AI Product Development: When to Add Generative AI Features

Maintaining a product roadmap is a difficult task, as there are many factors involved, with moving targets and continuously changing circumstances. It gets harder with a technology the team has no experience of. In AI product development the deciding question is no longer whether generative AI works. It is whether the product, the data and the commercial model can carry it.

When this post first appeared in June 2024, that judgement rested on immature models and on one hard price that has since gone stale. The four things that decide feasibility have not changed: readiness, cost, pricing structure and responsible use. Each now has published evidence attached, plus the question an ERP buyer actually asks: how much of this already ships inside Microsoft Dynamics 365 Business Central?

Is generative AI ready to build product features on?

Yes, for a defined set of tasks, and the adoption figures make that case better than any vendor can. In 2025, 20.0% of EU enterprises with 10 or more employees used AI technologies, up from 13.5% in 2024, according to Eurostat data published on 11 December 2025. Size separates the field sharply: Eurostat’s figures for 2025 put usage at 17% of small enterprises, 30.36% of medium and 55.03% of large.

Output reliability has improved substantially since 2024, and wrong answers still occur, so they have to be designed around, not assumed away. Some of that is prompt work and the rest is interface work: an interface that presents a suggestion as a suggestion makes a wrong one cost a click rather than the user’s confidence.

Where the models come from, and what they are called now

The names have moved, which matters when reading anything written before this year. Microsoft’s hosted platform is now Microsoft Foundry, the name that replaced Azure AI Studio and Azure AI Foundry, with Azure dropped from it. Azure AI Services is now Foundry Tools, and the catalogue is the Foundry Models catalog, which Microsoft says carries more than 10,000 models. Azure OpenAI has not been retired: Microsoft states that both services are generally available before and after an upgrade, and that upgrading an Azure OpenAI resource is opt-in.

A large provider removes availability risk and gives a contractual position on data handling. It does not remove incorrectness, and the guarantee is narrower than the 2024 version of this post implied. On its Copilot overview page for Business Central, Microsoft states that it “doesn’t read your prompts or use your data to train AI models without your explicit permission”, and that “for customers within the EU Data Boundary, data doesn’t move outside that boundary”, while separately listing the categories that do still leave it.

What changed most is context, not capability

The biggest shift since 2024 is how much you can put in front of a model at once. Anthropic’s published pricing states that “Claude 4.6 and later models ... include the full 1M token context window at standard pricing”, so features that failed in 2024 because the document would not fit are now ordinary. Meanwhile gpt-3.5-turbo, this post’s original cost benchmark, no longer appears on OpenAI’s pricing page at all.

What generative AI features cost to run in 2026

Text processing is still cheap enough to build product features on, but the rate to plan against is the one published on the provider’s own pricing page on the day the estimate is built, not a figure quoted in an article. What decides a feature is not the headline rate but the spread: the gap between the cheapest capable model and the frontier one is wide enough that the model you pick, not the feature you design, sets the unit cost.

This version prints no per-token price on purpose. Rates change often and a wrong one stays online for years, so read them at the provider on the day the estimate is built: OpenAI publishes its API rates here and Anthropic publishes its rates here. Two structural details survive any price change. OpenAI prices its long context variants at double the short context rate for the same model. And other modalities are not priced on the token unit at all, so a token estimate does not transfer: image generation is billed per token of output and video per second of footage.

Where the cost still runs away

A feature exposed to the public in an unrestricted way can still see costs climb fast. What decides it is the model, the amount of prompting involved, which can grow quite large, and how the feature reaches end users. Generating a product description from product details runs behind the scenes and is predictable. Helping a user draft text is not. Cap the second kind before launch, not after the first surprising invoice.

Crafting the right pricing and subscription structure

Two pricing models are available, and choosing between them is a cost-control decision before it is a revenue decision: a fixed subscription, or consumption-based charging. A subscription is a good source of revenue, but only if people subscribe, and once a customer has paid for a few separate per-user AI add-ons the next one meets resistance regardless of quality. The structural problem is heavy users: they raise the average per-user cost, which raises the price, which makes the majority feel overcharged. Tiers with different usage allowances help. Consumption-based charging creates the opposite worry, an unpredictable bill, answered with transparent pricing and generous caps under fair use.

Microsoft runs both methods side by side in Business Central: core Copilot sits inside the licence, and the agents that do unattended work bill by consumption. Whichever method you choose, price on a concept the customer already understands instead of on tokens, which end users neither want to think about nor need to understand. They need to know how many articles they can get help with, or whether their plan covers it.

