The AI Tools We Actually Use at Softomate Solutions (And the Ones We Stopped Using)

At Softomate Solutions we run a deliberately small AI stack: Claude for long-context analysis, strategy and code, ChatGPT for quick drafts and brainstorming, Perplexity for cited real-time research, GitHub Copilot inside our editors, Zapier and n8n for the automation layer, Canva for fast design, and per-client AI agents we build ourselves. A focused stack like ours sits within the £50 to £100 per seat per month range that typically delivers ROI for UK SMEs. We dropped Jasper once the general-purpose models we already used covered the same job for us, abandoned single all-in-one "AI marketing suites" that hallucinated in our tests, and keep a human review on anything client-facing, whichever tool drafted it. The honest rule we follow: match the tool to the function, never the hype, and never let AI touch legal, financial or contractual output without a human signing it off. This article is the full account, including the tools we stopped paying for and exactly why.

Last updated: June 2026

• What Is the Core AI Stack Softomate Actually Uses Every Day?
• Claude or ChatGPT: Which One Do We Use for Which Job?
• Which AI Tools Do We Trust for Research and Fact-Checking?
• What Sits in Our Automation and Integration Layer?
• Which AI Tools Did We Stop Using, and Why?
• What Does an AI Stack Actually Cost, and Is the ROI Real?
• When Do We Deliberately Not Use AI?
• What Does the Softomate AI Implementation Process Look Like?
• Frequently Asked Questions

What Is the Core AI Stack Softomate Actually Uses Every Day?

Our core AI stack is six tools, not sixteen. We use Claude as the workhorse for analysis, strategy and code; ChatGPT for fast first drafts; Perplexity for cited research; GitHub Copilot for in-editor code completion; Zapier and n8n for connecting everything together; and Canva for quick visual assets. Around that core we build bespoke AI agents for individual clients, but those are products we deliver, not subscriptions we maintain for ourselves. The discipline matters. Most agencies we talk to are paying for a dozen overlapping tools, half of which do the same job badly, and nobody can tell you the actual monthly bill.

We landed on this stack the hard way, by trialling far more than we kept. The test we apply to any tool is simple: does it save measurable time on real work, does it do one thing genuinely well, and does it survive a price rise without becoming a rip-off? If the answer to any of those is no, it gets cut at the next billing cycle. The result is a stack a small team can actually master rather than a graveyard of half-learned dashboards.

Here is the honest breakdown of what each tool does, who uses it, and how it is billed in 2026. Prices change often, so check each vendor's current pricing before you budget.

Our view, stated plainly: for our work, a stack like this has served us better than the single "do everything" platforms we have tried. The all-in-one suites promise to replace this whole table with one login, and every time we have tested that promise it has fallen apart on the work that actually matters. Specialised tools that each do one job well, glued together with automation, win. That is the principle behind every AI automation project we deliver for clients, and we eat our own cooking.

Claude or ChatGPT: Which One Do We Use for Which Job?

We use Claude for anything that needs long context, careful reasoning or code, and ChatGPT for anything fast and disposable. That split is not tribal loyalty, it is the result of running both on the same jobs for two years and seeing which suits which job for our team. Claude holds a long document in its head without losing the thread, which matters when we are reviewing a 60-page client brief, refactoring a large codebase, or drafting a technical proposal that has to stay internally consistent across 4,000 words. ChatGPT gets us to a usable first sentence quickly and is handy for riffing through twenty rough ideas when we do not care about polish yet.

The practical difference shows up most in development work. When a developer pastes an entire service class and asks for a security review, Claude works through the whole file and reasons about the edge cases. When the same developer just wants a regex or a quick boilerplate function, either tool is fine and ChatGPT is often faster to the answer. We are not precious about it. The job picks the tool.

Here is how we actually allocate work between the two:

1. Long-context analysis (client briefs, audits, large code files, multi-document strategy): Claude, every time.
2. Technical writing that must stay consistent (proposals, documentation, this article): Claude for the draft, because it holds the structure.
3. Fast brainstorming and idea volume: ChatGPT, because we want quantity then we curate.
4. Throwaway copy and reformatting (turn these bullet points into an email): either, usually whichever tab is open.
5. Code review and refactoring of real files: Claude, for the reasoning depth.
6. Image generation and quick visual concepts: ChatGPT, for the built-in image model.

