AI for Technology and Telecom
Buyers who will read your architecture before they read your case study. Ten verticals, written for people who can evaluate the engineering — and who will notice if the claims do not survive contact with it.
You are the only sector that might be a provider.
Everywhere else in our industry list, an organisation deploys AI. In technology you may also be the one placing it on the market, which is an entirely different position under the EU AI Act — different obligations, different documentation, different liability. And the line is easier to cross than product teams expect: fine-tuning someone else's model on your own data, or putting your name on a system built from a third-party model, can move you from deployer to provider without anyone deciding to.
That question should be settled before the roadmap rather than during a customer's security review. It changes what you must document, what you must disclose, what your customers can demand of you, and what you can honestly promise them. Answer it in week one and the rest is ordinary engineering. Answer it during a procurement process and you will be answering it under time pressure with a deal attached.
Where AI actually earns its place in a technology business
Ranked by evidence rather than by launch-post frequency. The maturity column is our own read; the catch column is the part that does not make the announcement.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Support and service automation | Deflects, drafts and routes customer support contacts. | Proven | Genuinely reliable. Deflection rate is the vanity metric; resolution rate and escalation quality are the real ones. |
| Documentation and content operations | Generates and maintains technical documentation, release notes and knowledge bases. | Proven | Works well when grounded in the codebase and specifications. Ungrounded, it produces confident documentation of features that do not exist. |
| Network and infrastructure optimisation | Predicts load, routes traffic, allocates capacity. | Proven | Mature in telecoms and cloud. The gains are real and bounded by physical capacity, which no model creates. |
| Security detection and triage | Prioritises alerts, correlates events, drafts investigation context. | Strong | Alert triage is the clear win. Autonomous response is where organisations get hurt, and adversaries adapt deliberately. |
| In-product AI features | AI capability shipped inside your own product. | Strong | Where most technology companies should focus — and where the provider question, evaluation discipline and unit economics all land at once. |
| Developer assistance | Code generation, review support, test writing. | Mixed | Large gains on routine work, much smaller on complex engineering, with measurable increases in review time and defect rates. |
| Sales and go-to-market automation | Scores leads, drafts outreach, summarises calls. | Mixed | Call summarisation and CRM hygiene are reliable. Automated outreach at scale is mostly a deliverability problem waiting to happen. |
| Autonomous operations | Agents that remediate infrastructure or security incidents without a person. | Early | Attractive and premature for most. The failure modes are correlated with the moments you least want a surprise. |
The developer productivity numbers, stated honestly
A randomised controlled trial found experienced open-source developers were roughly 19 per cent slower with AI tools in early 2025 while believing they were 20 per cent faster — and the same group's follow-up estimated around 18 per cent faster by early 2026, a swing attributed largely to learning and tooling. A study across 4,500 developers found routine task time down 46 per cent and complex task gains under 10 per cent. Separately, AI-authored pull requests have been observed taking substantially longer to review, and AI-generated code has been measured with security vulnerabilities at several times the human rate. All of that is consistent: large gains on boilerplate, small gains on hard problems, real costs downstream, and a perception gap that makes self-reported productivity useless as evidence. Measure your own throughput to merged, reviewed, working code — not to first draft.
The sector that has to read the AI Act twice
Technology companies sit on both sides of most AI regulation: you deploy it internally and you may place it on the market. The obligations differ substantially. This is our reading as at September 2026 and we work alongside your legal and compliance functions rather than in place of them.
The question that changes everything
A deployer uses an AI system. A provider places one on the market or puts it into service under their own name — and providers of high-risk systems carry the conformity assessment, technical documentation, quality management and post-market monitoring obligations. Putting your name on a system built from someone else's model can make you the provider. So can fine-tuning a third-party model on your own data. Settle this per product before the roadmap, not during a customer's procurement review.
In force since 2 August 2026
Article 50 applies now: people must be told when they are interacting with an AI system unless it is obvious, synthetic audio, image, video and text content must be marked in a machine-readable way, and deepfakes must be disclosed. These are product requirements with engineering consequences — provenance marking in particular is something to build in rather than to add when a customer asks.
Obligations since August 2025, enforcement from August 2026
Providers of general-purpose models carry transparency documentation, training data summaries and copyright compliance obligations, with additional systemic-risk duties above a compute threshold of 10^25 floating point operations. The Code of Practice offers a presumption of compliance to signatories; declining it means demonstrating equivalent measures. Commission enforcement powers over GPAI activated on 2 August 2026.
