The product team shows the board a video from the pilot: a sensor taped to a compressor on the shop floor, a temperature chart refreshing every few seconds on a laptop, a push notification when the reading crosses a threshold. Everything works. The pilot took six weeks and cost about as much as a few monitors. The board asks when it can order five thousand units and start selling the service to customers.
Six months later the team discovers that the modem in the pilot uses a network technology the operator plans to retire before the product’s intended end of life. That the microcontroller does not have enough flash to hold two firmware images and update safely. That the cloud credentials are shared by every device and compiled into the firmware. That the legal team is asking about the radio equipment delegated act, the Cyber Resilience Act support period, and how customers will retrieve their data under the Data Act. None of this came up in the pilot, because the pilot was testing something else.
This article is for CTOs and product leaders deciding whether and how to use IoT consulting services before building a connected product. The position it defends is this: an IoT consulting engagement that starts with choosing a cloud platform, or with a quick proof of concept, starts at the wrong end. Its first deliverable should be a decision record that answers what one device costs over its whole support period, how it will be updated, who manages its identity, how it will be retired, and which rules apply on the day it ships. The platform and the hardware should follow from those decisions, not precede them. A competent reader may disagree, because a fast pilot has real benefits. The rest of this piece argues that in IoT those benefits are smaller than they look.
Why does a working IoT pilot prove so little?
An IoT pilot tests the easiest part of the project: whether a sensor can take a reading and send it to a server. That question is rarely open. Off-the-shelf modules, evaluation boards and cloud services mean the answer is almost always yes. The hard questions come later. What does it cost to keep thousands of devices running for seven years? What happens when an update is interrupted halfway through? Who pays for data when a customer switches carriers? How do you retire the product when the company changes strategy?
In May 2017 Cisco published a survey of 1,845 IT and business decision-makers in the US, the UK and India. It found that 60% of IoT initiatives stall at the proof-of-concept stage and only 26% of companies considered an initiative a complete success. A third of completed projects were not considered successful. The top success factor respondents named was collaboration between IT and the business side, cited by 54%, not technology. The survey is several years old and the market has matured, but the mechanism has not changed: a pilot shows that something can be done, not that it can be done profitably.
The problem is not that pilots are bad. It is that their success is often treated as permission to scale. A team that spent six weeks on an evaluation board with mains power and steady Wi-Fi has not measured battery life in the field, transmission cost on a weak signal, or how the device behaves after three days offline. Those numbers drive the business model, and the pilot does not produce them.
So reverse the order. First write down the decisions that are expensive to change once devices are deployed, and estimate their cost. Then build a prototype that tests the riskiest of those decisions. Such a prototype makes a less impressive board video, because instead of a temperature chart it shows, for example, that an update survived a power cut in the middle of a flash write. But it is the one that answers the question the board actually cares about.
What should IoT consulting deliver, and what should it not sell as the first step?
The market for IoT consulting is broad: platform integrators, module vendors with support engineering teams, and software companies that build device firmware, cloud back ends and applications. All of them can add value, but their first recommendation often reflects what they sell. A platform integrator will propose a platform. A module vendor will propose its module. That is not an accusation, just a reason to know what you should expect from the pre-build stage.
A good pre-build engagement should deliver five things. First, a lifetime cost model per device: hardware, installation, connectivity, cloud, updates, support, replacements and decommissioning. Second, an architecture decision record with rationale and rejected alternatives: connectivity technology, the split of processing between device, gateway and cloud, the data model, the update mechanism. Third, a map of legal obligations for each target market, with dates. Fourth, a risk register with a plan for testing each risk in a prototype. Fifth, a prototype plan that tests those risks rather than just the sensor.
What is not worth buying as the first step? A cloud platform rollout before you know how much data will flow and in what shape. A batch of devices before you know whether the microcontroller can fit a secure update mechanism. A polished customer dashboard before you know which data customers are entitled to retrieve and in what format. Each of these may be needed eventually, but bought too early, each one closes options before anyone has costed them.