The honest stance: most of the work that pays our bills is analysis and code rather than throwaway copy, so Claude carries the bulk of it for us. ChatGPT earns its place for speed and image generation, so we keep both. Anyone telling you there is a single objectively best chatbot is selling something. There is only the right chatbot for a specific job, and the skill is knowing which is which. That same logic informs the AI agents inside the chatbots we build for clients, where we pick the underlying model per use case rather than defaulting to whatever is fashionable.

Which AI Tools Do We Trust for Research and Fact-Checking?

For research we trust Perplexity, because it cites its sources and pulls real-time data, and we trust nothing without checking those citations ourselves. This is the single most important discipline in our entire AI workflow. General chatbots will state a wrong figure with total confidence, and that confidence is the dangerous part. Perplexity is built differently: it searches the live web, summarises, and links every claim to a source you can click. That does not make it infallible, but it makes its errors catchable, which is the whole game.

We use Perplexity for competitor research, market sizing, regulatory checks, current pricing of third-party tools, and any factual claim that will end up in client-facing work. When a strategist needs to know the current VAT registration threshold, the latest Companies House filing requirement, or whether a competitor has launched a feature, Perplexity gets us a cited answer in seconds. The cited part is non-negotiable. A statistic without a source is a rumour, and we do not put rumours in client decks.

Our hard rule on AI research has three steps, and we never skip them:

• Get the answer with a citation. If the tool cannot show its source, we treat the answer as a starting hypothesis, not a fact.
• Open the source. We click through and confirm the original actually says what the AI claims it says. AI tools routinely paraphrase a source into something it never stated.
• Check the date. Pricing, regulations and statistics rot fast. A 2023 figure quoted in 2026 is often simply wrong, and AI tools love quoting stale data as current.

Where does this matter most in pounds? When research feeds a quote, a strategy or a build estimate, a single wrong number can cost a project. We have seen agencies size a market off a hallucinated statistic and pitch an entire campaign on a number that never existed. The cost of fact-checking is two minutes. The cost of not fact-checking can be a lost client. That asymmetry decides it.

Be sceptical if any AI tool gives you a precise figure with no link. Precision without a source is the classic hallucination signature: the model invents a number that sounds authoritative because authoritative-sounding numbers are what it was trained to produce. We have trained the whole team to feel suspicious of clean round confidence. It is usually the tell.

What Sits in Our Automation and Integration Layer?

Our automation layer is Zapier for quick, off-the-shelf connections and n8n, self-hosted, for anything that touches client data or needs to run at scale. This layer is where AI stops being a chat window and starts being infrastructure. A chatbot that answers a question is mildly useful. A chatbot that answers a question, logs the lead in the CRM, books the callback in the calendar, and notifies the right person on Slack, all without a human touching it, is a system that earns money while everyone sleeps. The automation layer is what turns models into that.

Zapier is useful for fast wins. Its entry-level plans are enough to wire together a contact form, an email tool and a spreadsheet in an afternoon. For a lot of small jobs that is all you need, and we are not too proud to use the simple tool when the simple tool fits. Zapier bills by usage, so at high volume the running cost needs watching, and as a rule we are cautious about routing sensitive UK client data through any third-party cloud we do not control.

That is why our heavier automations run on n8n, self-hosted on infrastructure we manage. Self-hosting means client data stays on servers we control, which matters enormously under UK GDPR, and the cost is the server rather than a per-task metre. For a client moving thousands of records a month, that distinction is the difference between a sensible bill and a runaway one. This is the backbone of every business process automation system we deliver, and it is exactly how we build the AI agents that handle support, qualification and routing for clients running on GoHighLevel automation.

1. Trigger. Something happens: a form submission, a new lead, an inbound call, a calendar event.
2. Enrich. An AI step reads the input, classifies it, extracts the key fields, and decides what to do.
3. Act. The workflow writes to the CRM, sends the email, books the slot, or escalates to a human.
4. Notify. The right person is told, with context, so nothing falls through a gap.
5. Log. Everything is recorded for audit, because UK businesses need a trail.

Our honest view on the automation layer: most businesses underinvest here and overinvest in the shiny chat tools. The model is the engine, but the automation is the gearbox and the wheels. A brilliant model wired to nothing produces a clever conversation and zero outcomes. The leverage is in the plumbing, which is unglamorous and exactly why it is undervalued. When we tell a client where their first AI pound should go, the answer is almost always a well-built workflow, not another subscription.

Which AI Tools Did We Stop Using, and Why?

