The AI tools market in 2026 is crowded with options, conflicting reviews, and marketing that exaggerates capability. The most useful signal is not what a tool promises but what practitioners who depend on it for client work actually use. This is the honest account of the AI tools we use at Softomate Solutions in our daily development work, client project delivery, and internal operations, alongside the tools we evaluated and stopped using and the specific reasons why.
We build AI systems for clients, which means we test tools more rigorously than most businesses do: not just whether a tool works in a demonstration, but whether it works reliably under production conditions, whether it is worth the cost at scale, and whether it creates more problems than it solves through integration complexity or inconsistent output quality.
GitHub Copilot is the AI tool that has most changed the day-to-day experience of software development on our team. We use it across PHP, JavaScript, Python, and SQL. Its primary value is autocomplete and boilerplate generation: it reduces the time spent on repetitive code patterns significantly. For a developer who knows what they want to build, Copilot handles much of the typing without requiring careful prompt engineering.
Its limitation is the same limitation all code generation tools share: it generates plausible code, not necessarily correct code. Our team treats Copilot output as a first draft that requires review, not as production-ready code. Developers who treat it as the latter introduce bugs at a higher rate than developers who treat it as accelerated typing. The productivity gain is real (our estimates suggest 20% to 30% reduction in time on standard implementation tasks) but only when review discipline is maintained.
We do not use Copilot for security-critical code or for complex architectural decisions. The risk of subtly wrong security implementations generated convincingly is too high. Security-critical code is written manually by the developer responsible for it.
When we need to review a complex function, debug a non-obvious issue, or think through an architectural decision, we use Claude 3.5 Sonnet. Its ability to hold a large amount of code in context and reason about it coherently makes it the most useful tool for these tasks. We paste a function, ask it to identify potential issues or to suggest a more efficient approach, and use its response as a structured second opinion.
We have found Claude more reliable than GPT-4o for the type of analytical reasoning that code review requires: identifying edge cases, reasoning about state across a multi-step function, and explaining why a particular implementation choice creates a specific type of risk. For code generation tasks (write me this function), the difference between Claude and GPT-4o is smaller. For code analysis tasks, Claude is consistently stronger in our practical experience.
Cursor is an IDE built around AI assistance, using GPT-4 and Claude as its underlying models. Its advantage over using Copilot in VS Code is its ability to make multi-file changes in response to a natural language instruction and to maintain context across a larger codebase. For feature development that touches multiple files and requires consistent changes across them, Cursor reduces the time spent on cross-file coordination significantly.
Its limitation: it is slower than VS Code with Copilot for individual file work, and it sometimes makes changes that are technically correct but contextually wrong (changing a file that was not intended to be changed). We use it selectively for larger refactoring tasks and new feature development, not as the primary day-to-day IDE.
When building LLM-powered features into client applications, we default to the OpenAI API with GPT-4o for most use cases. The primary reasons: the most mature API ecosystem with the widest third-party integration support, the most consistent uptime track record for production applications, and the enterprise agreement infrastructure that most of our UK clients need for GDPR compliance.
We use GPT-4o-mini for high-volume, lower-complexity tasks where cost is a significant consideration and the simpler queries do not require the full model's capability. The cost difference (roughly 50 times cheaper per token) justifies routing appropriate queries to the smaller model even when GPT-4o is available.
When a client project requires processing long documents, maintaining consistent behaviour across detailed system prompts, or producing well-structured written output at scale, we use the Anthropic Claude API. We have found Claude more reliable than GPT-4o for applications where the system prompt is complex and detailed, and where consistent adherence to instruction across thousands of interactions is critical.
The specific use cases we build on Claude: legal document analysis, financial report summarisation, compliance document processing, and any application where the client has a detailed brand voice or behavioural specification that the AI must maintain consistently.
For RAG implementations, we use Pinecone as the primary vector database in production. It is managed (no infrastructure to maintain), scales reliably, and has a clean API that integrates well with LangChain and custom retrieval implementations. We evaluated Weaviate and Qdrant and found Pinecone the best balance of managed convenience and production reliability for the project scale we typically work with.
For smaller projects where managed cost is a concern, we use pgvector (PostgreSQL extension for vector search) which runs within the client's existing database infrastructure with no additional service cost. The performance at smaller scale is comparable; the scaling ceiling is lower.