Agree up front who owns the output of this stage. The decision record and cost model should belong to you and be usable by a different build vendor if you choose one. If the consulting firm is also a candidate for the build, write into the contract that the stage’s output is not conditional on further work. The same logic applies when choosing an MVP development partner: a discovery stage should reduce the buyer’s risk, not lock the buyer in.
What does connectivity really cost per device over its lifetime?
Connectivity is the first decision that looks technical and turns out to be financial. Choosing between Wi-Fi, Bluetooth plus a gateway, LoRaWAN, cellular IoT technologies (LTE-M, NB-IoT), full LTE or satellite affects module cost, power draw, range, the billing model, and the risk that the network disappears before the product reaches end of life.
Wi-Fi costs nothing per byte but depends on the customer’s network. Every router swap, password change or security policy update on the customer side becomes a ticket for your support team. Bluetooth with a gateway moves the problem to the gateway, which must itself be managed, updated and powered. LoRaWAN offers long range at very low power, but in unlicensed spectrum it limits payload size and airtime, so it is a poor fit for frequent large messages or full firmware images. LTE-M and NB-IoT, introduced in 3GPP specifications for IoT devices, offer power-saving modes (PSM, eDRX) that let the modem sleep most of the time, but availability and roaming differ by country and operator. Full LTE gives you bandwidth that you pay for in energy and module price.
The second issue is network lifespan. Operators retire older generations to reclaim spectrum. AT&T shut down its 2G network in the United States at the start of 2017, and Swisscom ended guaranteed 2G availability in Switzerland at the end of 2020. Dates in other countries vary and sometimes move. If your device needs to work for ten years and its modem relies on a technology whose shutdown the operator has already announced, you are buying a problem with a due date. Ask each target operator about its plans for the technology your module uses, and put the answer in the decision record.
The third issue is transmission cost, which compounds with fleet size and years. Take an illustrative example. A device sends a 200-byte payload every five minutes, which is 288 messages and about 58 KB of payload per day. Add protocol overhead: IP, TCP and MQTT headers, TLS encryption and, on every reconnect, a full TLS handshake that can weigh several kilobytes. With messages this small, overhead easily exceeds the payload. Then add firmware updates: a 1 MB image shipped four times a year is 4 MB per device per year, before retries after interrupted downloads.
Now multiply by the fleet. Assume, hypothetically, that an IoT SIM plan costs EUR 1 per month and the fleet is 10,000 devices over seven years. That is 10,000 × 12 × 7 × EUR 1, or EUR 840,000 for connectivity alone, before cloud, support and replacements. The rate is a placeholder to replace with a real operator quote, not a market price. The point is that a difference of a few cents per device per month becomes hundreds of thousands at this scale, and that decisions about reporting frequency, payload format and connection handling affect the total more than the choice of cloud provider does.
What follows for the pre-build stage? Message format (binary or text), reporting frequency, on-device aggregation, connection keep-alive and update strategy should all be costed in a spreadsheet before anyone selects a module. Processing part of the data on the device or gateway rather than shipping everything to the cloud often lowers transmission cost. We cover that approach in more depth in our article on edge computing and where computing power is needed most.
How do you design the device data model so you are not rewriting the back end in three years?
In IoT the data model outlives every other part of the system, because part of it is baked into devices in the field that cannot always be updated. You can rewrite a back end in a quarter. You cannot rewrite five thousand devices sending data in an old format unless you planned for that from the start.
The first decision is to separate three kinds of data. Telemetry is measurements over time: temperature, voltage, location. State is the current configuration and status of the device: firmware version, thresholds, operating mode. Events are things that happened: an alarm, a reboot, a failed update, an opened enclosure. Each kind has different requirements for frequency, retention and delivery guarantees. Mixing them in one message means that a year later nobody can tell whether missing data means a sensor fault, lost connectivity or a deliberate configuration change.