We stopped using Jasper once the general-purpose models we already paid for covered the same job, abandoned the "all-in-one AI marketing suites" we trialled because they hallucinated in our tests and made our data hard to export, and standardised on one default tool for client-facing output. The tools we drop teach us more than the tools we keep, because a cut is always triggered by a specific failure, and that failure is a lesson. Here is the candid list.

Jasper was the big one. We used it early for marketing copy. Over the same period the general-purpose models we were already paying for improved to the point where they covered the copy work our team needed, so a separate specialist subscription stopped making sense for how we work. That is a decision about our own stack and usage, not a verdict on the product, and your needs may differ. When two subscriptions do the same job for your team, you keep one. We did.

The all-in-one suites were a different failure. These platforms promise to replace your entire stack with one login: writing, images, automation, chat, the lot. Every one we tested was mediocre at all of it and good at none of it. Worse, they hallucinated freely and made exporting your own data awkward, which is the classic lock-in play. We have a name for the pitch: AI-washing, where a thin wrapper around someone else's model is sold as a revolutionary platform. Be sceptical of any tool that claims to do everything. Doing everything is how you end up doing nothing well.

ChatGPT we did not drop, but we changed its role. It stays in the stack for drafts and brainstorming, but we removed it from our client-facing pipeline. The reason was consistency on our side: for copy that goes to a client under our name, we want one default tool, one set of prompts and one review process, so results are predictable across the team. Claude became the default for anything that ships, and every piece still goes through human review.

The pattern across every cut is the same: a tool gets dropped when its price stops matching its value, when it does many things instead of one thing well, or when its output cannot be trusted without redoing the work. Our honest rule is that every subscription has to justify itself at every renewal, and "we have always paid for it" is not a justification. Audit your stack quarterly. You will be surprised how much you are paying for tools nobody opens.

What Does an AI Stack Actually Cost, and Is the ROI Real?

A focused AI stack for a small UK business costs roughly £50 to £100 per user per month, and the ROI is real when, and only when, the tools are tied to actual workflows. Our own core stack is Claude, ChatGPT, Perplexity, Copilot, the shared automation infrastructure and Canva. That is a coffee-a-day budget for tools that, measured honestly, save our team hours every week. The trap is not the price of any one tool. The trap is owning twelve overlapping tools and using three.

The UK data backs the case for sensible adoption. Surveys in 2026 put SME AI adoption around 54%, with roughly 60% of medium-sized firms using AI against about 36% of micro-businesses, so the smallest firms are the furthest behind and have the most to gain. The productivity figures are striking: firms actively deploying AI report large net productivity expectations while non-adopters expect to go backwards, and reported SME productivity gains span a wide range depending on how well the tools are applied. Crucially, the same research shows the vast majority of AI-using SMEs report no reduction in workforce size: the tools augment people rather than replace them. That matches what we see. AI does not fire your team, it lets a small team do the work of a bigger one.

Here is a realistic monthly budget for a small UK firm building a sensible stack from scratch:

To work out whether that spend pays back, we use a blunt calculation, and we recommend every client run it:

1. Measure hours saved. Pick three tasks AI now does faster and time them honestly, before and after.
2. Cost the hours. Multiply hours saved per month by the loaded cost of the person doing them.
3. Subtract the stack cost. If hours-saved value comfortably exceeds the subscription bill, the stack pays for itself.
4. Re-measure quarterly. Tools and prices change. So does your usage. Audit it.

Our stance on ROI is unromantic: the spend is trivial, so the only real risk is buying tools you never wire into a workflow. A stack under £100 per seat that saves one hour a week of a £40-an-hour person has paid for itself by month-end. The failure mode is never the price, it is the shelf-ware. If you are paying for an AI tool nobody has opened in a fortnight, that is your real cost, and it is hidden on the invoice as a line you have stopped reading.

When Do We Deliberately Not Use AI?

We deliberately do not use AI as the final word on anything legal, financial, contractual, medical or client-facing without a qualified human reviewing it first. This is not caution for its own sake, it is the single boundary that keeps AI a tool rather than a liability. AI is excellent at drafting, summarising and accelerating. It is unreliable as the last line of defence on anything where being confidently wrong has consequences, and in law, finance and compliance, confidently wrong is the most expensive sentence in business.

The reasoning is about accountability. When an AI tool produces a contract clause, a tax figure or a compliance statement, the model carries no liability for getting it wrong. You do. So the human review is not a nicety, it is where the responsibility lives. We let AI draft a contract summary in seconds, then a person who can be held accountable checks it before it goes anywhere near a client. The speed-up is enormous and the safety net is intact. Remove the net and the speed becomes a hazard.