LangChain is the most widely used orchestration framework for LLM applications. We use it for standard RAG pipelines and simple agent implementations where its abstractions save development time. We do not use it for complex agent systems or production applications where the abstraction layer has introduced unexpected behaviour in the past.
Our honest assessment of LangChain: it is excellent for getting a working prototype up quickly and for standard use cases where its abstractions match the problem. It becomes a liability when the problem requires behaviour that the abstraction does not expose well, when debugging requires understanding what the framework is doing underneath, or when the abstraction layers add latency that matters at production scale. For these cases, we build on the raw API clients directly.
We use Notion for internal documentation and Notion AI for drafting and refining process documentation, meeting notes, and project briefs. Its primary value is reducing the friction of documentation: a developer who has just completed a piece of work can describe it verbally or in rough notes and use Notion AI to produce a clean, structured documentation entry. Documentation that previously went unwritten because it felt too time-consuming to format properly now gets written.
Every client meeting and internal planning session is recorded and transcribed by Otter.ai. The transcription accuracy on UK-accented English is high enough for practical use. We use the AI summary feature to produce a structured meeting summary within minutes of the meeting ending. This summary goes into the client's Notion workspace and serves as the source of truth for action items and decisions.
We stopped using meeting transcription tools that required installing a meeting bot that the client could see joining the call. Too many clients found it intrusive. Otter.ai's native integrations with Google Meet and Zoom handle transcription without a visible bot participant.
We trialled Jasper for content production assistance. The output quality for technical B2B content was consistently below what we produced with Claude or ChatGPT using well-crafted prompts. The additional structure Jasper provides over a raw LLM interface was not worth the premium cost for our use case. Teams producing high-volume commodity content may find more value; for technical, specialist content, a well-prompted general LLM produces better output at lower cost.
We evaluated AutoGPT during its early popularity for potential use in client agent deployments. In testing, the system's reliability degraded significantly on tasks involving more than five sequential steps, and the debugging experience when something went wrong mid-task was poor. We concluded that the abstraction over the agent loop provided by AutoGPT did not deliver enough value over building agent logic directly on the API to justify the loss of control and observability. We build agent logic directly and have not returned to AutoGPT or similar high-abstraction agent frameworks for production use.
We trialled seven or eight AI writing tools over 2023 and 2024. All of them produced output that was a thin wrapper over GPT-3.5 or GPT-4 with additional prompting and structure. None produced output quality better than Claude or GPT-4o accessed directly with a well-crafted system prompt. We consolidated to two tools rather than maintaining subscriptions to eight. The cost saving and the simplification of the toolchain were both significant.
The majority of AI tools are wrappers around a small number of underlying models: OpenAI's GPT family, Anthropic's Claude, Google's Gemini, and Meta's Llama. The value of a tool is in what it does to structure the interaction with those models: better prompting, relevant context, workflow integration, and a user interface suited to a specific task type. Very few tools have capability that genuinely exceeds what you can get from a well-designed direct API implementation.
The practical implication: choose tools based on workflow fit and integration quality, not on claims about unique AI capability. A tool that integrates well with your existing workflow and produces consistent, usable output at a price point that makes sense for your volume is the right tool, regardless of which underlying model it uses.
For a team of six to ten developers using the tools above: GitHub Copilot Business (£19 per user per month), Claude for Work team plan (£24 per user per month), OpenAI API usage (£300 to £1,500 per month depending on project volume), Pinecone (£70 per month for starter, scales with index size), Notion Team (£16 per user per month), Otter.ai Business (£20 per user per month). Total for a 10-person team: approximately £2,000 to £3,500 per month, depending on API usage volume.
Buy for internal productivity (writing assistance, code completion, meeting transcription). Build custom for client-facing applications where the AI capability is specific to the client's data and context. The investment in building a general-purpose AI writing tool or code assistant makes no sense when better ones already exist. The investment in building a client's bespoke AI system that no existing tool covers makes complete sense.
For more information about how we approach AI development for client projects, see our AI and Machine Learning Solutions service and our AI Projects page.
More from the blog
Let us help
Talk to our London-based team about how we can build the AI software, automation, or bespoke development tailored to your needs.
Softomate Solutions Limited
Deen Dayal Yadav
Online