The second decision is schema versioning. Every message should carry its format version, and the back end must handle several versions at once, because a fleet never updates in a single day. Adding a field should be safe for older consumers, and removing one needs a transition period. It is the same discipline teams apply to public APIs, except that in IoT some of the API’s clients are devices you cannot update quickly.
The third decision is the cloud-side representation of the device, often called a device shadow or digital twin. It stores the state the device reported and the desired state set by the system or the user. That way an application does not have to wait for the device to wake up before changing its configuration: it writes the desired state, and the device picks it up on its next connection. If you do not design this mechanism at the start, every application team will build its own inconsistent version.
The fourth decision is the basic attributes of every reading: unit, time of measurement (not time of receipt by the server), time source, device and sensor identifiers, and a quality flag. This sounds trivial, but without device-side timestamps, data buffered during a three-day outage arrives with the wrong times. Without explicit units, after a sensor supplier change part of the fleet reports Celsius and part reports Fahrenheit.
It is worth looking at existing device management standards, such as LwM2M from OMA SpecWorks, before designing a protocol from scratch. They will not always fit, but their objects and resources show what is easy to forget: power, connectivity, firmware version, update status. The data model also has a legal dimension, which we return to under the Data Act: you must be able to describe the data a product generates and make it available to the user.
Why must OTA updates be designed before you choose a microcontroller?
Over-the-air (OTA) firmware updates are the only way to fix a bug or a vulnerability in a device sitting at a customer’s site. Without them, every fix means a truck roll or a device swap. Yet the update mechanism is often designed last, after the hardware is chosen, when it turns out there is no room for it.
RFC 9019, published in April 2021 and describing a firmware update architecture for IoT devices, sets out requirements that make a useful checklist. The firmware image must be authenticated and integrity-protected. The manifest describing the update must be signed, and the device must verify that signature before acting on it. The device must not be tricked into installing an older, validly signed version, because that version may contain known vulnerabilities. The bootloader must verify the image at every boot and have a recovery strategy if the new version fails to start, for example by keeping the previous image as a backup. And the device must not fail if a power cut or network loss interrupts the update.
Every one of these requirements has a hardware consequence. Keeping two firmware images needs roughly twice the flash of one. Verifying signatures needs compute or a hardware crypto accelerator. Storing keys safely needs a secure element or the right microcontroller features. If you pick the cheapest microcontroller that was enough for the pilot, you may discover that a secure update mechanism does not fit, and the per-unit saving on that component was trivial compared with the cost of replacing a fleet.
Then there is the operational side. Who signs releases, and where are the signing keys kept? What does a staged rollout look like: a few test devices first, then one percent of the fleet, then the rest? What happens to a device that was switched off for six months and wakes up four releases behind? How much does pushing an update to the whole fleet over cellular cost, and can you ship deltas instead of full images? Ask these questions before choosing hardware, and record the answers in the decision record.
Firmware development and testing is a separate subject. If you plan to outsource firmware, see how to commission firmware and device testing without losing control, including why signing keys and the test rig should stay on your side. Here we are concerned with the decision that comes before that engagement: whether the chosen hardware can update itself securely for the whole support period at all.
Who owns the device, its identity and its keys across the lifecycle?
Every networked device needs an identity: a unique identifier and the credentials it uses to authenticate to the cloud. A single fleet-wide key compiled into firmware is convenient in a pilot and disastrous in production, because a leak from one device opens access to all of them. Unique per-device credentials, though, have to be generated, injected and registered somewhere, and that process touches the factory, logistics and customer support.
The first question is where the device’s identity is created. In the factory, during programming on the production line? At an integrator? On first boot at the customer’s site? Each option has different security requirements and a different cost. Injecting certificates at the factory requires trust in the factory and a secure channel to it. Registration on first boot requires a mechanism that tells a genuine device from a clone.
The second question is rotation and revocation. Certificates expire, keys leak, devices get stolen. If a device certificate is valid for less than the product’s life, you must be able to renew it remotely. If a device is stolen, you must be able to cut it off without affecting the rest of the fleet. If the certificate authority that signed the whole fleet’s credentials expires and nobody remembered, every device can lose its connection on the same day.
