---
title: "The £495 AI agent vs the £75,000 AI build: when each one is right (and when both are wrong)"
description: "UK pricing data shows AI projects ranging from £10/month SaaS tools to £100,000+ full custom machine-learning builds. Most SMBs pay the wrong price for their actual problem. Here's the five-tier cost framework, what each tier is built for, and the three failure modes that hit owners who buy the tool when they needed a process redesign or pay enterprise prices for what was always a SaaS-shaped problem."
canonical: https://richardbatt.com/blog/495-vs-75000-ai-cost-decision-uk
date: 2026-05-05
author: Richard Batt
tags: [AI Cost, AI Implementation, UK Business, Decision Framework]
type: blog_post
---

# The £495 AI agent vs the £75,000 AI build: when each one is right (and when both are wrong)

_UK pricing data shows AI projects ranging from £10/month SaaS tools to £100,000+ full custom machine-learning builds. Most SMBs pay the wrong price for their actual problem. Here's the five-tier cost framework, what each tier is built for, and the three failure modes that hit owners who buy the tool when they needed a process redesign or pay enterprise prices for what was always a SaaS-shaped problem._

**Richard Batt** — AI implementation specialist. 120+ projects across 15+ industries, serving SMBs (5-200 employees) worldwide from Middlesbrough, UK (working globally). Contact: richard@richardbatt.com · https://richardbatt.com

A UK accountancy firm I spoke with last month had two open quotes on the desk for the same problem. The problem was invoice extraction. The first quote was a £495-a-month pre-built AI agent from a firm in Reading. The second was a £75,000 fixed-fee build from a London consultancy. The brief was identical, the source documents were the same, and the desired outcome was the same. Both quotes came with confidence and case studies.

The accountancy partner asked me which one to pick. But the honest answer was neither, because the firm had not done the upstream work that determines which tier is the right tier. After 120+ AI projects across 15+ industries, that has been the story almost every time an SMB shows me two quotes that differ by a factor of 150.

This is the five-tier cost framework I use to sort the question, the kind of work each tier is genuinely built for, and the three failure modes that drain SMB budgets every quarter.

**The short version**

- Five real cost tiers, from £10/month SaaS to £75,000+ full custom machine-learning.
- Each tier is built for a specific shape of problem. Mismatches are expensive.
- The three failure modes: paying enterprise prices for SMB problems, paying SaaS prices for enterprise problems, and buying any AI tool when you needed a process redesign.
- Most UK SMBs in the 5 to 200 employee range live in tiers 2 and 3. Anything else is usually wrong.
- The £495 quote and the £75,000 quote can both solve the same surface problem. They cannot both solve the same underlying problem.

## The five tiers

Tiers are based on real UK pricing data: media-pack research from agencies like mediaffy.com and gigcmo.com, off-the-shelf AI agent platforms like automation-ai.uk (£495 entry), Microsoft's Copilot Studio published pricing (£23.10 per user per month), and the bracket I see in 80% of the SMB quotes I review.

| Tier | Price band | What you get | Right when... | Wrong when... |
| --- | --- | --- | --- | --- |
| 1. SaaS AI tools | £10 to £50 / month | Off-the-shelf product features (ChatGPT, Notion AI, GitHub Copilot, Zapier AI Steps) | One person needs a productivity boost on a known task | The work is shared across a team and needs governance |
| 2. Pre-built AI agents | £495 to £2,000 / month | Vendor-built agent templated for a vertical (invoice bots, support bots, scheduling agents) | The workflow is bounded, the documentation is clean, and a templated agent fits | Your data, edge cases, or compliance need can't be templated |
| 3. Small custom builds | £2,000 to £6,000 one-off | A specialist wires existing tools (n8n, Make, Zapier, custom GPT) to your stack | The workflow is yours and isn't templated, but the components exist | The problem requires model training, novel data pipelines, or live retraining |
| 4. Mid-range builds | £15,000 to £75,000 | Custom-built build by an agency: integrations, hosting, light model fine-tuning | A team-wide workflow needs custom integration with messy systems | The team isn't ready to redesign work around it |
| 5. full-stack ML and enterprise | £75,000 to £100,000+ | Data science work, model training and MLOps, plus an ongoing retraining contract | You have a defensible data asset and a regulated or differentiated outcome | You're using it to solve a problem that ChatGPT and a clean SOP would handle |

Most UK SMBs I work with in the 5 to 200 employee range need tier 2 or tier 3. Tier 1 is for individual productivity. Tier 4 is rare. Tier 5 is genuinely scarce in the SMB market and almost always over-prescribed.

## What each tier is genuinely built for

### Tier 1: SaaS AI tools (£10 to £50 / month)

ChatGPT Team at £20 per user per month. Microsoft Copilot at £23.10 per user, Notion AI at £8 per user, and the Zapier Tables AI step on top. These are built to make one person more productive on a task they already do.

Right call when: a salesperson wants to draft cold-email follow-ups faster, the head of operations wants help summarising long PDFs, the marketing manager wants to brainstorm campaign concepts. The cost is low, the upside is per-person, and the failure mode is small.