There is a data-protection dimension too, and for UK businesses it is sharp. Putting client data, personal data, or anything covered by UK GDPR into a public AI tool can mean handing it to a third party in a way you cannot control or retrieve. That is why our heavier client automations run on self-hosted infrastructure and why we never paste sensitive client records into a public chatbot. Where client data is involved, the default is on-our-servers, not in-someone-else's-cloud. The Information Commissioner's Office has clear guidance on this, and we treat data residency as a hard constraint on the CRM systems and automations we build.

• Legal text (contracts, terms, clauses): AI drafts, a qualified human approves. Never the reverse.
• Financial figures (quotes, tax, forecasts): AI assists the working, a human verifies every number.
• Compliance and regulatory claims: checked against the official source, never trusted from a model alone.
• Anything with personal data: kept off public AI tools, run on controlled infrastructure under UK GDPR.
• Final client-facing copy: AI drafts, a human edits and signs it off under our name.

The honest rule we live by: AI does the first 80% in a fraction of the time, and a human owns the last 20% that carries all the risk. Anyone selling fully autonomous, no-human AI for legal or financial work is selling you their liability dressed up as your efficiency. Where AI still fails us is exactly here, on the judgement calls, and pretending otherwise is how businesses get burned.

What Does the Softomate AI Implementation Process Look Like?

The Softomate AI implementation process is five stages, runs to a fixed quote with no hourly surprises, and typically takes four to ten weeks from first call to a live, working system. We do not sell you a pile of subscriptions and wish you luck. We work out which functions in your business genuinely benefit from AI, build the tools and automations around those functions, train your team on the small stack that remains, and stay on to support it. The principle from this whole article, match the tool to the function, is exactly how we scope your project.

We quote fixed prices because UK business owners deserve to know the cost before they commit, not after. Discovery is where we map your workflows and identify the high-leverage automations, so the build is targeted rather than speculative. Most clients are surprised how few tools they actually need once we have stripped out the overlap.

On pricing, here is the honest range. A simple single-workflow automation starts from £1,500, and a controlled AI agent from £6,500. A broader build with multiple integrated agents, CRM connections and a self-hosted automation layer is scoped and quoted on complexity. Ongoing support and optimisation is available on a fixed monthly retainer rather than an open-ended hourly metre. Every quote is fixed before we start, so you are never billed for surprises.

Our promise is the same discipline we apply to our own stack: we will not sell you twelve tools when three will do, and we will not build an autonomous system to touch anything that needs a human signature. If your business is exploring AI, the most valuable first conversation is usually a free discovery call where we tell you honestly what is worth automating and what is not. That candour is the point. You can talk to us through our AI automation agency or our software development service, and the first call costs nothing.

Frequently Asked Questions

The Softomate stack is deliberately small: Claude for analysis and code, ChatGPT for fast drafts, Perplexity for cited research, Copilot in the editor, Zapier and n8n for automation, and Canva for design. We dropped Jasper once general-purpose models covered the same job for us, walked away from all-in-one suites that hallucinated in our tests, and standardised on one default tool for client-facing work. The principles that survived every cut are simple: match the tool to the function not the hype, fact-check every AI claim against its source, keep sensitive UK data off public tools, and never let AI have the final word on legal or financial output. A focused stack tied to real workflows pays for itself in hours saved within the first month. The opportunity for UK SMEs is real and the cost of entry is trivial. The only genuine risk left is buying tools you never actually use.

If you want a stack built around your business rather than the hype, talk to our team about a fixed-quote AI build through our business process automation service in London, or book a free discovery call via our contact page.

Written by Deen Dayal Yadav, Founder of Softomate Solutions, a London-based AI automation and software development agency in Stanmore (HA7). With over 12 years building software and automation systems for UK businesses, Deen leads a team that builds AI agents, custom CRMs and process automation for SMEs across London and the UK. Softomate Solutions is registered at Companies House and works with clients who want practical, accountable AI rather than expensive shelf-ware. Learn more about Softomate Solutions and how we approach AI honestly.

• Department for Business and Trade: UK SME productivity and AI adoption guidance
• Information Commissioner's Office: guidance on AI and data protection under UK GDPR
• Department for Science, Innovation and Technology: UK AI policy and adoption
• British Chambers of Commerce: UK SME technology adoption research