The third question is change of ownership. A device sold to one company may be resold to another, moved to a different site or leased out. Who then has access to its historical data? How does the new user take over the device, and how does the previous one lose access? This has a legal dimension: the European Commission’s Data Act FAQ explicitly discusses subsequent users of a device and the possibility of erasing data before handover. If your account model assumes a device belongs to one customer forever, rebuilding it after rollout will be expensive.
These decisions rarely make it into a pilot, because a pilot has a handful of devices and one user. In production they form a process spanning several departments, which needs to be designed before the first device order. The output is a device lifecycle diagram, from manufacturing through registration, activation, change of owner and service to decommissioning, with an owner assigned to each step.
How do you plan a device’s end of life before you sell it?
The end of life of an IoT product does not depend only on hardware durability. A device also stops working when the cloud service it depends on is switched off, when a certificate expires, when an operator retires a network, or when the manufacturer goes out of business. Two well-known cases show what this looks like from the customer’s side.
In April 2022 Insteon, a home automation company, abruptly shut down its servers, leaving users without cloud and app features. Service only returned in October 2022, after a group of users bought the company. Spotify announced in May 2024 that it would end support for its Car Thing device, which stopped working on 9 December 2024; after a class action was filed, the company began offering refunds to customers who had bought directly from its website. In both cases the hardware was fine. It stopped working because the service it relied on disappeared.
For a company building an IoT product, the lessons are practical. First, define the support period up front and what its end means: will the device stop working, fall back to a local mode without cloud features, or can the customer move it to another service? Second, design an offline mode for functions whose loss would be critical to the customer. Third, plan a secure decommissioning process: erasing data on the device and in the cloud, revoking credentials, and collecting and recycling the hardware.
The support period is no longer purely a business decision. The Cyber Resilience Act requires the manufacturer to set a support period that reflects how long the product will be used, and to state its end date, with month and year, at the time of purchase. Outside the EU the direction is similar: the UK product security regime under the PSTI Act, in force since 29 April 2024, requires manufacturers of consumer connectable products to publish how long security updates will be provided, including an end date. What matters before the build is that the support period in your cost model matches the one you will be able to declare to customers and regulators. If the business model assumes three years and the device will realistically be used for eight, the mismatch will surface at the least convenient moment.
A useful exercise is to draft, before the build, the end-of-support notice for a product that does not exist yet. What will you tell customers? What happens to their devices and data? What will keeping that promise cost the company? If the answers are “we don’t know”, those decisions have not been made yet.
What has the RED delegated act, Regulation (EU) 2022/30, changed since 1 August 2025?
The Radio Equipment Directive 2014/53/EU (RED) has long set requirements for devices that transmit radio signals, covering among other things safety and electromagnetic compatibility. Its Article 3(3) also contains cybersecurity requirements, which did not apply until the Commission activated them. It did so with Delegated Regulation (EU) 2022/30, adopted in October 2021, which activates Article 3(3)(d), (e) and (f) for certain categories of radio equipment.
The three activated requirements cover network protection (the device must not harm the network or its functioning, or misuse network resources), protection of the user’s personal data and privacy, and protection from fraud for devices that enable the transfer of money or monetary value. The regulation covers, among others, radio equipment capable of communicating over the internet. In practice that includes a large share of IoT devices with Wi-Fi, Bluetooth or a cellular modem, but whether a specific product falls into a category must be checked against the regulation itself.
According to the European Commission’s Cyber Resilience Act FAQ (version 1.4, 4 September 2026), these requirements became applicable on 1 August 2025 and apply to in-scope radio equipment placed on the market from that date. Harmonised standards in the EN 18031 series have been cited in the Official Journal of the EU in support of them. Applying a harmonised standard gives a presumption of conformity with the requirements it covers, but check the scope of each standard before your conformity assessment.