Wrong call when: you need three people to use the same workflow and produce comparable output. SaaS AI tools are personal productivity. The output drifts across users because there's no shared prompt, no shared standard, no audit trail. If you need consistency, you need at least tier 2.

### Tier 2: Pre-built AI agents (£495 to £2,000 / month)

This tier has exploded in the last 12 months. UK platforms like automation-ai.uk start at £495 a month for vertical-specific agents: invoice extraction, lead qualification, customer support tier one, appointment scheduling, content drafting. The agent is templated for a vertical and configured for your specific tenant.

Right call when: your workflow is bounded (you know the inputs, you know the outputs, you know the edge cases), your documentation is clean enough for the agent to ground itself in, and the vertical match is good. A 20-person UK accountancy firm running invoice extraction is a textbook tier 2 buyer. The agent is built for that shape, the cost is small relative to the saving, and time-to-value is days not months.

Wrong call when: your data has unusual structure, your compliance constraints are tight (regulated advice or clinical or safeguarding work), or your edge cases sit outside the templated pattern. The £495 agent will work for the 70% of cases that fit the template and fail visibly on the 30% that don't. If the 30% includes the moments that matter to your customers, you've bought the wrong thing.

### Tier 3: Small custom builds (£2,000 to £6,000 one-off)

A specialist wires together n8n, Make, Zapier, a Custom GPT, and your existing stack. They write the prompts, build the routing rules, set up the monitoring, and hand it over with documentation. Ongoing cost is the SaaS licences plus a £200 to £400 a month support retainer if you want one.

Right call when: the workflow is specific to your business, the building blocks exist, and you need integration with your CRM, accounting software, or document store. A 30-person professional services firm wanting AI-assisted proposal writing tied to their Pipedrive instance is a textbook tier 3 buyer. The components exist. The work is wiring.

Wrong call when: the problem requires training a model on your private data, handling live retraining as the data drifts, or producing outputs that need defensible accuracy under audit. You're at the edge of what tier 3 can deliver. Pushing through to tier 4 or 5 is sometimes correct.

### Tier 4: Mid-range builds (£15,000 to £75,000)

A specialist agency builds you a custom workflow: hosted infrastructure, custom integrations with on-premises systems, light model fine-tuning, often a small team for 8 to 12 weeks. The consultancies live in this tier.

Right call when: a team-wide workflow needs custom integration with systems that don't have public APIs, the volume is high enough to justify the build cost (typically 100+ hours a week of human time at risk), and you have an internal owner who can take on the ongoing operation.

Wrong call when: the team isn't ready to redesign work around the tool. I have watched four tier-4 builds in the last year ship on time, on spec, and on budget, and then sit unused because the manager who should have redesigned the workflow around the new capability didn't. The tool is fine. The change management isn't there. Tier 4 without managerial conviction is the most expensive way to learn a free lesson.

### Tier 5: full-stack ML and enterprise (£75,000 to £100,000+)

Data science, model training, ongoing retraining, MLOps infrastructure, regulatory documentation. Genuinely necessary in tightly regulated industries (finance, pharma, healthcare imaging) and in firms with a defensible data asset whose value comes from being differentiated.

Right call when: you have a data asset competitors don't, the outcome is regulated or differentiated enough that off-the-shelf models can't deliver, and you have the operational capacity to retrain and govern the model over time.

Wrong call when: you're using it to solve a problem that ChatGPT and a clean SOP would handle for £200 a month. I see this most often in mid-market firms that have hired a head of data who wants to build internally. The build is technically correct. It is also almost always overengineered for the actual problem.

## The three failure modes

These are the three patterns I see drain UK SMB budgets every quarter.

### Failure mode 1: Paying enterprise prices for SMB problems

A 40-person UK manufacturer signed a £62,000 fixed-fee build with a London consultancy for a "predictive maintenance AI." The actual problem was that their three production lines weren't being checked on a regular schedule and the production manager wanted alerts when downtime crossed thresholds. The £62,000 build was tier 5. The actual problem was tier 2 plus a Google Sheet. After eight weeks of integration work, the build delivered a dashboard nobody used. We replaced it with a £180-a-month Tidio variant for the alerting and a cleaned-up checklist process for the team. Result: better outcome, £61,820 less spent.

The pattern: an enterprise consultancy is selling tier 5. The SMB buys tier 5 because the consultancy is impressive and the problem feels important. The problem is tier 2 with a process fix.

### Failure mode 2: Paying SaaS prices for enterprise problems

A UK SaaS company with 80 staff bought ChatGPT Team for everyone and called it their "AI strategy." Twelve months in, the head of operations couldn't tell which workflows had actually changed. Individual productivity was up across the team. The company's operating model was identical. Customer support tickets were still routed manually, and sales qualification was still done by an SDR with a spreadsheet. Compliance documentation was still drafted from scratch every time.