What “adding AI” means when the system is an ERP

For a Business Central customer, most of the AI being discussed at board level has already shipped inside the product. In Zoosh Digital’s experience, Business Central’s out-of-the-box functionality covers around 80% of SMB requirements without any bespoke development, and an AI requirement starts in the same place: establish what is switched on, what release status it carries, and what bills separately.

What Copilot in Business Central does today

Microsoft groups the current capabilities on its AI and Copilot in Business Central page. For everyone: Chat, Analysis assist, Summarize and Autofill. For finance: Payables Agent, E-document matching, Bank account reconciliation assist and the Finance solution in Microsoft 365 Copilot. For sales: Sales Order Agent, sales line suggestions and marketing text suggestions. Inventory gets item substitute suggestions, company setup gets number series suggestions.

Two constraints shape any plan built on that list. Copilot is exclusive to Business Central online, and Microsoft states that it “isn’t available for other deployment types, such as on-premises or private cloud”.

The second is release status, and it is worth checking, not assuming, because Microsoft’s own pages are not consistent about which of these capabilities carry a preview label. The documentation page for the capability itself is the reliable check, and where Microsoft is explicit it is explicit there: Chat is described in its own FAQ as “a production-ready preview feature”, the consumption billing page labels Expense Agent and the designing and coding agents “(preview)”, and the administration page lists custom agent development as “Custom Agent (preview)”. That status is fine for evaluation and poor as the foundation of a customer promise.

What it costs, and where the bill actually starts

Microsoft states on the same Copilot overview page that “Copilot in Business Central is included with your Business Central license at no extra cost”, while noting that “fair-use policies, quotas, or pricing might be introduced later”. The agents are the exception. Per Microsoft’s documentation on consumption billing, Expense Agent (preview), Payables Agent, Sales Order Agent and the designing and coding agents (preview) bill in Copilot Credits: 50 credits per uploaded receipt for Expense Agent, 50 credits per invoice plus 5 per invoice line for Payables Agent, and 2 to 5 credits per action for Sales Order Agent. Microsoft’s worked examples on that page put 100 invoices a month at three lines each at 6,500 credits, and 100 sales requests with a 50% attachment rate at 1,650 credits.

The assistive features are included, the autonomous ones are metered, and the volumes are predictable enough to estimate first. Microsoft publishes no euro or sterling price per Copilot Credit for Ireland or the UK, so the credit pack price is a question for your implementation partner.

One availability gap worth checking from Ireland

On Microsoft’s country and language availability list, Expense Agent is limited to Australia, Austria, Canada, Denmark, France, Germany, New Zealand, Spain, the United Kingdom and the United States. Ireland is not on that list. Payables Agent and Sales Order Agent are published as available in all countries and regions.

Built-in Copilot or custom AI: how to tell which you need

Work through four layers in order and stop at the first one that meets the requirement. Most requirements stop earlier than the person raising them expects. In Zoosh Digital’s experience, ungoverned customisation requests can represent 20 to 30% of total project budget for businesses that approve every request without a business case process, and an AI request is a customisation request with a model attached.

• Enable what is already there. An administrator controls Copilot and agent capabilities, and most of them use permission sets for access control. Nothing is built.
• Configure and extend inside the Microsoft platform. Business Central data reaches Power Apps and Power Automate through the Business Central connector, and Copilot Studio agents connect through the Business Central MCP server and the same connector. Low-code work.
• Build a Business Central extension. The right layer when the behaviour must live inside the ERP and follow its posting logic. Custom agent development through the AI development toolkit is still labelled preview.
• Build bespoke software. For when the feature is the product itself, not the ERP, and the model, the interface and the pricing all belong to the business that ships it.

What matters commercially is who owns the unit cost. At the first three layers the business consumes a priced Microsoft capability and the exposure is a credit rate. At the fourth it owns the model choice, the prompt size, the caps and the margin. Zoosh Digital’s position on this is deliberately neutral, because Zoosh Digital builds bespoke software, builds Business Central extensions, and implements off-the-shelf Business Central. There is no single answer it is incentivised to sell. The same logic for non-AI requirements is set out in Business Central customisation versus configuration, and what makes a good ERP system covers where the AI layer sits in a modern ERP architecture.

Transparency and responsible AI implementation

Telling people when they are dealing with AI is now partly a legal obligation in the EU, not only a design preference. According to the European Commission’s AI Act implementation timeline, the “transparency rules (Article 50) start to apply” from 2 August 2026, with enforcement starting at national and EU level on the same date. Rules for high-risk systems in Annex III follow from 2 December 2027.