A further change came in 2026. On 16 February 2026 the Commission adopted a delegated regulation repealing 2022/30, published in the Official Journal on 29 April 2026 as Delegated Regulation (EU) 2026/339. The repeal takes effect on 11 December 2027, the day the full Cyber Resilience Act requirements start to apply, to avoid two overlapping regimes. Radio equipment placed on the market between 1 August 2025 and 10 December 2027 is subject to the RED cybersecurity requirements; equipment placed on the market from 11 December 2027 is subject to the CRA.
What does this mean before the build? If your device ships in the EU before December 2027, it must meet the RED requirements, and if you keep selling it afterwards, later batches must meet the CRA. A design that assumes unique credentials, no default passwords, encrypted communication, personal data protection and a working update mechanism from the start will cross that line without rework. A design that defers these issues will face conformity assessment twice, each time with a different set of fixes. If you also sell in the United States, the EU rules do not carry over: plan separately for FCC equipment authorisation of the radio and check the state-level IoT security laws in the states where you sell.
What does the Cyber Resilience Act mean for design decisions before the build starts?
The Cyber Resilience Act, Regulation (EU) 2024/2847, entered into force on 10 December 2024 and covers products with digital elements placed on the EU market, including IoT devices. It has two key dates. From 11 September 2026 manufacturers must report actively exploited vulnerabilities and severe incidents, and from 11 December 2027 the remaining obligations apply, including the essential cybersecurity requirements in Annex I and vulnerability handling throughout the support period.
The reporting obligations are already in force as this article is published. The Commission describes the timelines as follows: an early warning within 24 hours of becoming aware, a full notification within 72 hours, and a final report no later than 14 days after a corrective measure is available for an actively exploited vulnerability, or within one month of the notification for a severe incident. Reports go through a single reporting platform run by ENISA. This applies to products already on the market too. If your company sells connected devices, your team must be able to detect an exploited vulnerability and report it within a day, which requires fleet monitoring and knowing the software components of every firmware version.
For design decisions, the most important element is the support period. Under Article 13(8) the manufacturer sets it to reflect the time the product is expected to be in use, and it must be at least five years unless the product is expected to be used for less. The Commission’s FAQ stresses that products used for longer, such as network equipment or systems used in industrial settings, should get correspondingly longer support. The end date of the support period, including month and year, must be clearly stated at the time of purchase.
What does this change before the build? First, the cost model must include vulnerability handling for the whole declared support period: monitoring, patches, testing, releases and update transmission. It is a fixed cost that does not shrink when sales slow down. Second, the choice of third-party components such as the device operating system, the communications stack and cryptographic libraries must account for their own support periods. The FAQ lists the support period of third-party components providing core functions as a factor manufacturers may take into account. If the supplier of a core library stops supporting it after three years and you have declared seven, the gap is yours. Third, the update mechanism from the earlier sections stops being good practice and becomes a condition for meeting legal obligations.
Fines for breaching the essential requirements and manufacturer obligations reach EUR 15 million or 2.5% of total worldwide annual turnover, whichever is higher. The point is not to frighten anyone with fines, but to get the order right: CRA requirements shape hardware selection, update architecture and the cost model, so they belong in the decision record before the build, not on a to-do list before certification.
How does the Data Act change the data architecture of connected products from 12 September 2026?
The Data Act, Regulation (EU) 2023/2854, has applied since 12 September 2025. For IoT manufacturers the key part is Chapter II, which gives users of connected products and related services a right to access the data generated by their use. A user here can be a consumer or a business that bought or leased the device, which matters for B2B products sold into the EU from anywhere, including the United States.
Two elements have applied since 12 September 2025. The first is a pre-contractual information duty (Article 3(2) and (3)): the seller, lessor or provider of a related service must tell the user, among other things, what type of data the product will generate, its volume and collection frequency, and how to access it. The second is the right of access to readily available data on request (Article 4), and the right to have it shared with a third party of the user’s choosing (Article 5). The Commission’s FAQ explains that since 12 September 2025 both products already on the market and new ones must allow users to access their data, with the manufacturer deciding whether access will be direct, indirect or a combination.
