27.631 24.428 L 27.631 34.605 Z M 32.848 15.658 C 29.875 18.657 27.278 20.967 27.256 20.989 L 27.249 20.989 L 26.623 21.546 L 25.792 22.294 C 24.586 23.372 23.173 24.633 22.187 25.52 C 21.996 25.689 21.701 25.535 21.724 25.285 C 21.841 24.171 21.915 22.756 21.885 20.945 C 21.804 16.743 21.378 15.511 21.194 14.983 C 21.186 14.969 21.179 14.954 21.179 14.939 C 21.157 14.844 21.135 14.778 21.12 14.741 C 21.106 14.705 21.091 14.668 21.083 14.639 C 21.047 14.536 21.076 14.426 21.157 14.353 C 22.239 13.326 29 6.954 33.252 2.951 C 33.422 2.789 33.694 2.899 33.716 3.127 C 33.826 4.293 33.966 6.214 33.995 8.23 C 34.062 13.971 33.135 15.365 32.848 15.658 Z M 17.302 27.955 C 17.03 26.957 16.522 25.095 16.596 20.996 C 16.669 17.146 17.044 16.076 17.162 15.731 C 17.184 15.658 17.214 15.585 17.236 15.504 C 17.368 15.152 17.449 14.8 17.471 14.47 C 17.493 14.147 17.39 13.737 17.228 13.444 C 17.206 13.407 17.184 13.378 17.162 13.341 L 17.155 13.326 C 17.081 13.224 17.008 13.121 16.86 12.952 C 16.824 12.908 16.743 12.828 3.698 0.538 C 3.125 -0.004 2.278 -0.144 1.543 0.172 C 0.866 0.465 0.439 1.074 0.41 1.778 C 0.358 2.21 0.049 5.041 0.005 8.193 C -0.061 12.996 0.52 15.951 1.793 17.234 C 4.316 19.786 6.568 21.839 7.267 22.47 L 7.171 22.572 C 6.825 22.932 6.369 23.511 6.369 24.428 L 6.369 36.211 C 6.369 36.922 6.788 37.545 7.465 37.831 C 7.715 37.941 7.973 37.993 8.23 37.993 C 8.709 37.993 9.172 37.817 9.533 37.479 L 16.912 30.499 C 17.044 30.382 17.133 30.25 17.184 30.184 C 17.221 30.132 17.25 30.081 17.272 30.044 C 17.346 29.92 17.442 29.722 17.486 29.472 C 17.581 28.974 17.478 28.6 17.434 28.446 C 17.398 28.284 17.353 28.13 17.302 27.955 Z M 3.389 15.658 C 3.102 15.365 2.168 13.979 2.249 8.237 C 2.278 6.221 2.426 4.3 2.529 3.134 C 2.551 2.907 2.823 2.797 2.992 2.958 C 7.245 6.962 14.006 13.334 15.087 14.36 C 15.161 14.433 15.19 14.543 15.161 14.646 C 15.154 14.675 15.139 14.712 15.124 14.749 C 15.109 14.785 15.087 14.851 15.065 14.947 C 15.058 14.961 15.058 14.976 15.051 14.991 C 14.867 15.519 14.44 16.75 14.359 20.952 C 14.322 22.763 14.396 24.178 14.521 25.293 C 14.55 25.542 14.249 25.696 14.057 25.527 C 13.071 24.64 11.659 23.379 10.452 22.301 L 9.621 21.553 L 8.988 20.989 C 8.871 20.893 6.325 18.62 3.389 15.658 Z M 9.077 34.81 C 8.9 34.979 8.606 34.854 8.606 34.612 L 8.606 24.42 C 8.606 24.34 8.606 24.296 8.782 24.112 C 8.79 24.105 8.797 24.098 8.797 24.09 L 8.929 23.951 L 13.476 28.028 L 13.792 28.314 L 13.851 28.365 C 13.866 28.38 13.881 28.394 13.895 28.402 C 13.932 28.438 13.976 28.475 14.013 28.512 C 14.035 28.534 14.057 28.548 14.072 28.57 C 14.109 28.607 14.153 28.644 14.19 28.68 C 14.212 28.695 14.227 28.717 14.249 28.732 C 14.293 28.776 14.337 28.812 14.381 28.856 L 14.41 28.886 C 14.462 28.937 14.513 28.981 14.558 29.025 L 14.587 29.054 C 14.624 29.091 14.661 29.12 14.69 29.157 L 14.727 29.194 C 14.756 29.223 14.793 29.26 14.815 29.289 L 14.837 29.311 L 14.852 29.326 Z" fill="rgb(0, 0, 0)" height="38px" id="bH388uk2E" transform="translate(1 1)" width="118.00000395534607px"/></svg>)