The design problem is older than the law. What we need to avoid is the so called “Uncanny Valley”. People don’t like automated features which are too simple, which just “don’t get it”: everyone dreads call centre menu structures where all they want is a human agent. Present something as if it came from a human, almost but not exactly, like a robot which doesn’t blink, and we feel unease instead.

We shouldn’t present things in a form which they are not. In Zoosh Digital’s experience AI supported features are accepted the most when it is shown transparently what comes from an AI, whether that is a robot image, a cute name for the product’s “robot assistant” or a label on the generated field.

Being responsible also means being prepared for abuse and for incorrect output. Abuse can be prevented with proper prompts: we need to sanitise user inputs, much as avoiding SQL injections was a hot topic some years ago. The major model providers offer basic content filtering against hate, violence or sexual inputs. Incorrect output can be recognised with grounding, where generated output is referenced against other materials.

When not to add generative AI yet

Four situations make waiting the right decision, and recognising them early saves more than any of the cost work above.

• The data is not ready. A generative feature inherits the quality of the records it reads, so if reporting is assembled, not generated, an assistant will describe those problems back fluently. The five patterns in The Zoosh ERP Readiness Indicators are a reasonable pre-check: if three or more apply, data foundations come first.
• Nobody owns the wrong answer. Every generative feature needs a named person who reviews, corrects or overrides output. Microsoft builds that into the Payables Agent, which it describes as showing “invoice drafts to agent supervisors for review” (Microsoft Learn).
• The exposure is unbounded. If the feature is public, free and unlimited, the unit cost is theoretical.
• The dependency is a preview. Committing a customer promise or a launch date to one puts the roadmap on someone else’s release timetable.

Where that leaves a roadmap in 2026

In 2024 the debate was whether generative AI was good enough to put in front of a customer. In 2026 it usually is, and the decision is which layer you buy it at: enable it, configure it, extend the ERP, or build it yourself. That is a commercial question about who owns the unit cost and the correctness risk, and the published rates and credit tables can answer it. Zoosh Digital is a Microsoft Business Central and technology partner working with organisations across the UK, Ireland and Europe, supporting cloud ERP, custom development and long-term optimisation. If you are weighing an AI feature against an ERP capability you may already own, that conversation works best with the licence details and the volumes in front of you.

FAQ

Is it worth adding AI to our ERP now, or should we wait?

For a Business Central customer on the online service, evaluating costs nothing, because core Copilot is included in the licence at no extra cost. Waiting makes sense when data quality is poor, when nobody owns reviewing AI output, or when the capability you need is one Microsoft still labels preview.

What does Copilot in Microsoft Dynamics 365 Business Central cost?

Microsoft states that Copilot in Business Central is included with the licence at no extra cost, though fair-use policies, quotas or pricing might be introduced later. The agents bill separately in Copilot Credits: 50 per uploaded receipt for Expense Agent, 50 per invoice plus 5 per line for Payables Agent, and 2 to 5 per action for Sales Order Agent. Microsoft publishes no per-credit price for Ireland or the UK.

Can Copilot do what we need without custom development?

Often yes. Work through the layers in order: enable the built-in capability, then configure or extend through Power Apps, Power Automate or a Copilot Studio agent on the Business Central MCP server, then build a Business Central extension, and only then bespoke software.

How do we know our data is ready for AI features?

Check the records the feature would read before checking the feature. If month-end close is a recovery exercise, if spreadsheets overlap without syncing, or if reports are assembled by hand, a generative feature will report those problems back rather than solve them. The Zoosh ERP Readiness Indicators set out five such patterns.

Referenced sources

• Eurostat. ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2.
• Eurostat. ec.europa.eu/eurostat/statistics-explained/index.php.
• Microsoft Learn. learn.microsoft.com/en-us/azure/foundry/what-is-foundry.
• Microsoft Learn. learn.microsoft.com/en-us/azure/foundry/how-to/upgrade-azure-openai.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/copilot-overview.
• Anthropic. platform.claude.com/docs/en/about-claude/pricing.
• OpenAI. developers.openai.com/api/docs/pricing.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/ai-in-bc.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/chat-with-copilot-faq.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/dev-itpro/administration/tenant-admin-center-manage-consumption-billing.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/copilot-agents-region-language-availability.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/enable-ai.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/dev-itpro/powerplatform/powerplatform-integration-overview.
• ai-act-service-desk.ec.europa.eu. ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act.
• Wikipedia. en.wikipedia.org/wiki/Uncanny_valley.
• Microsoft Learn. learn.microsoft.com/en-us/dynamics365/business-central/payables-agent.