Since 12 September 2026, a few weeks ago, the design obligation in Article 3(1) has been added: connected products and related services placed on the market after that date must be designed and manufactured so that data is accessible to the user by default, easily, securely and free of charge, and, where relevant and technically feasible, directly. The Commission’s FAQ stresses that this does not mandate direct access in every case and leaves the manufacturer discretion, but where data is sent to a remote server, the data holder must provide appropriate means, such as an API, so that the user can easily access it.
For the architecture, this means four concrete decisions before the build. First, a data dictionary: the data the device and service generate, with format, frequency and storage location. Without it you cannot meet the information duty, and the data model from the earlier section becomes the basis for compliance. Second, an access mechanism: export, an API or direct access from the device, with user authentication. Third, an account and consent model that handles subsequent users of a device and data sharing with third parties the user nominates. Fourth, a separation between data and trade secrets: the Data Act provides safeguards for trade secrets, but you need to know up front which data and algorithms qualify.
There is also a business consequence. If the business model assumed that device data belongs exclusively to the manufacturer and is a source of competitive advantage, for example in aftermarket services, the Data Act changes that model: the user can share the data with a competing service provider. That is not a reason to abandon the product, but it is a reason to rethink where the value of the service lies before writing the business plan. The Data Act also covers matters not discussed here, such as switching between cloud providers; if your product depends on cloud services, review those rules separately.
What does the decision list look like before the first device order?
The table below pulls the decisions discussed above into a single working tool. The last column matters most: it shows what changing the decision costs once devices are with customers. Decisions with a high cost of change should be made and documented before the first production batch is ordered.
| Decision | Key question | Who decides | Output to record | Cost of changing after rollout |
|---|---|---|---|---|
| Connectivity technology | What range, power draw and transmission cost over the support period? Does the operator plan to retire it? | CTO, product, finance | Connectivity cost model per device and per fleet | Very high: modem or device replacement |
| Processing split | What runs on the device, the gateway, the cloud? | Architect, product | Data flow diagram with rationale | High: changes to firmware and back end |
| Data model | How are telemetry, state and events separated? How is the schema versioned? | Architect | Message specification with versioning | High: supporting several versions for years |
| Update mechanism | Can the hardware hold two images and verify signatures? Who signs releases? | CTO, embedded engineer | OTA design aligned with RFC 9019, memory budget | Very high: without OTA, every fix is a truck roll |
| Identity and keys | Where are credentials created? How are they renewed and revoked? | CTO, security, manufacturing | Provisioning and credential lifecycle process | High: recovering a fleet after a key leak |
| Support period | How long will the product be used? What are the support periods of core components? | Board, product, legal | Declared support period and its cost | High: commitments to customers and regulators |
| End of life | What happens to the device and its data when support ends? | Product, legal | Decommissioning plan and offline mode | Medium to high: depends on promises made |
| RED, CRA, US and UK rules | Which requirements apply, in which markets, from when? | Legal, CTO | Requirements map with dates and conformity plan | High: repeat conformity assessment |
| Data Act | What data does the product generate and how will users get it? | Product, legal, architect | Data dictionary and access mechanism | Medium: retrofitting an API and consent model |
| Data business model | Where is the value if users can share data with competitors? | Board, product | Value rationale in the business plan | Medium: changes the offer, not the code |
Fill in the table before talking to platform and hardware vendors, not after. That way the conversation is about whether a vendor’s offer meets your decisions, not about which decisions the vendor made for you. Empty cells in the “output to record” column mean those questions still need analysis or a prototype designed to settle them.
How do you choose an IoT consulting company and pay for its work?
Choosing a partner for the pre-build stage starts with asking what the partner sells besides consulting. A company that earns its money implementing one platform may know that platform well, but its recommendation will be skewed. A company that builds software across technologies has less stake in any particular choice, but may have a stake in a large build scope. There is no partner without interests. There is a partner whose interests you know and control in the contract.