Tier 1 is the right answer for individual productivity. It is the wrong answer when you need an operational change. The company eventually stepped up to a tier 3 build for support routing and a tier 2 agent for sales qualification. The ChatGPT Team licences stayed in place for personal use. Total cost across all three tiers landed around £18,000 a year, down from a planned tier 4 build the prior CTO had been pitching at £55,000.

The pattern: SaaS is the obvious starting point. It looks like the AI strategy. The reality is that SaaS is only the productivity layer. Operational change happens at tier 2 or above.

### Failure mode 3: Buying a tool when you needed a process redesign

The most common failure I see, and the most expensive over time. The owner buys a £495 agent or a £6,000 custom build to fix a problem the team is still going to do the old way. The tool runs in parallel with the existing process. Nothing meaningful changes. The licence lapses 11 months later when someone notices nobody opens it any more.

The Gallup 2026 State of the Global Workplace report put a number on this. 65% of US workers at AI-adopting firms say AI has personally made them more productive. Only 12% strongly agree the company has actually changed how it operates because of AI. The gap between those two numbers is almost entirely about whether a manager redesigned the workflow around the new capability. Buying a tool without that redesign is the most expensive way to confirm that AI works at the individual level and changes nothing at the organisational level.

The fix is not a different tool. The fix is a 90-minute meeting with the manager who owns the workflow, a whiteboard with the before-and-after process diagram, and a clear retirement of the old steps. If you can't get that meeting, don't buy the tool yet.

## How to pick the right tier

Three questions, in order, before you sign anything.

1. Is this an individual productivity problem or an operational one? If individual, you are at tier 1. If operational, you are at tier 2 or above.
2. Is the workflow bounded and templatable? If yes, tier 2 (a pre-built agent) is the default. If your specifics push you outside the template, tier 3.
3. Are you redesigning the work around the new capability? If yes, tier 2 to 4 in proportion to the integration complexity. If no, do not spend the money yet. Buy back the manager's time first.

And tier 5 is the answer for maybe 1 in 30 SMB cases I see, almost always in regulated industries with a defensible data asset. If the consultancy quoting tier 5 hasn't asked you about the data asset and the regulatory frame, they're selling tier 5 for a problem that probably isn't tier 5.

## What I'd tell the accountancy firm

The accountancy firm with the £495 and £75,000 quotes ended up at tier 3. We spent £4,200 with a UK specialist on a Custom GPT and an n8n workflow tied to their Iplicit instance. Time-to-value was nine working days. The senior partner was running 60% of their monthly invoice extraction through the workflow within a fortnight. The £495 quote would have worked for 70% of their invoices and failed on the 30% that come from international suppliers with non-standard formats. The £75,000 quote would have worked perfectly and been useless to the firm because it would have arrived four months later and required a project manager they didn't have.

Both quotes were correct for someone. But neither was correct for them.

That is almost always how this goes.

## FAQ

**What's the cheapest realistic AI project for a UK SMB?**

A ChatGPT Team subscription at £20 per user per month, plus 4 hours of someone's time writing a clean prompt library for the team. Total cost in month one: £200 to £400. That answers the question if the problem is individual productivity.

**When does a £495 AI agent make sense?**

When you have a bounded workflow (invoice extraction, lead qualification, support deflection), clean documentation for the bot to ground itself in, and a vertical match. The 20-person UK accountancy firm running invoice extraction is the textbook example.

**When is a £75,000 build actually justified?**

When the workflow has team-wide impact, requires integration with messy or on-premises systems, the volume is high enough to justify the build cost (100+ hours of human time at risk per week), and a manager is genuinely committed to redesigning the work around the new capability. Without that fourth condition, the build will sit unused.

**How do I know if I'm being oversold?**

Three signs. The consultancy hasn't asked about your existing process. They haven't asked who will own the change internally. They haven't asked what your data assets look like. If any of those three is missing, you're being sold a tier above the one you need.

**What's the typical payback period?**

Tier 1 typically pays back in weeks to a month. The pre-built agent layer (tier 2) lands in 3 to 6 months on average. Custom builds at tier 3 usually break even in 4 to 8 months, with tier 4 stretching to 9 to 18 months. Tier 5 takes 18 months or more, if it's measurable at all. The data point I'd pin to is "AI projects break even within 3 to 6 months on average," which is true for tier 2 and tier 3 and not really true above that.

If you want a structured way to figure out which tier fits your one workflow worth redesigning, the AI Roadmap audit is the fastest path. We map your operations, identify the workflow with the highest payback, and tell you which tier is genuinely correct before you take any quotes. https://richardbatt.co.uk/roadmap

---

## More about Richard Batt

Richard Batt is an AI implementation specialist who helps businesses deploy working AI automation in days, not months. 120+ projects across 15+ industries.

### Key pages

- [Home](https://richardbatt.com/)
- [About Richard](https://richardbatt.com/about)
- [Blog](https://richardbatt.com/blog)
- [Contact](https://richardbatt.com/contact)
- [Subscribe](https://richardbatt.com/subscribe)

### Contact

- Email: richard@richardbatt.com
- Location: Middlesbrough, UK (working globally)
- Website: https://richardbatt.com