Already enforceable, and they catch product features
Since February 2025: social scoring, untargeted facial image scraping, emotion inference in workplace and education contexts, certain biometric categorisation, and manipulative or vulnerability-exploiting techniques. Penalties reach €35 million or 7 per cent of global turnover. Several of these describe features that have shipped in consumer and enterprise products, so they are worth checking against your backlog rather than assuming they apply to someone else.
Reporting now, full obligations December 2027
For products with digital elements — which includes most commercial software — vulnerability reporting obligations began on 11 September 2026 and the main cybersecurity requirements apply from 11 December 2027. Model files, weights and configuration are part of the attack surface and part of the compliance surface.
500 kW and above
Data centres with at least 500 kW of installed IT power demand report energy performance indicators to a European database under the Energy Efficiency Directive, with a common rating scheme now operational. For anyone running significant training or inference capacity in the EU this is a reporting obligation attached to a cost line that AI workloads are increasing.
What this means for a build
Three things are architectural. Establish per product whether you are a provider or a deployer, because the documentation burden differs by an order of magnitude and cannot be reconstructed later. Build content provenance marking into anything that generates media, because Article 50 already requires it and retrofitting marking across a content pipeline is unpleasant. And treat model files, weights and configuration as part of your security posture, because the Cyber Resilience Act does. See EU AI Act compliance readiness and LLM security.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
SaaS & software
In-product AI features with evaluation and unit economics — plus the provider question settled per product.
Read more →Startups
Defensibility, evaluation before scale, and what diligence will ask — including when not to build.
Read more →IT services & system integrators
Estimation, knowledge reuse and proposal production — plus building capability you can genuinely sell.
Read more →Telecommunications
Network optimisation, fault root cause, churn and field operations — at the scale telecoms actually runs.
Read more →Cybersecurity
Alert triage, investigation support and detection engineering — with autonomous response held back deliberately.
Read more →Data centres & cloud
Cooling and energy optimisation, GPU scheduling and capacity — plus EU energy reporting from 500 kW.
Read more →Gaming
Player modelling, matchmaking, moderation and live-ops — with generated content and disclosure handled honestly.
Read more →Developer tools
Evaluation for code-facing features, benchmark honesty and unit economics — for buyers who will check.
Read more →Hardware & devices
On-device inference, fleet telemetry and OTA model updates — with CRA and Data Act obligations designed in.
Read more →Managed service providers
Ticket triage, alert noise reduction and per-client margin — with client segregation enforced architecturally.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Are we a provider or a deployer under the AI Act?
It depends on what you place on the market under your own name. Using a third-party model inside your internal tooling makes you a deployer. Shipping a product that incorporates it, under your brand, generally makes you a provider — and fine-tuning someone else's model on your own data can do the same. Providers of high-risk systems carry conformity assessment, technical documentation, quality management and post-market monitoring obligations. Settle it per product in week one, because it is the question that determines everything else.
Do AI coding assistants actually make our team faster?
On routine work substantially; on complex engineering much less; and the honest answer for your team requires measuring your own throughput. The evidence is genuinely mixed and moving: a randomised trial found experienced developers 19 per cent slower in early 2025 and around 18 per cent faster a year later, while a large enterprise study found 46 per cent time savings on routine tasks and under 10 per cent on complex ones. Self-reported productivity is not usable evidence — developers in that trial believed they were 20 per cent faster while being slower. Measure time to reviewed, merged, working code.
What does Article 50 require of our product?
That people are told when they are interacting with an AI system unless it is obvious, that synthetic audio, image, video and text is marked machine-readably, and that deepfakes are disclosed. It has applied since 2 August 2026. The provenance marking requirement is the one worth acting on first, because building it into a generation pipeline is straightforward and retrofitting it across an existing content system is not.
Should we build AI features into our product?
Usually yes, and the discipline that separates the ones that work is evaluation. Ship an AI feature without an evaluation suite and you cannot tell whether a model update improved it or broke it, which means you cannot safely change anything. Build the evaluation harness before the feature, measure quality against it on every change, and track cost per request from the first week — because unit economics is where in-product AI features most often fail commercially rather than technically.
What is the most common way technology companies get this wrong?
They ship an AI feature with no evaluation, no cost tracking and no version pinning, and then discover that their provider changed the model, quality moved, and nobody can say by how much or in which direction. That is not a modelling failure; it is a missing engineering practice, and it is the single most common thing we are called in to fix.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.