When evaluating IoT consulting companies, a few questions quickly separate experience from slideware. How would you design an update for a device that can lose power mid-write? How would you cost connectivity for our fleet over seven years, and what assumptions would you use? How would you separate telemetry, state and events in the data model? How would you set a support period that satisfies the CRA, and which third-party components constrain it? How will users get their data under the Data Act? The answers need not be complete at the first meeting, but they should show the person has met these problems in practice, not just in a presentation.
The second issue is the team. The pre-build stage needs several skills at once: distributed systems architecture, embedded firmware, connectivity, security and regulation. One person rarely has all of them. Ask who exactly will work on your project and what experience each brings in each area. If nobody on the team understands firmware, hardware and update decisions may be made without knowing their consequences.
The third issue is how you pay. The decision stage suits a short engagement with a fixed scope and timeline and clearly defined deliverables: the decision record, cost model, regulatory map, risk register and prototype plan. Building an IoT product, with its uncertainty around hardware, connectivity and certification, rarely suits a fixed price for the whole thing, and forcing one moves the cost of risk into the quote. We explain this in more detail in our article on who really pays for risk transfer in fixed-price software.
At ARDURA the discovery phase for custom software development projects takes one to two weeks and defines scope and architecture. For an IoT product it makes sense to extend that phase with the items in the table above. If you are taking over an existing IoT product and want to assess its software first, an alternative is a technical due diligence audit; ARDURA’s Code Health Audit starts at EUR 4,500, takes one to two weeks and has a fixed scope.
When is pre-build IoT consulting not needed?
The position in this article has limits, and they are worth naming. Not every project with a sensor needs a multi-week decision stage. In some situations a quick prototype is the right first step and a decision record would be overkill.
The first is a small internal project: a few dozen mains-powered sensors in one plant, on the corporate network, not sold to customers. The cost of getting it wrong is low, because the devices are within reach and replacing them does not mean sending technicians across the country. RED and CRA obligations sit primarily with the manufacturer, so for off-the-shelf devices they fall on whoever made them. If you build your own radio device, even for internal use, check whether the rules treat you as the manufacturer.
The second is market research, where the question is “do customers want this service at all” rather than “how do we build it”. The prototype can then be even less technical: a few devices operated by hand, data copied into a spreadsheet, customer interviews. What matters is that everyone knows it is research, not the first version of the product, and that its code does not quietly become the foundation of production.
The third is a product built entirely on another manufacturer’s device, where you only build an application that consumes its data. Hardware, update and device compliance decisions then sit with that manufacturer, and your questions are about its support period, its data access interface, and what happens if it changes strategy.
In every other case, when you design your own device, sell it to customers or plan a fleet in the thousands, the pre-build decision stage costs little compared with the mistake it helps you avoid. For the technical foundations of IoT, such as system layers and protocol choices, see our IoT software development guide and the glossary entry on the Internet of Things (IoT).
What are the key takeaways before you commit to building an IoT product?
The most expensive mistakes in IoT projects are not made in code, but in decisions taken before anyone costed them. A pilot in which a sensor streams data to the cloud tests the easiest part of the project. The hard part, the cost and risk of keeping thousands of devices running for years, needs decisions on paper before the first batch goes into production.
Remember seven things. First, do not treat a successful pilot as permission to scale: build a prototype that tests the riskiest decisions, not just the sensor. Second, cost connectivity and updates for one device over its whole support period, then multiply by the fleet. Third, separate telemetry, state and events in the data model and version the schema from day one. Fourth, design secure updates before choosing a microcontroller. Fifth, decide where device identity is created and how it will be renewed, revoked and transferred. Sixth, decide the support period and end of life before you sell the first unit, allowing for the CRA’s five-year minimum. Seventh, map obligations under the RED delegated act, the CRA and the Data Act, plus US and UK rules where you sell, with dates, and treat them as design requirements rather than pre-certification tasks.
If you are preparing an IoT product and want to start with decisions rather than a platform, see how we work in custom software development, or book a call. Discovery takes one to two weeks and ends with a scope, an architecture and a decision list you can take to the board before ordering the first devices.